Application of Generative AI in Health Systems: A Comprehensive Review of Innovations and Implications

Disse a ion Ti le:
"Applica ion o Gene a i e AI in Heal h Sys ems: A Comp ehensi e
Re iew o Inno a ions and Implica ions"
Mas e i le: MSc. In e na ional Heal h Managemen
Name: Mada Daniel Nkhoma
Yea : 2023/2024
ABSTRACT
“The e olu ion o Gene a i e A i icial In elligence has unlocked an un apped U$1 illion
imp o emen po en ial in he heal hca e indus y.” Globally, he e seems o be a pa adigm shi
ega ding he u u e o heal h sys ems, especially wi h inc eased in e es o digi isa ion, and he
elease o a Gene a i e AI known as Cha GPT in 2022. Gene a i e AI de elopmen s ha e
ma e ialised including hose ailo ed o he heal h sec o . The likelihood ha GenAI oge he
wi h o he echnological ad ancemen s will expand apace and become a i al equi emen
globally is ce ain. The e is a g owing ce ain y ha people and machines may collabo a e. This
is pa icula ly ue wi h a ansposi ion o Indus y 4.0 (4IR) and Indus y 5.0 (5IR). I seems
Gene a i e AI’s exis ence in oday’s and u u e wo kplaces is una oidable, bu a wha cos ?
G aphical Abs ac
2
CONTENTS
ABSTRACT 2
CONTENTS 3
ACKNOWLEDGEMENTS 5
DISSERTATION THESIS 7
INTRODUCTION 8
CHAPTER ONE – LITERATURE REVIEW I 14
1.1 In oduc ion 14
1.2 Theo e ical F amewo k 14
1.3 Di e en Ca ego ies o Gene a i e AI Models and hei Applica ions 16
1.4 Bene i s o in eg a ing GenAI in heal h sys ems 19
1.5 Challenges and E hical Implica ions o In eg a ing GenAI in heal h sys ems 24
1.6 Po en ial Fu u e Resea ch Di ec ions and he Implica ions o Heal hca e O ganisa ions
Conside ing he Adop ion o GenAI 27
1.7 Summa y 30
CHAPTER TWO – LITERATURE REVIEW II 31
2.1 In oduc ion 31
2.2 Gene a i e AI Models: Use Cases, Applica ions and Bene i s in Heal h Sys ems 32
2.3 Gene a i e AI Tailo ed o Heal h Sec o : Use Cases, Applica ions, Bene i s and
Challenges in Heal h Sys ems 35
2.4 Summa y 43
CHAPTER THREE – METHODOLOGY 44
3.1 In oduc ion 44
3.2 Resea ch Design 44
3.3 Popula ion 44
3.4 Ta ge popula ion 45
3.5 Sampling P ocedu e 45
3.6 Sampling Technique 45
3.7 Sample Size 46
3.8 Inclusion C i e ia 47
3.9 Exclusion C i e ia 47
3.10 Da a Collec ion Ins umen s 47
3.11 Da a Collec ion Ins umen 47
3.12 P e- es ing o Resea ch Ins umen s 48
3.13 Reliabili y 48
3.14 Replicabili y 48
3.15 Validi y 49
3
3.16 Da a Collec ion P ocedu e 49
3.17 E hical Conside a ions 49
3.18 S udy limi a ions 50
3.19 Summa y 50
CHAPTER FOUR – FINDINGS / ANALYSIS / DISCUSSION 51
4.1 FINDINGS 51
Sec ion A: Demog aphic Da a & Knowledge Assessmen 52
Sec ion B: Gene a i e AI Applica ions and hei bene i s in Heal h Sys ems 55
Sec ion C: Challenges and E hical Implica ions o Gene a i e AI in Heal h Sys ems 57
Sec ion D: Fu u e Di ec ion o GenAI in Heal hca e Managemen 58
4.1.1 Summa y 60
4.2 ANALYSIS 61
4.2.1 Da a Analysis Tools 61
4.2.2 Da a Analysis Tool 61
4.2.3 Analysed Da a 62
Sec ion A: Demog aphic Da a & Knowledge Assessmen 62
Sec ion B: Gene a i e AI Applica ions and hei bene i s in Heal h Sys ems 63
Sec ion C: Challenges and E hical Implica ions o Gene a i e AI in Heal h Sys ems 64
Sec ion D: Fu u e Di ec ion o GenAI in Heal hca e Managemen 64
4.2.4 Summa y 65
4.3 DISCUSSION 66
4.3.1 Demog aphic Da a 66
Knowledge o Gene a i e AI 68
4.3.2 Gene a i e AI Applica ions and hei Bene i s in Heal hca e Sys ems 68
4.3.3 Challenges and E hical Implica ions o Gene a i e AI in Heal hca e Managemen 69
4.3.4 Fu u e Di ec ion o GenAI in Heal hca e Managemen 70
4.3.5 Summa y 70
CONCLUDING REMARKS 72
BIBLIOGRAPHY 74
APPENDIX 0: In o med Consen 86
APPENDIX 1: Pa icipan Ques ionnai e 88
APPENDIX 2: Indus y 4.0 s Indus y 5.0 95
APPENDIX 3: E olu ion o Gene a i e AI 96
APPENDIX 4: De ini ion o Key Te ms 97
4
ACKNOWLEDGEMENTS
I would like o acknowledge he ollowing ou g oups o people o suppo ing me h oughou
my academic jou ney, hey ha e shaped he excellence o my schola ship. These a e:
Fi s ly, G ena om he Umodzi Wa aMalawi Communi y in Ge many who helped me o explo e
a ious sou ces o eliable da a du ing he b ains o ming s age.
Secondly, I ex end my since es g a i ude o my supe iso , Zoi Kai ou. He expe ise,
encou agemen , and cons uc i e eedback ha e been ins umen al in shaping his disse a ion
and my g ow h as a schola .
Fu he mo e, I would like o exp ess my deepes g a i ude o my gi l iend, Esna Rumando, o
dis u bing me wi h phone calls mos o he ime. I was u ious, bu i u ned ou o be a
wonde ul e eshing ime, a ime o e-ene gise, as he so -swee - oice and he
ne e -ending-s o ies ha e been a cons an sou ce o s eng h, and I am p o oundly g a e ul o
he necessa y dis u bances e e y s ep o he way.
Finally, I would like o ex end my hea el app ecia ion o he diligen e o s and insigh ul
g amma ical checks and o e all eedback p o ided by Sammuel Shakunle, Jessica Shakunle and
Tayo Shakunle. Thei aluable con ibu ions ha e signi ican ly en iched he quali y and dep h o
his disse a ion endea ou , o which I am since ely g a e ul.
5

S a emen o compliance wi h academic e hics and he a oidance o plagia ism
I hones ly decla e ha his disse a ion is en i ely my own wo k and none o i s pa has been copied
om p in ed o elec onic sou ces, ansla ed om o eign sou ces and ep oduced om essays o o he
esea che s o s uden s. Whe e e I ha e been based on ideas o o he people's ex s I clea ly decla e i
h ough he good use o e e ences ollowing academic e hics.
(In he case ha i is p o ed ha pa o he essay does no cons i u e an o iginal wo k, bu a copy o an
al eady published essay o om ano he sou ce, he s uden will be expelled pe manen ly om he
pos g adua e p og am).
Name and Su name (Capi al le e s):
MADA DANIEL NKHOMA
Da e: 10/03/2024
6
DISSERTATION THESIS
(lea e his page emp y)
7
INTRODUCTION
“The e olu ion o Gene a i e A i icial In elligence has unlocked an un apped U$1 illion
imp o emen po en ial in he heal hca e indus y” (McKinsey & Company, 2023). Globally,
he e has been a pa adigm shi ega ding he u u e o heal h sys ems, especially wi h he elease
o a Gene a i e AI (GenAI) known as Cha GPT in 2022. The e is a g owing ce ain y ha people
and machines may collabo a e on speci ic adminis a i e asks (Wee a a hna e ., al. 2023). This
is pa icula ly ue wi h he ansposi ion o Indus y 4.0, which began in 2010 and Indus y 5.0,
which s a ed ui ion in 2020 (Xu e al. 2021; Appendix 2).
Acco ding o McKinsey & Company (2023), in eg a ing human skills wi h Gene a i e AI ools
imp o es pe o mance. As he applica ion o echnological models o managemen inc eases,
heal h sys ems managemen is no spa ed, and hese inno a ions a e no only bene icial bu also
pose challenges and implica ions (Yamin, 2018).
This chap e ocuses on a b ie in oduc ion, backg ound o he s udy, s a emen o he p oblem,
esea ch objec i es, esea ch ques ions, me hodology, he signi icance o he s udy, delimi a ions
o he s udy and he limi a ions o he s udy.
Backg ound o he S udy
A i icial In elligence (AI) widens human abili ies. This is he ideal o ensu ing compe i i e
heal h ins i u ions. Th ee main ypes o AI include A i icial Na ow In elligence (ANI),
A i icial Gene al In elligence (AGI) and A i icial Supe In elligence (ASI) (Figu e 1). ANI is
designed o pe o m a speci ic ask, and p e-de ined pa ame e s limi hei abili ies. Fo example,
medical diagnos ics (Kuusi & Heinonen, 2022). On he o he hand, AGI also e e ed o as
“Human-Le el AI” aims a c ea ing in elligen machines ha can au onomously sol e any
in ellec ual ask which a pe son can do, including gene a ing no el o c ea i e in o ma ion
(Goe zel, 2014).
8
Alan Tu ing designed “ he Tu ing Tes ” in 1950 and in es iga ed he po en ial o Human-Le el
AI. He is c edi ed wi h being he i s o p opose he concep o hinking machines o Gene a i e
AI (GenAI). Subsequen ly, in 1960 Joseph Weizenbaum c ea ed he i s Gene a i e AI known as
he “Eliza cha bo .” The Tu ing es is used as an impo an benchma k o he ongoing
de elopmen o AGI. ASI is he hi d and inal main ype o AI. I is a ype o a i icial
in elligence an icipa ed o exceed human-le el in elligence in b oad cogni i e spec ums. ASI
will ha e he capaci y o pe o m asks beyond human abili y (Bos om, 2014).
This pape in es iga es he ad an ages o inco po a ing A i icial Gene al In elligence,
speci ically Gene a i e AI in o heal h sys ems while in es iga ing he po en ial implica ions o
wo k low, p ocesses and he heal h sys em. Gene a i e AI u ilises a i icially in elligen sys ems
o desc ibe algo i hms which can be used o gene a e new and o iginal con en like ex , images,
ideo, o audio as a esponse o p omp s ini ia ed by use s (Va ghese & Chapi o, 2023). The
majo i y o GenAI is de eloped using neu al ne wo ks which a e designed o imi a e he
s uc u es o he b ain like Deep Lea ning and La ge Language Models (LLMs). No able GenAI
applica ions a e Cha GPT, Google Ba d, Midjou ney and DALL-E. Acco ding o Bi an (2023),
Gene a i e AI ensu es heal hca e managemen e iciency by s eamlining he wo k low
p ocesses and acili a ing he c ea ion o digi al wins (AWS, 2024).
9
1.3 Di e en Ca ego ies o Gene a i e AI Models and hei Applica ions
Gene a i e AI models con inue o ad ance, and hei applica ions span a wide ange o domains,
om c ea i e a s o p ac ical p oblem-sol ing in heal hca e, economics, and so o h. Acco ding
o Law on (2023), he choice o model o en depends on he speci ic use cases, he equi emen s
o asks and he ype o da a in ol ed (Mille , 2022). Gene a i e AI models gained ac ion in he
mid-2010s a e he de elopmen o Va ia ional Au o-Encode s (VAEs), Gene a i e Ad e sa ial
Ne wo ks (GANs) and di usion models (Law on, 2023; Choudha y e al. 2022). B eak h oughs
in gene a i e neu al ne wo k models like T ans o me s which can analyse la ge da ase s o
au oma ically c ea e La ge Language Models (LLMs) came abou in 2017 (Lin, 2022). In 2022,
Neu al Radiance Fields (NeRFs), a echnique o gene a ing 3-dimensional (3D) con en om
2-dimensional (2D) images was in oduced (Appendix 3). 3-dimensional images a e key in
a ious simula ions, media, gaming and he In e ne o Things (IoT). In heal hca e, NeRFs can
be use ul in medical imaging by acili a ing he c ea ion o comp ehensi e ana omical and
physiological s uc u es om 2-D scans like Magne ic Resonance Imaging (MRIs). Tha is, i
can econs uc ealis ic ep esen a ions o body issues and o gans, gi ing doc o s and clinic
manage s use ul isual con ex (Co ona-Figue oa e al. 2022).
VAEs a e o en used o gene a ing ealis ic images, ex , speech o ideos and o asks equi ing
p obabilis ic ep esen a ions (Singh & Ogun unmi, 2021). GANs a e mo e applicable when
gene a ing ealis ic da a and in image p ocessing (Good ellow e al. 2020). The di usion model,
on he o he hand, pe o ms be e use cases ela ed o image syn hesis, ideo gene a ion, and
molecule designing (Yang e al. 2023).
Fu he mo e, T ans o me s can be applied in na u al language p ocessing, compu e ision,
speech syn hesis, music gene a ion, and mul imodal applica ions such as isual ques ion
answe ing and commonsense easoning (Lin, 2022). T ans o me -based models, such as GPT
(Gene a i e P e- ained T ans o me ) and BERT (Bidi ec ional Encode Rep esen a ions om
T ans o me s) a e a b eak h ough well known o hei abili y o gene a e cohe en and
con ex ually ele an ex , and we e popula ised by OpenAI and Google espec i ely (P a im,
2023). Recen ly, T ans o me s ha e been used by DeepMind o AlphaS a (Zhang e al. 2023).
16

O he no able GenAI models include Au o eg essi e Models, Recu en Neu al Ne wo ks
(RNNs) and Long Sho -Te m Memo y (LSTM) (She s insky, 2020). Au o eg essi e models
gene a e sequences one elemen a a ime, condi ioning on p e ious elemen s. Acco ding o
Bu an (2023), Au o eg essi e models a e use ul when o ecas ing u u e e en s based on
his o ical da a. Fo example, p edic ing pa ien ou comes and assessing he ela ionship be ween
physical heal h and psychological well-being. RNNs and LSTMs a e o en used o gene a ing
da a sequences, such as ex o music. Besides, RNNs and LSTMs can be applied in na u al
language p ocessing, c ea i e con en gene a ion, speech- o- ex ansc ip ion, and machine
ansla ion (Bandi e al. 2023; Bu an , 2022; She s insky, 2020).
Table 1: Examples o GenAI and I s Applica ions
Wha i Gene a es
Desc ip ion
Applica ion Example (s)
Tex
Cha bo s o AI w i ing ools
gene a e new ex as a
esponse o a p omp om he
use (e.g. an answe o a
ques ion, a summa y, a
pa aph ase o a ansla ion)
Cha GPT, DeepL T ansla o ,
QuillBo Pa aph ase , Sc ibb
Tex Summa ize , Ba d &
Bing AI
Code
Na u al language p ocessing
(e.g. English; Chinese),
Ou pu ex in di e en
p og amming languages (e.g.
Ja aSc ip ).
OpenAI Codex, Gi Hub
Copilo
Images
LLMs can be used o gene a e
images a he han ex . These
apps ake a ex -based p omp
(e.g. “The Measles pa ien
wi h symp oms”) and u n i
DALL-E, P isma,
Midjou ney, and S able
Di usion
17
in o an image. Some modi y a
use -submi ed image.
Videos
These can c ea e whole
ideos. The echnology o
his GenAI is s ill imp o ing
Syn hesia, Gen-2,
Make-a-Video, OpenAI’s
SORA
Audio
GenAI may be used o
gene a e he apeu ic music
and syn hesised oices.
Tex - o-speech wi h a
e sa ile AI oice gene a o
MusicLM, MusicGen,
MuseNe , Mu AI
O he
GenAI has po en ial in o he
a eas, o example, in ha d
sciences (e.g. p edic ing
p o ein s uc u es) and in
obo ics (e.g. u ning ex
p omp s in o ac ions done by
obo s).
UniPi, AlphaFold
Sou ce: (Caul ield, 2023; OpenAI, 2024)
Acco ding o Caul ield (2023), he e a e a ange o GenAI applica ions which can be u ilised in
a ious ields including heal h managemen (Table 1). Simila ly, o he de elopmen s ha e
e ealed GenAI ools speci ically designed o imp o e he e iciency o heal h sys ems such as
Azu e AI Heal h Bo (Bi an, 2023). In addi ion o ha , Epic and Mic oso pa ne ed o in eg a e
AI in o Elec onic Heal h Reco ds (EHR), allowing heal hca e manage s and ca e p o ide s o
imp o e p oduc i i y & pa ien communica ion wi h AI-enabled solu ions (Epic, 2023).
Mo eo e , he Mayo Clinic, he la ges in eg a ed medical g oup p ac ice in he wo ld has
adop ed Gene a i e AI App Builde o enhance he capaci y o i s heal h sys ems (Landi, 2023).
18
1.4 Bene i s o in eg a ing GenAI in heal h sys ems
Acco ding o Yu e al. (2023), GenAI can e olu ionise how heal h da a and in o ma ion a e
handled, p ocessed and managed. Mo eo e , a sys ema ic e iew o 60 selec ed pape s ha
assessed he u ili y o GenAI in heal hca e wi h Cha GPT as a case s udy e ealed ha 85% (51
o 60) o pape s e iewed ci ed bene icial applica ions like s eamlining he wo k low,
cos -sa ing, imp o ed documen a ion, pe sonalised medicine, and imp o ed heal hca e li e acy,
while a s agge ing 97% (58 o 60) aised conce ns (Sallam, 2023). On he o he hand,
adminis a i e expendi u e in he U.S.A., he la ges and mos complica ed heal hca e model in
he wo ld, accoun s o 15 o 30% o heal hca e spending (Heal h A ai s ci ed in Adne &
Weins ein, 2023). As such, GenAI is poised o imp o e billing and claims p ocessing as well as
esou ce managemen and quali y. McKinsey & Company (2023) no e ha GenAI is poised o
imp o e he e iciency and e ec i eness o heal hca e managemen ope a ions due o i s abili y
o au oma e and summa ise da a, e en big da a, wi hin a sho pe iod. This educes u na ound
ime as heal h adminis a o s ocus on o he impo an asks.
Fo ins ance, BioNTech, a bio echnology company, acqui ed Ins aDeep, a ype o GenAI, and
de eloped an ea ly-wa ning sys em o no el COVID-19 a ian s (Bos on Consul ing G oup,
2023). S uc u al modelling o he SARS-CoV-2 p o ein in eg a ed wi h Ins aDeep’s Gene a i e
AI capabili ies enables he sys em o o ewa n and ale esea che s, accine de elope s, heal h
au ho i ies, and policymake s (Bos on Consul ing G oup, 2023). In addi ion, heal hca e could be
e olu ionised by imp o ing documen a ion accu acy and le e aging s uc u ed and uns uc u ed
da a o c ea e accessible clinical and adminis a i e eco ds, gene a ing simula ions aimed a
educa ing employees and pa ien s in c i ical a eas and c ea ing mee ing o wo kshop summa ies.
Fu he mo e, Na u al Language P ocessing (NLP) models can be used o unde s and and ex ac
in o ma ion om uns uc u ed heal hca e eco ds. This acili a es e icien da a managemen ,
enables end analysis, and suppo s clinical and managemen decision-making (A anade, 2024;
Dash e al. 2019). Acco ding o Dash e al. (2019), an NLP-based algo i hm known as
“Linguama ics” u ilises an in e ac i e ex mining algo i hm (I2E).
19
I2E is capable o ex ac ing and analysing as amoun s o in o ma ion use ul o bo h clinical
and adminis a i e p oblem-sol ing (Figu e 3). The s uc u ed da a gene a ed by ex mining
could be in eg a ed in o da a wa ehouses, da abases o business in elligence dashboa ds and may
be u ilised o p esc ip i e, desc ip i e o p edic i e analy ical decision-making (Linguama ics,
2023). Resul s ob ained using his echnique a e en old as e han o he simila ools and do no
equi e expe knowledge o da a in e p e a ion (Dash e al. 2019; Linguama ics, 2023).
Figu e 3: An NLP-based GenAI Sys em Used in Big Da a Re en ion and Analysis
A schema ic ep esen a ion o he wo king p inciple o an NLP-based AI sys em used in gigan ic
da a e en ion and analysis in Linguama ics.
Sou ce: (Dash e al. 2019)
Many heal h sys ems a e g appling wi h high cos s o ca e, s a sho ages and employee
bu nou (Deloi e Cen e o Heal h Solu ions Su ey, 2023). To sol e his, Value-Based Ca e is
key. I aims a ensu ing quali y o ca e while p omo ing employee wellbeing, and pe sonalised
and a o dable ca e. Tex p omp s in o he GenAI like GPT-4 could d a s anda d Value-Based
Ca e and ca e ou con ac s based on ma ke ends and cha ac e is ics (McKinsey & Company,
2023; Jayakuma e al. 2023).
Acco ding o Deloi e Heal h Ca e Consume Su ey (2023), o e hal o he Uni ed S a es
popula ion (53%) is hope ul ha Gene a i e AI could imp o e access o heal h ca e; while 47%
su eyed esponded ha GenAI has he po en ial o make heal hca e mo e a o dable. On he
o he hand, esponden s who had al eady expe ienced Gene a i e AI we e mo e op imis ic.
20
Tha is; 69% belie e i could imp o e access o heal hca e. And 63% indica ed ha GenAI has
he po en ial o p omo e Value-Based Ca e. Such esul s a e based on a na ionally ep esen a i e
su ey o 2,014 US adul s (Deloi e Heal h Ca e Consume Su ey, 2023). Hence, GenAI may
ca alyse he implemen a ion o Value-Based Ca e (Jayakuma e al. 2023). Mo eo e , i is
belie ed ha A i icial In elligence can imp o e all kinds o p ocesses wi hin bo h heal hca e
ope a ions and ca e deli e y (Da enpo & Kalako a, 2019). Fo example, he cos sa ings ha
AI could b ing o he heal hca e ecosys em is a pi o al d i e o he implemen a ion o AI
applica ions, especially wi h he eme gence o Gene a i e AI which is capable o c ea ing new
and o iginal con en c i ical o quick decision-making and educ ion o u na ound ime
(A anade, 2024). I is es ima ed ha in eg a ing AI sys ems in heal hca e managemen could cu
annual heal hca e cos s in he Uni ed S a es by U$150 billion by 2026 (Boh , & Mema zadeh,
2020). Acco ding o Boh & Mema zade (2020), a big pa o hese educ ions in cos s s ems
om shi ing he heal hca e model om a eac i e one o a mo e p oac i e app oach, wi h an
emphasis on heal h managemen a he han disease ea men .
Fu he mo e, Robo ic P ocess Au oma ion (RPA) associa ed wi h GenAI pe o ms s uc u ed
( epe i i e) digi al asks o adminis a i e pu poses. Fo example, hose in ol ing in o ma ion
sys ems. RPAs handle adminis a i e asks e icien ly, as i hey we e a human use ollowing a
pa icula sc ip o some ules. Compa ed o o he ypes o AI hey a e inexpensi e, easy o
p og am wi h a high deg ee o anspa ency in hei ac ions. Mo eo e , inco po a ing Robo ic
P ocess Au oma ion (RPA) applica ions concu en ly wi h GenAI sys ems in heal h sys ems
could signi ican ly educe cos s and imp o e adminis a i e p ocedu es (MSG Su ey, 2023).
These echnologies a e needed in heal hca e because, o ins ance, he a e age nu se in he
Uni ed S a es spends 25% o wo k ime on egula o y and adminis a i e ac i i ies which could
o he wise be done by RPAs (Be g ci ed in Boh & Mema zadeh, 2020). RPAs can be used o a
my iad o applica ions in heal hca e, including he p ocessing o claims, clinical documen a ion,
managemen o e enue cycle as well as medical eco ds.
21

No ably, RPA does no in ol e obo s, bu a he , only compu e p og ams on se e s (Boh , &
Mema zadeh, 2020). Hence, i elies on a combina ion o business ules, wo k low and
“p esen a ion laye ” in eg a ion wi h heal h in o ma ion sys ems o ac like a semi-in elligen
use o he sys ems. In heal hca e sys ems, RPAs a e used o epe i i e asks such as p io
au ho isa ion, upda ing pa ien da a eco ds and billing among o he asks. When combined wi h
o he echnologies like OpenAI’s DALL-E, a Gene a i e AI o image c ea ion and ecogni ion,
RPAs can be used o e ie e da a om, o ins ance, axed images o inpu i in o ansac ional
sys ems (Hussain ci ed in Boh , & Mema zadeh, 2020). This makes heal hca e p ocesses mo e
e icien , e ec i e and inexpensi e.
GenAI could os e echnical ope a ional e iciency as hey can be used o d a Reques Fo
P oposals (RFPs) and equisi ions, gene a e epo s and Key Pe o mance Indica o s, d a
endo communica ions, and c ea e pu chase o de s based on supply le els (McKinsey &
Company, 2023). The e o e, since heal h acili ies and depa men s ha e ope a ed in silos o
gene a ions, GenAI has he po en ial o imp o e in e ope abili y o heal hca e sys ems. In heal h
sys ems, GenAI may use uns uc u ed pu chasing and accoun s payable da a and wi h he aid o
cha bo s, can add ess common hospi al employee in o ma ion and human esou ce ques ions.
Hence, i could imp o e employee expe iences and may educe ime and money spen on
hospi al adminis a i e cos s (McKinsey & Company, 2023).
GenAI p omises bene icial impac s in Popula ion Heal h Managemen (GenHeal h.ai, 2023).
Gene a i e AI models like P e- ained T ans o me s a e spon aneous. The in eg a ion o
p edic i e analy ics may p o ide mo e accu a e and imely iden i ica ion o high- isk pa ien s
and a - isk popula ions, enhancing ca e coo dina ion, educing adminis a i e bu den and
imp o ing o e all popula ion heal h ou comes. (Ya aghi, 2024).
In o he de elopmen s, heal hca e o ganisa ions ha e also expe imen ed wi h cha bo s o
pa ien in e ac ion, men al heal h p omo ion and wellness p og ams, including eleheal h (Boh
& Mema zadeh, 2020). These NLP-based applica ions could be use ul o simple ansac ions
like scheduling mee ings o appoin men s.
22
In addi ion, Doo e al. (2023) indica es ha Gene a i e A i icial In elligence is eliable when
de e mining popula ion heal h ou comes and ends e ospec i ely, as such, may be help ul when
analysing pas public heal h challenges (Doo e al. 2023). Howe e , as GenAI models a e s ill
de eloping, hey a e unable o p edic cu en and u u e popula ion heal h- ela ed da a.
Simila ly, GenAI could help o add ess dispa i ies in heal hca e as i p omises o bene i he
uninsu ed popula ion (Dha , Fe a & Ko enda, 2023). People who a e no co e ed a e mo e likely
o use Gene a i e AI when seeking heal h ca e like men al heal h suppo , o ind a doc o , o o
iden i y he mos app op ia e o closes ca e loca ion (Fo ins ance, an eme gency oom o a
doc o ’s o ice). This is because ulne able popula ions like he uninsu ed end o be
p ice-sensi i e and may no be able o a o d as li le as a doc o consul a ion ee (Kaise Family
Founda ion ci ed in Dha , Fe a & Ko enda, 2023).
Mo eo e , AI-powe ed ools could p o ide access o li e-sa ing ca e swi ly and eliably,
including heal h expe ise o millions o ulne able people in need (WeFo um, 2024c). Despi e
ha , c i ics a gue ha GenAI could exace ba e dispa i ies in heal hca e as high-income coun ies
can in es in and upscale he ele an echnologies apidly (WeFo um, 2024b; Khan e al. 2023;
Omiye e al. 2023). To make ma e s wo se, an a e age pe son in u al a eas o low-income
loca ions expe iences in e mi en access o elec ici y and an a e age pe son does no ha e
access o in e ne connec i i y, ende ing GenAI applica ions ha de o adop inclusi ely.
Compa ed o adi ional AI sys ems, which a e gene ally ule-based o ely on p e-de ined
da ase s, GenAI models possess a unique abili y o c ea e new con en ha is o iginal and
unp og ammed. This can esul in ou pu s ha a e simila o he use p omp in one, s yle, o
s uc u e. Hence, i designed hough ully and de eloped esponsibly, GenAI can ampli y heal h
manage s' capabili ies in a ious domains such as suppo o decision-making, knowledge
e ie al, answe ing ques ions, language ansla ion when supe ising mul icul u al eams, and
au oma ic epo ing (Yu e al. 2023).
23
1.5 Challenges and E hical Implica ions o In eg a ing GenAI in heal h sys ems
While GenAI p esen s p omising esul s, ha m ul ou comes in he managemen o heal h se ices
can a ise. These may include a endency o LLMs o hallucina e. Acco ding o he In e na ional
Business Machines Co po a ion (IBM), hallucina ion is a scena io whe eby GenAI cha bo s o
compu e ision ools pe cei e pa e ns o objec s ha a e non-exis en o incomp ehensible by
human obse e s, gene a ing ou pu s ha a e nonsensical o al oge he alse (Ji e al. 2023). This
leads o inco ec o misleading esul s gene a ed by he AI models (Ha em, Simmons, &
Tho n on, 2023). E o s such as insu icien aining da a, inco ec AI model assump ions made,
o biases in he da a which is used o ain he model con ibu e o he hallucina ion p oblem
(Google Cloud, 2024; Alkaissi & McFa lane, 2023).
Bang e al. (2023) ind ha GenAI, like Cha GPT, a e o e all 63.41% accu a e in 10 di e en
easoning classi ica ions unde non- ex ual easoning, logical easoning, and commonsense
easoning. In his ega d, I can be a gued ha Gene a i e AI is an un eliable easone . Besides,
GenAI has an issue o algo i hmic bias (Oniani, 2023). Acco ding o Panch, Ma ie & A un
(2019), algo i hmic biases exace ba e dispa i ies in heal hca e. The e o e, i may be de imen al
o heal hca e manage s o ely solely on GenAI o p oblem-sol ing and c i ical answe s.
Con a y o Dha , Fe a & Ko enda's (2023) belie ha GenAI could educe dispa i ies in
heal hca e, a s udy by Omiye e al. (2023) shows ha La ge Language Models could po en ially
pe pe ua e ha m ul p ac ices by p omo ing debunked acis ideas (Figu e 3). Besides, mino i y
g oups may be p esen ed wi h dis o ed ou comes which migh be he esul o biases when
mining big da a used o in o m he de elopmen o such models. Apa om ha , he
unde - ep esen ed mino i ies as a consequence o acial biases in he de elopmen o da ase s
could be p esen ed wi h poo p edic ions and un ep esen a i e esul s (Khan e al. 2023; Omiye
e al. 2023). Technological biases ha e been p o en in medical p ac ices. Fo example, he e is
sys emic and s uc u al acism in Ge man medicine, ools like Oxime e s ha e been ound o
eco d misleading diagnoses o black pa ien s in Ge many and he e is he po en ial o GenAI o
exace ba e such dispa i ies i due diligence is no done (Deu sche Welle, 2024).
24
Figu e 4: La ge Language Models P opaga e Debunked Race-based Medicine
Fo each ques ion and each model, he a ing ep esen s he numbe o uns (ou o 5 o al uns)
wi h dis u bing ace-based esponses. Red indica es a highe numbe o dis u bing and
po en ially a al ace-based esponses.
Sou ce: (Omiye e al. 2023)
Al hough Dash e al. (2019) and Linguama ics (2023) claim he as a ay o bene i s o
NLP-based models in heal hca e, a su ey in ol ing 500 use s o he bes i e cha bo s u ilised
in heal hca e in he Uni ed S a es o Ame ica indica es ha clien s exp essed deep conce ns as
ega ds discussing complex heal h condi ions h ough he aid o cha bo s o e ealing
con iden ial in o ma ion. Besides, a lack o use - iendliness was also ci ed as a de e en
(MedNews, 2019).
Apa om ha , he e a e egula o y challenges associa ed wi h Gene a i e AI (Hacke , Engel &
Maue , 2023). Acco ding o he p oposed Eu opean Union A i icial In elligence Ac o 2021, a
lack o a p ope egula o y amewo k o guide he de elopmen and deploymen o AI agen s in
he human ma ke is a se ious issue (Eu opean Pa liamen , 2023). As A i icial In elligence
echnology is e ol ing e y as , he ques ions o us wo hiness a e ine i able.
25
2.2 Gene a i e AI Models: Use Cases, Applica ions and Bene i s in Heal h Sys ems
As he ield o A i icial In elligence con inues o e ol e, Gene a i e AI models and hei
applica ions could ans o m how adminis a i e and o he heal h managemen p ocesses a e
done (Mille , 2022). Fo example, om c ea ing execu i e epo s o p ac ical p oblem-sol ing in
heal hca e, heal h economics and so o h (WeFo um, 2024c). As Law on (2023) and Mille
(2022) no e, he choice o model o en depends on he speci ic use cases, he equi emen s o
asks and he ype o da a in ol ed. Following his backg ound, Va ia ional Au oencode s
(VAEs) can be applied in he managemen o heal h se ices, pa icula ly in he domains o
human esou ce managemen and he execu ion o adminis a i e asks. In his ega d, VAEs can
gene a e syn he ic da a ha can esemble eal-wo ld da a, no wi hs anding pa e ns and
dis ibu ions ound in heal h se ice managemen p ocesses (Giu è & Shung, 2023; Bandi e al.
2023). Such da a can be aluable o aining machine-lea ning models used in he managemen
o human esou ces o heal h.
Addi ionally, VAEs can be applied o analyse his o ical human esou ce da a, o example,
s a ing le els, wo kloads, and pe o mance app aisal me ics due o hei abili y o ecognise
pa e ns (Choudha y e al. 2022). The e o e, by lea ning he unde lying pa e ns in his and
ela ed da a, VAEs can assis in his o ical wo k- ela ed assessmen s like pe o mance e iews,
aining eco ds, and public heal h p ojec e alua ions and ou comes. Besides, i could be able o
p edic u u e s a ing needs, op imise esou ce alloca ion, and ensu e adequa e wo k o ce
a ailabili y.
Mo eo e , VAEs a e adep a de ec ing anomalies in da a (Gonzalez e al. 2021). As such, in he
con ex o heal h se ices adminis a ion, Va ia ional Au oencode s can be e icien ly u ilised o
iden i y unusual pa e ns o de ia ions in adminis a i e asks, ale ing managemen o po en ial
issues such as billing e o s, da a en y mis akes, o i egula i ies in p ocess execu ion (Gonzalez
e al. 2021). Mo eo e , since VAEs a e ained on big da a, hey could model complex
co ela ions be ween esou ce alloca ion decisions and ou comes in heal h se ice managemen
o adminis a ion (Law on, 2023). Hence, VAEs could p o ide insigh s ela ed o esou ce
alloca ion, whe he i be inancial esou ces, equipmen dis ibu ion, human capi al o indi idual
employee assignmen s.
32

As ega ds Human Capi al Managemen , VAEs could analyse he skill se s, quali ica ions, and
indi idual aining eco ds o heal hca e p o essionals (Gonzalez e al. 2021). This could help in
wo k o ce planning by iden i ying ea u es ep esen ing speci ic skills. Hence, hey could assis
in ma ching heal h sec o pe sonnel o asks. Addi ionally, VAEs could p ocess da a ela ed o
employee sa is ac ion su eys, eedback, and wellness p og ams.
On he o he hand, Gene a i e Ad e sa ial Ne wo ks (GANs) could be employed o c ea e
simula ions o employee aining ac i i ies (L , 2023). This may include simula ed scena ios o
hospi al adminis a i e asks, pa ien in e ac ions, o eme gency esponse, allowing heal h
pe sonnel o p ac ise and enhance hei skills in a con olled i ual en i onmen . Acco ding o
L (2023), GANs a e capable o gene a ing di e se and ealis ic con e sa ional da a. This da a
could be used o ain ailo ed heal hca e cha bo s and i ual assis an s. These AI-d i en ools
could hen assis in adminis a i e asks like documen ing p esen a ions, answe ing employee
que ies, o p o iding in o ma ion abou heal h se ices which a e o e ed by a pa icula heal h
uni . No only ha , GANs could also enhance acial ecogni ion sys ems used o access con ol
in heal hca e acili ies (Shahbakhsh & Hassapou , 2023). This is pa icula ly ele an in
managing human esou ces as i ensu es secu e access o es ic ed a eas, moni o s a endance
and enhances he o e all secu i y o heal h uni s.
Apa om ha , GANs could augmen da ase s ela ed o adminis a ion wo k by gene a ing
a ia ions o a ailable da a (Wu, S ou s, & Biljecki, 2022). This can be bene icial o asks such
as documen classi ica ion, da a en y, and wo k low op imiza ion. Besides, i could also be used
o oice syn hesis o c ea e ealis ic oice models. In he managemen o heal h se ices, his
can imp o e he quali y o elephony se ices in e nally wi h employees o ex e nally wi h
s akeholde s, o ins ance, au oma ed appoin men eminde s. No wi hs anding, GANs could
assis in gene a ing ealis ic and in o ma i e epo s o adminis a i e pu poses quickly and
eliably, he eby educing u na ound ime, as hey can be u ilised o au oma e he c ea ion o
managemen epo s, inancial summa ies, and o he documen s essen ial o a well unc ioning
heal h sys em (Wu, S ou s, & Biljecki, 2022).
33
Ano he no able GenAI model is T ans o me s, a ype o Deep Lea ning model (L , 2023;
P a im, 2023)). As discussed in he p eceding chap e , an ad anced ype o T ans o me s
cu en ly shi ing he pa adigm is called Gene a i e P e- ained T ans o me s (GPTs).
T ans o me s ha e demons a ed g ea e sa ili y in Na u al Language P ocessing (NLP) and
sequen ial da a asks (L , 2023). Due o ha , T ans o me s could be used o build cha bo s o
i ual assis an s capable o unde s anding and esponding o na u al language que ies abou
adminis a i e asks. In he con ex o heal h managemen , his could include inqui ies abou
upcoming schedules, policies, lea e eques s, o gene al in o ma ion abou heal h se ices o he
wo kplace. Fo example, by unde s anding and p ocessing na u al language eques s,
T ans o me s may be able o in e ac wi h scheduling sys ems o ind a ailable slo s, con i m
appoin men s, and send eminde s o pa ien s and hospi al s a (Nah e al. 2023).
Fu he mo e, T ans o me s can assis in s eamlining he onboa ding p ocess o new heal hca e
pe sonnel by p o iding in o ma ion abou o ganisa ional policies, p ocedu es, and aining
ma e ials (Doo e al. 2023). Addi ionally, hey can answe ques ions and guide employees
h ough he onboa ding p ocess. Also, hey may be u ilised o c ea e in e nal memos and epo s,
and can summa ise leng hy adminis a i e epo s, con ac s, policies, and guidelines (Nah e al.
2023). GPTs can also acili a e he gene a ion o accu a e billing codes and could assis wi h
medical coding asks in clinical adminis a ion (Luo e al. 2022; New on, 2023). This could help
heal hca e adminis a o s o quickly ex ac key in o ma ion and make in o med decisions
spon aneously (A anade, 2024).
Simila ly, ans o me s could also acili a e e icien e ie al o a e and old heal h in o ma ion
om la ge documen eposi o ies, o example, when needed by audi o s o o e e ence (Luo e
al. 2022; New on, 2023). No only ha , bu T ans o me s could also enhance email managemen
by au oma ically ca ego ising, summa ising, and esponding o ou ine emails (L , 2023). This
can help heal hca e adminis a o s ocus on c i ical asks while ensu ing ha ou ine
communica ion is handled e icien ly. As a as human capi al is conce ned, T ans o me s could
be u ilised o esume sc eening and candida e ma ching du ing he ec ui men p ocess. Fo
example, hey can analyse esumes by iden i ying ele an skills and expe iences while ma ching
candida es o speci ic job equi emen s. This could s eamline he hi ing p ocess.
34
In heal hca e uni s wi h mul ilingual eams o which se e pa ien s om a mul icul u al
backg ound, T ans o me s could p o ide eal- ime ansla ion se ices, acili a ing e ec i e
communica ion and collabo a ion among di e se s a membe s (Luo e al. 2022; New on, 2023).
Simila ly, T ans o me s could be in eg a ed in o backend sys ems o au oma e ou ine
adminis a i e asks, such as upda ing da a, eco d keeping, and da a econcilia ion, imp o ing
he o e all ope a ional e iciency o heal h se ices adminis a ion (Doo e al. 2023).
2.3 Gene a i e AI Tailo ed o Heal h Sec o : Use Cases, Applica ions, Bene i s and Challenges
in Heal h Sys ems
Acco ding o Wald on (2023), he e a e ou b oad a eas whe e A i icial In elligence and
Gene a i e AI a e le e aged in he heal hca e ecosys em: The i s one is adminis a i e as well
as ope a ional e iciency. Secondly, AI is le e aged in clinician and ca e eam
suppo . Fu he mo e, inpa ien and consume engagemen , and inally ad ancing esea ch and
de elopmen s a egies o esilien heal h sys ems (Wald on, 2023). The wo no able AI ools
which a e challenging he s a us quo in heal hca e a e:
(i) Azu e AI Heal h Bo : Applica ions and Bene i s
To begin wi h, Azu e AI Heal h Bo is a cloud-based con e sa ional A i icial In elligence
se ice o e ed by Mic oso Azu e, designed speci ically o heal hca e o ganisa ions (Hoo ne,
2023). Acco ding o Hoo ne (2023), i enables heal hca e wo ke s o build, deploy, and manage
AI-powe ed cha bo s and i ual heal h assis an s o engage wi h bo h pa ien s and colleagues.
Azu e Heal h Bo inco po a es ad anced na u al language unde s anding capabili ies, allowing i
o in e p e and unde s and use inpu s in na u al human language (Bi an, 2023). This enables
hospi al o clinic adminis a o s o in e ac in eal ime wi h he bo using con e sa ional
language (Mic oso , 2024). Azu e is p o ound in he managemen o heal h se ices because
Heal hca e o ganisa ions can easily cus omise he design ea u es and beha iou o he cha bo o
align wi h he speci ic needs o a heal h uni and p e e ences o he ask a hand (Bi an, 2023).
This could include de ining con e sa ional lows, designing use in e aces, and in eg a ing
a ious depa men al wo k lows o ensu e seamless communica ion ac oss depa men s (Hoo ne,
2023).
35
In addi ion, Azu e Heal h Bo suppo s deploymen ac oss mul iple channels, including
websi es, mobile apps, messaging pla o ms (like Mic oso Teams, Skype, and Facebook
Messenge ), and oice-enabled de ices like Mic oso Co ana and Amazon Alexa (Mo e -Ta ay,
Radawski, & Gua iglia, 2022). I p o ides a cen alised pla o m o scheduling appoin men s,
managing inqui ies and dissemina ing in o ma ion. By eaching each esponsible employee using
he pla o m o his o he choice, Azu e could ensu e e ec i e and imely communica ion o
u gen adminis a i e ma e s like low s ocks o clinical sund ies and supplies which need an
u gen e ill (Bi an, 2023; Mic oso , 2024).
Wi h Azu e Heal h Bo , scalable and lexible se ices can be adap ed o changing demands and
he changing needs o he heal hca e landscape (Hoo ne, 2023). By au oma ing ou ine
adminis a i e asks, Azu e Heal h Bo is capable o op imising esou ce alloca ion wi hin
heal hca e o ganisa ions. This could ee up hospi al adminis a i e and clinic s a o ocus on
mo e complex o specialised asks, he eby imp o ing ope a ional e iciency. Au oma ion could
enable heal hca e o ganisa ions o educe cos s associa ed wi h adminis a i e asks, such as
swi chboa d call managemen , appoin men scheduling, and s a aining. This could enable
o ganisa ions o achie e cos sa ings while main aining se ice quali y.
Simila ly, Azu e Heal h Bo collec s aluable da a and insigh s om in e ac ions wi h pa ien s
and use s. I ga he s in o ma ion on equen ly asked ques ions, se ice inqui ies, use
p e e ences, and eedback (Bi an, 2023). This in o ma ion can be use ul o heal h manage s,
policymake s and disease su eillance o ice s o sa egua d popula ion heal h.
Mo eo e , Azu e AI Heal h Bo can seamlessly in eg a e wi h exis ing as well as upscaled
heal hca e sys em in as uc u e, elec onic heal h eco ds (EHRs), pa ien managemen sys ems,
and heal h in o ma ion communica ion channels (Mic oso , 2024). As i consolida es
in o ma ion om mul iple sou ces, s eamlines wo k lows, and enhances da a in e ope abili y
ac oss he o ganisa ion, a well-s eng hened heal h sys em can be enhanced (Mo e -Ta ay,
Radawski, & Gua iglia, 2022). Besides, i p io i ises compliance wi h heal hca e egula ions and
indus y s anda ds, such as he Heal h Insu ance Po abili y and Accoun abili y Ac (HIPAA) in
he Uni ed S a es and he Gene al Da a P o ec ion and Regula ion (GDPR) in he Eu opean
Union (Mic oso , 2024).
36
Th ough his compliance, Azu e add esses e hical and o he ea s ha use s may ha e, as i
ensu es da a secu i y, p i acy, and con iden iali y, sa egua ding sensi i e in o ma ion sha ed
du ing in e ac ions wi h pa ien s and use s.
Challenges Associa ed wi h Azu e AI Heal h Bo
While Azu e Heal h Bo has a wide ange o bene i s as i p o ides se e al buil -in
unc ionali ies and p e-buil empla es ailo ed o he heal h sec o , con igu ing and se ing up
Azu e Heal h Bo may equi e ad anced echnical expe ise and amilia i y wi h Azu e se ices
(Mic oso , 2024). This can be e y expensi e in he long e m. Besides, heal hca e o ganisa ions
wi hou dedica ed echnical esou ces may ind i challenging o deploy and manage he bo
e ec i ely.
On he o he hand, In eg a ing Azu e Heal h Bo wi h exis ing heal hca e sys ems, such as EHRs
o heal h managemen sys ems, could be complex, and may equi e addi ional de elopmen
e o as special de elope s a e equi ed (Mic oso , 2024). Acco ding o Mic oso (2024),
Azu e Heal h Bo o e s a pay-as-you-go p icing model. Howe e , cos s can escala e based on
usage and he scale o deploymen (Mic oso , 2024). Hence, he e is a high p obabili y o
incu ing exo bi an cha ges. As such, heal hca e uni s need o ca e ully moni o usage and
conside po en ial cos implica ions, especially o la ge-scale deploymen s.
Ano he Azu e AI Bo challenge is he po en ial dependency on he Azu e Ecosys em (Epic,
2023; Bi an, 2023). Azu e Heal h Bo is closely in eg a ed wi h he Mic oso Azu e ecosys em.
In his case, heal hca e o ganisa ions al eady in es ed in al e na i e cloud pla o ms o
in as uc u e may ace challenges in adop ing Azu e Heal h Bo due o dependencies on
Azu e-speci ic se ices and echnologies (T us Radius, 2024). This could c ea e a monopoly
capi al o Mic oso , which has he po en ial o ende heal h uni s ulne able. Apa om ha ,
in o ma ion asymme y be ween he echnology gian , he ac ual Gene a i e AI, and adop e s
aises conce ns o mis ep esen a ion, in ended o unin ended (Ba haee, 2018).
37

Al hough Mic oso (2024) claims compliance wi h HIPAA and GDPR, p o iding checks and
balances o moni o i s con inued compliance could be a challenge due o he black-box na u e o
AI as well as he complexi y o he big da a ha GenAI p ocesses (Dash e al. 2019). Also,
aining and main aining he Azu e Heal h Bo equi es an ongoing, almos unb eakable,
commi men , as heal hca e o ganisa ions need o alloca e esou ces o aining he bo and
upda ing i s knowledge base egula ly (Mic oso , 2023).
(ii) Gene a i e AI App Builde : Applica ions and Bene i s
Gene a i e AI Applica ion Builde (Gen AI App Builde ) necessi a es he de elopmen , as
expe imen a ion, and deploymen o Gene a i e AI applica ions in heal hca e wi hou he need
o deep expe ience in AI. No wi hs anding, Gen AI App Builde is no a s and-alone GenAI
applica ion, a he i acili a es he building o no el and ailo ed GenAI applica ions. Examples
include Amazon Web Se ices App Builde and Google Cloud App Builde .
Gen AI App Builde could be u ilised o build ailo ed GenAI applica ions o he managemen
o heal h se ices. I has al eady s a ed making s ides as big heal hca e cen es like he Mayo
Clinic ha e adop ed he Google App Builde swi ly (Landi, 2023). I is capable o p ep ocessing
da a, model selec ion and aining, deploymen , as well as ongoing moni o ing. In da a
p ep ocessing, Py hon lib a ies (like Pandas, and NumPy), SQL da abases, and Ex ac ,
T ans o m, Load (ETL) can be used o ga he and p ep ocess heal hca e da a, which may include
Elec onic Heal h Reco ds (EHRs), s a ing needs, and in o ma ion abou he o e all heal h
sys em (Heal hca e Finance News, n.d.). Subsequen ly, du ing model de elopmen , Gen App
Builde can help in selec ing app op ia e as well as ailo ed gene a i e AI models based on he
speci ic applica ion equi emen s (such as GANs, VAEs, and au o eg essi e models) (Law on,
2023; Choudha y e al. 2022). I can ain he models using p ep ocessed da a o lea n pa e ns,
gene a e syn he ic da a, o pe o m o he gene a i e asks ele an o heal hca e managemen . A
his s age o ailo ed GenAI de elopmen , ools like Tenso Flow, PyTo ch, Ke as, OpenAI's GPT,
and Hugging Face's T ans o me s could be o signi ican ele ance.
38
No only ha , Gen App Builde could also be use ul du ing he deploymen phase o he model.
Fo ins ance, i can deploy ained models o heal h managemen as web se ices o as an
Applica ion P og amming In e ace (API) o make hem accessible o heal hca e manage s and
o he p o essionals, also e e ed o as end-use s (Heal hca e Finance News, n.d.). Besides, Gen
AI App Builde s may acili a e compliance wi h da a p i acy egula ions such as HIPAA o
GDPR and could help in implemen ing secu i y measu es o p o ec sensi i e heal hca e da a
(Wald on, 2023). Cloud pla o ms (such as Google Cloud o Mic oso Azu e), Fas API, and
Docke could assis in ensu ing he e ec i e deploymen o he model (T us Radius, 2024).
Addi ionally, Gen AI App Builde could be i al in in eg a ing Gene a i e AI applica ions wi h
exis ing heal h managemen sys ems, such as Elec onic Heal h Reco d (EHR) sys ems o heal h
in o ma ion su eillance sys ems (Epic, 2023). In his way, a seamless da a exchange and
in e ope abili y pla o m could be c ea ed o acili a e adop ion and usabili y. Fo example, he
App Builde could in eg a e GenAI wi h Fas Heal hca e In e ope abili y Resou ces (FHIR),
Heal h Le el Se en In e na ional (HL-7), and Subs i u able Medical Applica ions and Reusable
Technologies (SMART on FHIR) allowing heal h manage s o communica e, exchange
knowledge and adminis e du ies ac oss he en i e heal h sys em and wi h ex e nal s akeholde s
seamlessly (Lekkala, 2023).
Las ly, he Gen App Builde may acili a e implemen a ion, moni o ing and logging mechanisms.
This is pi o al when acking model pe o mance, usage pa e ns, and po en ial issues in
eal- ime. And i has he po en ial o con inuously upda e and e ain models as new da a
becomes a ailable o as heal hca e managemen equi emen s e ol e. Fo example, he GenAI
App Builde can in eg a e wi h Elas icsea ch, Logs ash Kibana (ELK s ack) o G a ana o
moni o heal h adminis a i e p ocesses in eal ime. Hence i can ale po en ial heal h sys em
cybe secu i y h ea s and abuse (ENISA, 2023).
39
Challenges associa ed wi h he Gene a i e AI App Builde
In spi e o se e al applica ions and bene i s ha he GenAI App Builde o e s, i has i s
d awbacks as well. Fo ins ance, i demands special de elope s wi h adequa e knowledge o
heal h sys ems. O e ime, de elope s may become elian on he ea u es and capabili ies
p o ided by he gene a i e AI app builde pla o m, he eby es ic ing hei abili y o in eg a e
ex e nal lib a ies, amewo ks, o ad anced unc ionali ies which may be mo e sui able.
Fu he mo e, depending on he Gene a i e AI App Builde pla o m chosen, GenAI de elope s
may ace endo lock-in, making i di icul o mig a e o al e na i e pla o ms o cus omise he
applica ion beyond he pla o m's o e ings (Opa a-Ma ins, Sahandi & Tian, 2014).
Mo eo e , GenAI App Builde s may in ol e subsc ip ion ees, usage-based p icing models, o
addi ional cha ges o p emium ea u es (Mic oso , 2024; Google Resea ch, 2024). This can
con ibu e o ongoing cos s o de elopmen and deploymen , u he s aining he heal h sys em
which al eady ope a es wi h sca ce esou ces. Also, de elope s could ace e aining challenges
when using a speci ic Gene a i e AI App Builde , pa icula ly i hey a e un amilia wi h he
pla o m's in e ace, wo k low, o unde lying echnologies. This can impac quali y and he e o e
he en i e heal h ecosys em. Fo ins ance, a de elope whose backg ound is based on Amazon
Web Se ices App Builde may ind i challenging o wo k on Google Cloud App Builde . This
can inc ease u na ound ime, oppo uni y cos s and unin ended consequences.
(iii) O he No able GenAI Tailo ed o he Heal h Sec o
To begin wi h, Ga o T on. I is a ype o NLP model (F ies e al. 2022). Also e e ed o as a
“la ge clinical language model” ained using o e 90 billion wo ds o ex , including o e 82
billion wo ds o de-iden i ied clinical ex (Yang e al. 2022). I is by a he la ges clinical LLM.
Acco ding o he Uni e si y o Flo ida News (2023), in es iga o s ained supe compu e s o
gene a e heal h- ela ed eco ds based on a no el model, Ga o T onGPT, ha unc ions simila ly
o Cha GPT. Despi e being expensi e o ain, Ga o T on can easily adap o new asks, making i
sui able o he e e -e ol ing heal hca e landscape (F ies, 2022).
40
Ga o T on excels in unde s anding and gene a ing human-like ex , hence is aluable o
p ocessing and in e p e ing uns uc u ed da a sou ces in heal hca e, such as clinical no es,
employee eedback, o biomedical li e a u e. In addi ion, Ga o T on can se e as a i ual
assis an , p o iding eal- ime in o ma ion e ie al, clinical decision suppo , and expe
ecommenda ions o clinic manage s, heal hca e p o ide s and policymake s. Besides, hey o e
pe sonalised esponses o que ies, helping heal h pe sonnel o access ele an medical and
logis ical knowledge seamlessly (Uni e si y o Flo ida News, 2023).
Ga o T on can also be in eg a ed in o pa ien - acing applica ions, cha bo s, o educa ional
pla o ms o deli e pe sonalised heal h in o ma ion, answe common ques ions, and p o ide
sel -ca e ecommenda ions (Yang e al. 2022). This could imp o e pa ien engagemen and heal h
li e acy apa om s eamlining adminis a i e asks and enhancing documen a ion quali y and
comple eness. Hence, he alue and quali y o ca e may imp o e o e all.
Figu e 7: An Illus a ion o Ga o T on; An NLP Model In Ac ion
Sou ce: (F ies e al. 2022)
41
3.12 P e- es ing o Resea ch Ins umen s
A p e- es is when a ques ionnai e is es ed on a s a is ically small sample o esponden s be o e
he ac ual s udy. P e- es ing o e s he oppo uni y o see wha ques ions wo k well, wha
ques ions sound s ange, wha ques ions can be elimina ed and wha needs o be added
(Co mack, 2001).
3.13 Reliabili y
Acco ding o Co mack (2001), eliabili y is he abili y o he esea ch s udy o p oduce simila
esul s consis en ly when edone o pe o med by ano he esea che o a e a ce ain pe iod.
The e should be essen ially li le o no di e ence in esul s i he same me hodology is applied o
he same pa icipan s in he u u e as long as he same app oach is used (B yman, 2021).
The esul s om he ques ionnai e adminis e ed as pa o his s udy consis ed o answe s
collec ed o e a o nigh . Since he esponden s we e no gi en any new in o ma ion o a
di e en ques ionnai e o ill ou h oughou his ime, i can be concluded ha his measu e is
dependable.
Howe e , suppose he su ey ques ionnai e was eadminis e ed a yea la e , i is possible ha he
pa icipan s' unde s anding o GenAI sys ems would ha e me amo phosed o e ime. Hence,
hei esponses could be di e en . This is some hing ha migh be con es ed. And indica es ha
he eliabili y o his pa icula esea ch echnique canno be sus ained.
3.14 Replicabili y
Acco ding o B yman (2021), eliabili y is s ongly linked o eplicabili y. Replicabili y is when
esea ch echniques employed in one s udy can be ep oduced o pe o m a di e en
in es iga ion o he same na u e. The e o e, i could be ha d o ano he esea che o ep oduce
he s udy i he i s in es iga o does no explain he s udy p ocedu e adequa ely. In his s udy,
he esea ch me hodology, design and p ocedu e ha e been desc ibed ex ensi ely in a ious
sec ions o his disse a ion.
48

The su ey ques ionnai e used in his s udy has been a ached as an appendix and is made
accessible o o he esea che s who would wan o conduc a simila in es iga ion. Following his
backg ound, he indings o his esea ch may be ega ded as highly dependable and
ep oducible.
3.15 Validi y
Poli & Hungle (1999) de ine alidi y as “ he abili y o he esea ch me hodology o measu e
wha i in ended o measu e.” The e is a consensus among Poli & Hungle (1999), Co mack
(2001), G ay & G o e (2020) and B yman (2021) ha a alid measu emen is p edominan ly
eliable. Tha is; i a es p oduces accu a e esul s, hey should be ep oducible. Al hough a alid
me hodology is always eliable, a eliable me hodology is no always alid. A e his
in es iga ion, i was decided ha he esea ch me hods employed o answe he p ima y esea ch
ques ion (s) we e e ec i e and conclusi e.
3.16 Da a Collec ion P ocedu e
A e ge ing in o med consen om he pa icipan s, da a was collec ed anonymously wi hin a
o nigh using Google Fo ms. The collec ed da a was kep secu ely in an online olde using
access c eden ials managed by he Be lin School o Business and Inno a ion (BSBI), a uni e si y
which complies wi h GDPR guidelines.
3.17 E hical Conside a ions
Pe mission was sough om BSBI’s Examina ions and Assessmen Depa men , a commi ee
ha ensu es academic quali y and e hical conduc .
The pu pose and bene i s o he s udy we e explained o he pa icipan s who ga e in o med
consen . P i acy, anonymi y, con iden iali y and no coe cion we e main ained h oughou he
s udy. The pa icipan s we e no o ced o answe ques ions and we e ee o wi hd aw om he
s udy a any s age wi hou being ic imised.
49
3.18 S udy limi a ions
The s udy was conduc ed o a sho pe iod as he in es iga o was ollowing a p e-planned
academic schedule.
The con enience sampling me hod was used which has an elemen o bias. Mo eo e , he
ins umen which was employed o collec da a was de eloped and used o he i s ime by he
in es iga o who has no expe ience in esea ch. As such, he ins umen may no yield accu a e
and de ailed in o ma ion despi e being p e- es ed o alidi y and eliabili y.
The p ima y in es iga ion is limi ed o Ge many as a geog aphic wo k a ea o he pa icipan s.
Besides, no all key in o man s we e eached, he language was a ba ie since he in es iga o
was no able o communica e wi h some key in o man s in Ge man, he o icial language o he
esea ch se ing.
3.19 Summa y
This chap e add essed he esea ch me hodology, esea ch design, popula ion and sampling.
Fu he mo e, he da a collec ion p ocedu e and esea ch ins umen ha was used when
collec ing da a ha e been p esen ed. On he o he hand, issues ega ding alidi y, eplicabili y,
eliabili y, and e hical conside a ions du ing he s udy we e also add essed.
50
CHAPTER FOUR – FINDINGS / ANALYSIS / DISCUSSION
4.1 FINDINGS
Acco ding o James Madson Uni e si y (n.d.), he indings sec ion o a esea ch pape desc ibes
wha he in es iga o ound a e analysing he da a. The p ima y pu pose o his sec ion is o use
he da a collec ed o answe he esea ch ques ions posed in he in oduc o y chap e , e en i he
indings challenge he ini ial hypo hesis.
This chap e ocuses on he p esen a ion o indings, he analysis o da a ob ained om he
esea ch pa icipan s as well as he discussion o such indings. Da a is p esen ed in he o m o
equency ables, ba g aphs and pie cha s. The s udy indings a e sys ema ically a anged in o
ou sec ions in ela ion o he esea ch objec i es. The sec ions a e as ollows; Demog aphic
Da a & Knowledge Assessmen , Gene a i e AI Applica ions and hei Bene i s in Heal h
Sys ems, Challenges and E hical Implica ions o Gene a i e AI in Heal hca e Managemen , and
he sec ion abou Fu u e Di ec ion o GenAI in Heal h Sys ems.
51
Sec ion A: Demog aphic Da a & Knowledge Assessmen
Figu e 8: Age Rep esen a ion o he Pa icipan s: N = 35 Pa icipan s, all esponded.
Table 3: Measu es o Cen al Tendency o Age o Pa icipan s
Mean
27 yea s
Mode
26-35 yea s
Median
26-35 yea s
Range
17 Yea s
Figu e 9: Gende o Pa icipan s: N = 35, o whom all esponded
52
Figu e 10 Composi ion o Heal h Wo ke s in he S udy
Figu e 11: Knowledge Assessmen : N = 35, all esponded
53

Figu e 12: U ilisa ion o GenAI Applica ions: N = 35, o whom all espond
Table 4: Wha is Gene a i e AI? N = 35, all esponded
Response
F equency
Pe cen age (%)
AI which can p oduce ex , images, ideo o o he
ma e ial
17
48.6%
AI which esponds o human p omp s and
p oduces new and o iginal con en
22
62.9%
AI which can gene a e elec ici y
1
2.9%
I don' know
7
20%
54
Figu e 13: Familia i y wi h Gene a i e AI: N = 35, o whom 32 esponded
Sec ion B: Gene a i e AI Applica ions and hei bene i s in Heal h Sys ems
Figu e 14: Bene i s o GenAI: N = 35, 28 o whom esponded
55
Figu e 15: GenAI in Heal hca e: N = 35, all esponded
Figu e 16: GenAI Use Case, and By Whom: N = 35, 34 o whom esponded
56
Sec ion C: Challenges and E hical Implica ions o Gene a i e AI in Heal h Sys ems
Figu e 17: E hical Conce ns o GenAI: N = 35, all esponded
Figu e 18: Cos -e ec i eness o GenAI: N = 35, all esponded
57
Sec ion C: Challenges and E hical Implica ions o Gene a i e AI in Heal h Sys ems
In Figu e 17 abo e, 35 pa icipan s we e asked abou hei ea s ega ding he in eg a ion o
GenAI in heal h sys ems. 15 pa icipan s (42.9%) indica ed p i acy conce ns, 10 esponden s
(28.6%) pinpoin ed us issues, and 25 pa icipan s (71.4%) highligh ed ha GenAI may lead o
o e - eliance and au oma ion bias, which may comp omise quali y o ca e. Howe e , 5
esponden s (14.3%) did no ha e any ea s; while 1 pa icipan epo ed o he ea s. In he
u he commen s sec ion p o ided, he o he ea s we e highligh ed as “GenAI may lead o loss
o job and quali y issues.”
Figu e 18 abo e illus a es ha 18 pa icipan s (51.4%) pe cei ed inco po a ing GenAI in heal h
managemen as cos ly. While 12 pa icipan s we e no su e abou he expenses. On he o he
hand, 5 pa icipan s (14.3%) pe cei ed GenAI as cos -e ec i e.
Sec ion D: Fu u e Di ec ion o GenAI in Heal hca e Managemen
Ou o 35 esponden s su eyed, 22 (62.9%) epo ed op imis ic p ospec s o GenAI in
heal hca e managemen , 9 (25.7%) had no idea abou how GenAI will impac heal h
managemen in he u u e. 4 pa icipan s (11.4%) we e pessimis ic abou he u u e bene i s o
GenAI in heal h managemen .
Table 5 abo e shows ha ou o 15 esponden s, 7 pa icipan s we e awa e o Azu e AI Heal h
Bo , 2 knew Gene a i e AI App Builde , 6 knew abou Med PaLM, ano he 6 demons a ed
amilia i y wi h BioGPT, 3 knew abou ClinicalBERT while only 1 knew Ga o T on. All
ep esen ing 48.7%, 13.3%, 40%, 40%, 20%, and 6.7% o he esponden s pe each ca ego y
espec i ely.
In Figu e 20, when asked how soon hey expec o wi ness GenAI ac i ely inco po a ed in he
heal h sys ems, only 1 (3.2%) indica ed wi hin 1 yea , 5 (16.1%) indica ed 2 yea s, 9 (29%)
expec ed GenAI in 3 yea s while he o he 5 (16.1%) hough 4 yea s was ealis ic. Howe e , 11
(35.5) pa icipan s ep esen ing he majo i y o answe s we e mo e pessimis ic and expec ed o
see GenAI being in eg a ed in o heal h sys ems in 5 o mo e yea s o come.
64

Figu e 21 abo e shows ha 5 esponden s we e scep ical abou he u u e in eg a ion o GenAI in
heal h sys ems, and 15 pa icipan s we e posi i e ha soone o la e GenAI will be in eg a ed in
heal h sys ems. Mo eo e , 14 pa icipan s we e no su e.
4.2.4 Summa y
This sec ion has highligh ed he meaning o da a analysis and di e en ways o analysing da a in
esea ch, and how he collec ed da a o his s udy was analysed.
65
4.3 DISCUSSION
This sec ion aims a in e p e ing and desc ibing he signi icance o he esea ch indings in ligh
o wha was al eady known abou Gene a i e AI in heal hca e, and enables a ho ough
explana ion o any new unde s anding o insigh s abou he esea ch opic a e he indings and
he analysis ha e been aken in o conside a ion (Co mack, 2001). This sec ion is w i en
ollowing a di ide-and-conque ac ic o app oaching a manusc ip adap ed om Şanlı, E dem &
Te ik (2013). The ou come o his disse a ion has shed ligh on he possible applica ion o
GenAI in heal h sys ems by e iewing GenAI inno a ions and hei implica ions. Howe e , he
indings ough o be in e p e ed ca e ully due o he limi a ions o his s udy.
This disse a ion in es iga ed he applica ion o GenAI in heal h sys ems. As discussed in he
p eceding chap e s, GenAI has he po en ial o ans o m how he heal h sec o ope a es. And
se e al s udies ha e been done abou how GenAI will impac he heal hca e indus y. Howe e ,
p e ious s udies ocused on he hospi al and clinical p ocesses wi h limi ed conside a ion o he
heal h sys ems as a whole, hence, his s udy aims o ill ha gap and make a con ibu ion o
u he esea ch. The hi d chap e is composed o a de ailed explana ion o he me hodology and
da a collec ion p ocedu e. Validi y, eliabili y and eplicabili y issues we e also discussed,
including he limi a ions o his disse a ion. This chap e summa ises key indings h ough he
in e p e a ions o he in es iga o ega ding he signi icance o he s udy. Besides, he
implica ions o his s udy a e also aken in o accoun h ough he lens o he in es iga o by
ocusing on he con ibu ion o his s udy o wha was al eady known. Following his chap e is
he i h chap e also e e ed o as “conclusion.” I p esen s a summa y o he indings discussed
and conclusions d awn. This is he inal sec ion o he disse a ion.
4.3.1 Demog aphic Da a
In his s udy, 35 heal h wo ke s in managemen o adminis a i e oles pa icipa ed, o whom 23
we e emale, ep esen ing 65.7% o esponden s in he s udy. On he o he hand, 12 we e males
ep esen ing 34.3% o pa icipan s in he s udy. These s a is ics con i m S a is a's (2023) indings
ha in Ge many, mo e emales occupy leade ship posi ions, especially in heal hca e which
accoun s o 36.7% o all women in manage ial oles.
66
The e o e, an open and inclusi e accep ance and applica ion o GenAI in heal h sys ems could
ensu e success ul implemen a ion. This u he indica es ha gende s e eo ypes in he
echnology ield can only de ail he e ec i e adop ion o Gene a i e AI which is swi ly
changing he de ini ion o wo k, as bo h men and women a e d i e s o change, e en, women
a guably, ha e mo e o con ibu e as shown by he s a is ics abo e.
The mean age o pa icipan s was 27 yea s, he mode was 26 - 35 yea s, and he median was also
26 - 35 yea s. The age ange was 17 yea s. This shows ha young adul s a e occupying
leade ship oles, and a e shaping he pe cep ion and eali y o he wo kplace. S a is isches
Bundesam (2024) concu s wi h hese indings by s a ing ha young pe sons a e becoming
highly quali ied and a e assuming leade ship oles in he wo kplace. The numbe o quali ied
young people in Ge many in c eased by 10.8 pe cen om 2002 o 2022 (S a is isches
Bundesam , 2024). In ligh o his, young people, when gi en adequa e oppo uni ies o h i e in
b idging he echnology accep ance gap in he heal h sec o , mo e can be achie ed. Simila ly,
young people a e be e quali ied in echnical pu sui s, especially in he u ilisa ion, de elopmen ,
and e alua ion o A i icial In elligence.
9 (25.7%) o heal h wo ke s su eyed we e in nu sing adminis a ion, 6 (17.1%) we e in heal h
educa ion, 5 (14.3%) we e heal h manage s, 5 (14.3%) we e heal h esea che s, 4 (11.4%) we e
in medical adminis a ion, 3 (8.6%) we e in gene al public heal h, and 3 (8.6%) we e clinic o
hospi al adminis a o s. These esul s sugges ha he nu sing ield has mo e heal h wo ke s
seconded by heal h educa ion, hen heal h managemen and heal h esea ch. Gene al public
heal h and hospi al o clinic adminis a ion we e he leas ep esen ed. The e o e, when
conside ing he adop ion o GenAI in heal h sys ems, i is i al o conside he alloca ion o
echnological esou ces and upskilling equi ably. Tha is; acco ding o need.
67
Knowledge o Gene a i e AI
Among he 35 heal h wo ke s su eyed, 17 we e amilia wi h Gene a i e A i icial In elligence
while 18 we e no . This indica es a knowledge gap ega ding ad ancemen s in echnology and
he u u e o heal hca e. Fu he mo e, when asked o name he Gene a i e AI hey know, a
whopping 90.6% o esponden s indica ed amilia i y wi h Cha GPT. Tha is; hey we e no
awa e o a leas less amilia wi h o he Gene a i e AI applica ions like Mu AI. Besides, 80%
o he pa icipan s epo ed ha hey ha e ne e used Gene a i e AI applica ions a wo k; while
20% indica ed ha hey ha e a leas used he GenAI echnology a wo k. These insigh s e eal a
lack o awa eness and echnological li e acy among heal hca e wo ke s. Mo eo e , knowledge o
GenAI applica ions ailo ed o he heal h sec o was also e y limi ed, p o ing ha e en i he
applica ion o GenAI is o be adop ed in heal h sys ems, i s impac s migh no be ealised ully
due o incompe en abili ies o lack o eadiness o ealise i .
4.3.2 Gene a i e AI Applica ions and hei Bene i s in Heal hca e Sys ems
28 esponden s amilia wi h GenAI applica ions we e asked o men ion he ad an ages o
GenAI. 20 (71.4%) indica ed ha GenAI helps hem o do asks quickly and p o ides use ul
insigh s and answe s. Such indings highligh he impo ance o GenAI in ask op imisa ion and
wo k low imp o emen . This ag ees wi h McKinsey & Company's (2023) and Bain &
Company's (2023) indings ha GenAI could imp o e he adminis a i e p ocesses and can
op imise o e all wo k low. Ne e heless, only 13 (46.4%) esponden s no ed ha GenAI
imp o es ask e iciency. This sugges s ha despi e i s bene icial impac s, use s a e s ill scep ical
abou he us wo hiness and accu acy o GenAI. Simila ly, 14 (50%) epo ed ha GenAI is
help ul in p oblem-sol ing; implying ha he o he 50% did no ega d GenAI as a good ool o
decision-making. These indings e u e he claims o A anade (2024) and Dash e al. (2019) ha
GenAI could imp o e decision-making. Human beings a e good a decision-making because o
he abili y o exe cise emo ions, mo al judgemen and empa hy. As such, GenAI ad ancemen s
can only be e ec i e wi h humans-in- he-loop.
68
Despi e ha , 30 esponden s ep esen ing 85.7% we e op imis ic abou in eg a ing GenAI in
heal h sys ems. On he o he hand, 5 esponden s ep esen ing 14.3% we e scep ical. This
indica es ha heal h wo ke s a e eady o emb ace Gene a i e AI wi h cos s. Fo ins ance, ou o
34 pa icipan s, 20 (58.8%) epo ed ha GenAI should be an agenda o heal h manage s.
Whe eas 14 (41.2%) pa icipan s indica ed ha heal h manage s ha e no hing o do wi h GenAI,
hence, GenAI should no be inco po a ed in heal h sys ems. These indings demons a e a gap
and a lack o consensus among heal h wo ke s as ega ds whe he o no GenAI should be
applied in heal h sys ems.
4.3.3 Challenges and E hical Implica ions o Gene a i e AI in Heal hca e Managemen
Ou o 35 esponden s, 15 (42.9%) indica ed p i acy conce ns, 10 (28.6%) pinpoin ed us
issues, 25 (71.4%) highligh ed ha GenAI may lead o o e - eliance and au oma ion bias. On he
con a y, 5 esponden s (14.3%) did no ha e any ea s; while 1 pa icipan epo ed o he ea s.
In he u he commen s sec ion o he ques ionnai e, he o he ea s we e highligh ed as “GenAI
may lead o loss o job and quali y issues.” These indings shed ligh on he se ious e hical
conce ns, bo h held in insically a indi idual and co po a e le els, associa ed wi h he
inco po a ion o GenAI in heal h sys ems. This con i ms he Deu sche Welle's (2024) epo ha
he sys emic design, he s uc u al design and echnological design o a heal h sys em can
exace ba e heal h dispa i ies.
Fo example, ools like Oxime e s ha e been seen calcula ing misleading diagnoses o black
people because hey we e no designed as acially inclusi e ools (Deu sche Welle, 2024).
The e o e, he implemen a ion o GenAI equi es a ho ough and me hodical inclusi i y in
aining da a and acial ep esen a ion. Apa om ha , accoun abili y issues, he ea o AI
becoming supe in elligen , and ques ions o mo ali y and empa hy we e also conce ns.
Mo eo e , o e hal o he pa icipan s (51.4%) ci ed challenges ela ed o a o dabili y. As
Mic oso (2024), one o he GenAI supplie s, no es, GenAI cos s o de eloping, implemen ing
and main aining inc eases wi h he inc ease in usage and o e a long pe iod. In his way, heal h
sys ems planning o o adop ing GenAI a e ine i ably going o incu huge ope a ional and ela ed
cos s. Besides, heal h sys ems also isk being locked o he ecosys em o he GenAI supplie ,
which can be e en mo e cos ly unless clea egula ions a e legally de ined.
69

This endo lock-in could be synonymous wi h holding heal h sys ems hos age. Fu he mo e,
he asymme ic in o ma ion be ween GenAI supplie s, GenAI i sel , and use s u he wo sen he
ea s associa ed wi h inco po a ing GenAI in he heal h sys em.
4.3.4 Fu u e Di ec ion o GenAI in Heal hca e Managemen
22 (62.9%) esponden s epo ed op imis ic p ospec s o GenAI in heal h sys ems, and 9 (25.7%)
had no idea abou how GenAI will impac heal h managemen in he u u e. Meanwhile 4
pa icipan s (11.4%) we e pessimis ic abou he u u e bene i s o GenAI in heal h sys ems.
Al hough ea s and unce ain ies pe sis , hese indings highligh he eadiness o he heal h
wo k o ce as ega ds he inco po a ion o GenAI in heal h sys ems. Howe e , he e is a lack o
consis en consensus. The e o e, as ega ds he u u e, he adop ion o GenAI in heal h sys ems
migh enable human-machine collabo a ion and accele a e he ansi ion om Indus y 4.0 (4IR)
o Indus y 5.0 (5IR). No ably, a he momen 5IR emains a u opia.
Fu he mo e, 15 esponden s demons a ed a lack o awa eness o ending GenAI applica ions
wi h he po en ial o impac he u u e o heal h sys ems. Fo example, only 7 pa icipan s knew
Azu e AI Heal h Bo , 2 knew Gene a i e AI App Builde , 6 knew abou Med PaLM, ano he 6
demons a ed amilia i y wi h BioGPT, 3 knew abou ClinicalBERT while only 1 knew
Ga o T on.
This shows ha GenAI may ake heal h wo ke s by s o m i hey a e no in en ional enough o
ge o know and assess hei applicabili y in heal h sys ems. GenAI is ine i ably ans o ming
how managemen indus ies in e ac wi h employees. The ques ion is, “A e heal h sys ems
spa ed o is he e a need o a p oac i e assessmen o hese ans o ma i e echnologies?”
4.3.5 Summa y
This chap e has p esen ed he indings o he s udy opic, "Applica ion o Gene a i e AI in
Heal h Sys ems: A Comp ehensi e Re iew o Inno a ions and Implica ions.” The indings we e
p esen ed in g aphical, abula and cha o ma s.
70
In addi ion, i has also highligh ed how he collec ed da a was analysed. Las ly, he indings
ha e been discussed wi h e e ence o he e iewed li e a u e and he analysed da a om p ima y
esea ch. Simila o de ia ing indings om o he s udies and he obse a ions and
in e p e a ions o he in es iga o we e jus i ied.
71
CONCLUDING REMARKS
“The e olu ion o Gene a i e A i icial In elligence has unlocked an un apped U$1 illion
imp o emen po en ial in he heal hca e indus y.” Globally, he e seems o be a pa adigm shi
ega ding he u u e o heal h sys ems, and Ge many is no spa ed, especially wi h he elease o
a Gene a i e AI known as Cha GPT in 2022. Subsequen ly, signi ican GenAI de elopmen s
ha e occu ed including hose ailo ed o he heal h sec o . The likelihood ha GenAI oge he
wi h o he echnological ad ancemen s will expand apace and become a i al equi emen
globally is ce ain. The e is a g owing ce ain y ha people and obo s may collabo a e on
speci ic adminis a i e asks. This is pa icula ly ue wi h a ansposi ion o Indus y 4.0 (4IR),
which began in 2010 and Indus y 5.0 (5IR), which s a ed ui ion in 2020. Based on he
in es iga o ’s seconda y and p ima y esea ch, i seems like Gene a i e AI’s exis ence in oday’s
and u u e wo kplaces is una oidable.
Wi h e e ence o he Technology Accep ance Model which p o ides a heo e ical amewo k o
unde s anding and p edic ing use s' accep ance and adop ion o no el echnologies, his s udy
has e iewed li e a u e ela ed o GenAI. The e iew concu ed wi h TAM ha use accep ance
o echnology hinges on ou ac o s, namely: pe cei ed use ulness, pe cei ed ease o use, social
in luence, and cogni i e ins umen al p ocesses. These ha e been ansla ed h oughou his
s udy as di e en ca ego ies o GenAI and hei applicabili y in heal h sys ems, bene i s o
inco po a ing GenAI in heal h sys ems, challenges and e hical conce ns o such inno a ion as
well as p esen and u u e implica ions espec i ely. Gene a i e AI inno a ions could ca alyse
he ansi ion om 4IR o 5IR wi h he aid o human-machine collabo a ion, helping s eamline
heal h sys em p ocesses, and acili a ing he imp o emen o e iciency a he wo kplace.
Whe eas he implica ions ha e been iden i ied as e hical conce ns ela ed o bias, mo ali y,
accoun abili y, empa hy and he ea o GenAI and associa ed echnology becoming supe
in elligen . Mo eo e , implica ions abou unsus ainable cos s, endo lock-in and main enance
demands ha e been e ealed. No wi hs anding, he asymme ic in o ma ion among GenAI
de elope s, Gene a i e AI i sel , and he end use s.
72
The con ibu ions o his s udy a e: (1) To p o ide a basic unde s anding o gene a i e a i icial
in elligence and e eal he po en ial alue o GenAI in suppo ing heal hca e sys ems,
no wi hs anding he d awbacks ha come wi h such inno a ion. (2) To b ing o a en ion and
discuss he la es s udies ela ed o he applica ion o GenAI in heal h sys ems, ha p o ide a
backg ound highligh ing he bene i s and challenges o GenAI and po en ial u u e di ec ion o
u he esea ch (3) To o eshadow an image o a u u e heal h sys em and he po en ial
ede ini ion o he ole o he u u e heal h manage , including he skills ha should be acqui ed
and insis on he ac ha he e will always be a need o heal h manage s, and human skills will
possibly be mo e aluable in he u u e.
As Gene a i e AI con inues o de elop, he echnology is becoming mo e sophis ica ed wi h
ime. The e migh be implica ions associa ed wi h such ad ancemen s. The e o e, he
in es iga o p o ides 3 ecommenda ions as ollows: To s a wi h, he e is a need o ensu e
heal h wo ke s, especially manage s and adminis a o s, ha e access o domain expe ise o a
leas ha e p io knowledge o on- he-job aining o make sense o GenAI ad ancemen s.
Secondly, he c ea ion o decisi e and adap able egula ions and ules ha mus be applied o he
da ase s o gene a e use ul insigh s wi hou comp omising he heal hca e’s undamen al
p inciples o “bene icence” and “nonmale icence.” Tha is, ules and egula ions o sa egua d
use s om biassed AI solu ions. Ques ions like, “Who will police he GenAI when a b each o
GDPR, accoun abili y o e hical misconduc is obse ed?” and “How will he issues o us ,
mo ali y and empa hy be add essed?” should be conside ed.
Finally, inco po a ing Gene a i e AI in heal h sys ems may come wi h huge in es men and
oppo uni y cos s bo h mone a ily and e hically. As such, a ho ough cos -bene i analysis mus
be done bo h quali a i ely (how in eg a ing GenAI will a ec heal h wo ke s and pa ien s alike)
and quan i a i ely (how in eg a ing GenAI will impac cos s and e u n on in es men (ROI))
be o e, du ing and a e he in eg a ion o GenAI in o heal h sys ems.
73
Landi, H. (2023) Mayo Clinic, Google Cloud pa ne on Gene a i e AI. Accessed om
Fie ceHeal hca e on 20.Jan.24
Lainjo, B. (2023) The Global Social Dynamics and Inequali ies o A i icial In elligence.
Accessed om DOI:Resea chGa e on 02.Feb.24
L , Z. (2023) Gene a i e A i icial In elligence in he Me a e se E a. Accessed om Cogni i e
Robo ics on 02.Jan.24
Mascha, E. & Ve e , T. (2018) Signi icance, E o s, Powe , and Sample Size: The Blocking and
Tackling o S a is ics. Accessed om PubMed on 12.Feb.24
McKinsey & Company (2023) Tackling Heal hca e’s Bigges Bu dens wi h Gene a i e AI.
[online] Accessed om McKinsey & Company on 11-Jan-24
Mahmud, M., Kaise , M., McGinni y, T.M. e al. (2021) Deep Lea ning in Mining Biological
Da a. Accessed om Sp inge on 02.Feb.24
MedNews (2019) Consume s Don’ Fully T us Heal hca e Cha bo s; S udy inds. Heal h IT,
A i icial In elligence. Accessed om Heal h IT on 03. Feb.24
Mille , K. (2022) Deploying AI in Heal hca e: Sepa a ing he Hype om he Help ul. Accessed
om h ps://hai.s an o d.edu/news/deploying-ai-heal hca e-sepa a ing-hype-help ul on 12.Feb.24
McKinsey & Company (2022) Ge many’s E-heal h T ans o ma ion Makes Une en P og ess.
Accessed om McKinsey & Company on 17.Feb.24
Mülle , J. (2020). Enabling Technologies o Indus y 5.0—Resul s o a Wo kshop wi h Eu ope’s
Technology Leade s. Di ec o a e-Gene al o Resea ch and Inno a ion. Accessed om
Eu opean Commission on 05.Jan.24
MSG Su ey (2023) O e 80% o Companies in Ge many A e Planning Mo e In ensi e Use o
Gene a i e AI. Accessed om MSG G oup on 10.Jan.24
Meskó, B. & Topol, E. (2023) The Impe a i e o Regula o y O e sigh o Gene a i e AI in
Heal hca e. Accessed om Na u e Digi al Medicine Jou nal on 14.Feb.24
80

Mo e -Ta ay, C., Radawski, H. & Gua iglia, C. (2022). Heal h P o essionals' Expe ience Using
an Azu e Voice-Bo . Accessed om MDPI Jou nal 7.Feb.24
Mic oso (2024) Azu e Heal h Bo : Pay-As-You Go. [online] Accessed om Mic oso Azu e on
07.Feb.24
Nah, F., Zheng, R., Cai, J. e al. (2023) Gene a i e AI and Cha GPT: Applica ions, Challenges,
and AI-Human Collabo a ion, Jou nal o In o ma ion Technology Case and Applica ion
Resea ch. Accessed om In o ma ion Technology on 13.Jan.24
New on, W. (2023) Wha is BioGPT and Wha Does i Mean o Heal hca e? Clinical T ials
A ena. [online] Accessed om Clinical T ials A ena on 07.Feb.24
Omiye, J., Les e , J., Spichak, S. e al. (2023) La ge Language Models p opaga e Race-based
Medicine.Accessed om Digi al Medicine on 03.Feb.24
Opa a-Ma ins, J., Sahandi, R. & Tian, F. (2014) A C i ical Re iew o Vendo Lock-in and I s
Impac on Adop ion o Cloud Compu ing. Bou nemou h Uni e si y, Uni ed Kingdom. Accessed
om Bou nemou h Uni e si y on 07.Feb.24
Oniani, D., Hilsman, J., Peng, Y. e al (2023) Adop ing and Expanding E hical P inciples o
Gene a i e A i icial In elligence om Mili a y o Heal hca e. Accessed om Digi al Medicine
on 20.Jan.24
OpenAI (2024) SORA Tex -To-Video: The Age o AI is He e. Accessed om
h ps://www.you ube.com/wa ch? =HK6y8DAPN_0 on 17.Feb.24
Poli , D. & Hungle , B. (1999) Nu sing Resea ch: P inciples and Me hods. ISBN-10:
0781715636
P a im, P. (2023) Cha GPT: A Comp ehensi e Re iew on Backg ound, Applica ions, Key
Challenges, Bias, E hics, Limi a ions and Fu u e scope. In e ne o Things and Cybe -Physical
Sys ems. Accessed om IoT on 21.Jan.24
Peng, C. e al. (2023) A s udy o Gene a i e La ge Language Model o Medical Resea ch and
Heal hca e. Accessed om Digi al Medicine on 12.Feb.24
81
Panch, T., Ma ie, H., & A un, R. (2019). A i icial In elligence & Algo i hmic Bias:
Implica ions o Heal h Sys ems. Accessed om Jou nal o Global Heal h on 20.Jan.2024
Qingxiong, M. & Liping, L. (2006). Technology Accep ance Model. Jou nal o O ganisa ional
and End Use Compu ing. Accessed om Resea chGa e on 6.Jan.24
Singh, A. & Ogun unmi, T. (2021). An O e iew o Va ia ional Au oencode s o Sou ce
Sepa a ion, Finance, and Bio-Signal Applica ions. Accessed om MDPI Jou nal on 27.Jan.24
Sallam M. (2023). Cha GPT U ili y in Heal hca e Educa ion. Sys ema ic Re iew on he
P omising Pe spec i es and Valid Conce ns. Accessed om MDPI Jou nal on 19.Jan.24
She s insky, A. (2020) Fundamen als o Recu en Neu al Ne wo k (RNN) and Long Sho -Te m
Memo y (LSTM) ne wo k. Accessed om Science Di ec on 27.Jan.24
Shok ollah, Y., Ya mohammad oosky, S., Nikahd, M. e al. (2023) A Comp ehensi e Re iew o
Gene a i e AI in Heal hca e. Accessed om Co nell Uni e si y on 30.Jan.24
Sage Resea ch Me hods (n.d.) Unde s anding Resea ch. Accessed om Sage Publica ions on
17.Feb.24
Sezgin E. (2023) A i icial In elligence in Heal hca e: Complemen ing, no Replacing, Doc o s
and Heal hca e P o ide s. Accessed om Sage Publica ions on 14.Feb.24
Shahbakhsh, M. & Hassapou , H. (2023) Empowe ing Face Recogni ion Me hods using a
GAN-based Single Image Supe -Resolu ion Ne wo k. Accessed om Jou nal o Enginee ing on
24.Jan.24
Singh (2023) Wha is Resea ch Design? Unde s anding Types o Resea ch Design. [online]
Accessed om Resea che Li e on 13.Feb.24
S a is isches Bundesam (2024) Quali y Employmen . [online] Accessed om Des a is.de on
20.Feb.24
S a is a (2023) Sha e o Women in Leade ship Posi ions in Ge many in 2023, by Indus y.
[online] Accessed om S a is a on 20.Feb.24
82
Sage Publica ions (n.d.) Li e a u e Re iew and Focusing he Resea ch. Accessed om Sage
Publica ions on 04.Feb.24
Tellhed, U., Bjö klund, F. & Kallio-S and, K. (2023) Men and Ca ing Women: Gende
S e eo ypes in Tech-Educa ion. Sex Roles. Accessed om Sp inge on 20.Feb.24
Tu chin, A., Masha sky, S. & Zi nik, M. (2023) Compa ison o BERT Implemen a ions o NLP
o Na a i e Medical Documen s. Accessed om Science Di ec on 07.Feb.24
Tahe doos , H. (2020) Di e en Types o Da a Analysis; Da a Analysis Me hods and Techniques.
In e na ional Jou nal o Academic Resea ch in Managemen . Accessed om HAL Science on
19.Feb.24
T us Radius (2024) Azu e Heal h: Compe i o s and Al e na i es. [online] Accessed om
T us Radius on 09.Feb.24
Uni e si y o Flo ida News (2023) Medical AI ool om UF, NVIDIA ge s Human Thumbs-up.
[online] Accessed om Uni e si y o Flo ida on 07.Feb.24
USC Lib a ies (2024) Resea ch Guides: O ganising You Social Sciences Resea ch Pape .
Re ie ed om h p://libguides.usc.edu/w i ingguide on 1.Jan.24
Vogel-Heuse , B., & Hess, D. (2016). Gues edi o ial Indus y 4.0–P e equisi es and Visions.
IEEE T ansac ions on Au oma ion Science and Enginee ing. Accessed om IEEE Xplo e on
05.Jan.24
Ve ghese, J. & Chapi o, J. (2023) Cha GPT: The T ans o ma i e In luence o Gene a i e AI on
Science and Heal hca e. Accessed om Jou nal o Hepa ology on 14.Feb.24
WHO (2023) Digi al Tools Posi i ely Impac Heal h Wo ke s’ Pe o mance. Accessed om
WHO on 6.Jan.24
Wee a a hna, I., Raymond, D. & Luha ia, A. (2023). Human-Robo Collabo a ion o
Heal hca e: A Na a i e Re iew. Accessed om Cu eus Sp inge G oup on 3.Jan.24
83
WHO, OECD & WBG (2018) Deli e ing Quali y Heal h Se ices: A Global Impe a i e o
Uni e sal Heal h Co e age. [online] Accessed om WHO, OECD & he Wo ld Bank G oup
Da abase on 4-Sep -23
WeFo um (2024b) How Gene a i e AI and La ge Language Models Can Close he Gap be ween
Da a and Ou comes in Heal hca e. [online] Accessed om Wo ld Economic Fo um on 31.Jan.24
WeFo um (2024c) 3 Ways Gene a i e AI is Reshaping he Pa ien Expe ience. Heal h and
Heal hca e. [online] Accessed om Wo ld Economic Fo um. On 03.Feb.24
WeFo um (2023) Jobs o Tomo ow: La ge Language Models and Jobs. A Whi e Pape . [online]
Accessed om Wo ld Economic Fo um on 20.Feb.24
Wu, A., S ou s, R. & Biljecki, F. (2022) Gene a i e Ad e sa ial Ne wo ks in he Buil
En i onmen : A Comp ehensi e Re iew o he Applica ion o GANs Ac oss Da a Types and
Scales. Accessed om Science Di ec on 05.Feb.24
Wald on (2023) Gene a i e AI in Heal hca e: “Google Cloud’s Amy Wald on on he Tech Gian ’s
Heal h Ambi ions” An In e iew. [online] Accessed om Med-Tech Di e on 07.Feb.24
Wilkinson, D. & Dok e , D. (2023) The Resea che 's Toolki : The Comple e Guide o
P ac i ione Resea ch. 2nd ed. ISBN978-1032018102
Wi s Uni e si y (2023) Resea ch Suppo Guide o Facul y and S uden s. Johannesbu g.
Accessed om Wi s Uni e si y Lib a y. On 13.Feb.24
WeFo um (2024) Pa ien -Fi s Heal h wi h Gene a i e AI: Reshaping he Ca e Expe ience.
Accessed om Wo ld Economic Fo um on 21.Fe.24
Xu, X., Lu, Y., Vogel-Heuse , B. & Wang, L. (2021) Indus y 4.0 and Indus y 5.0—Incep ion,
Concep ion and Pe cep ion. Accessed om Science Di ec on 3.Jan.24
Xiao, Y. & Wa son, M. (2019). Guidance on Conduc ing a Sys ema ic Li e a u e Re iew. Jou nal
o Planning Educa ion and Resea ch. Accessed om Sage Jou nals on 08.Feb.24
84
Ya aghi, N. (2024) Gene a i e AI in Heal hca e: Oppo uni ies, Challenges, and Policy. [online]
Accessed om B ookings.edu on 30.Jan.24
Yang e al. (2022) Ga o T on: A La ge Clinical Language Model. Accessed om Co nell
Uni e si y on 07.Feb.24
Yamin M. (2018). IT Applica ions in Heal hca e Managemen : A Su ey. In e na ional Jou nal
o In o ma ion Technology. Accessed om. Jou nal o In o ma ion Technology on 11.Jan.24
Yang, L., Zhang, Z., Song, Y. e al. (2023) Di usion Models: A Comp ehensi e Su ey o
Me hods and Applica ions. Accessed om ACM Digi al Lib a y on 23.Jan.24
Yu, P., Xu, H., Hu, X., & Deng, C. (2023). Le e aging Gene a i e AI and La ge Language
Models o Heal hca e In eg a ion. Accessed om MDPI Jou nal on 19.Jan.24
Zhang, E., Cheok, A., Pan, Z. e al. (2023) F om Tu ing o T ans o me s: A Comp ehensi e
Re iew and Tu o ial on he E olu ion and Applica ions o Gene a i e T ans o me Models.
Accessed om MDPI Jou nal on 25.Jan.24
Şanlı, Ö., E dem, S., & Te ik, T. (2013). How o W i e a Discussion Sec ion?. Tu kish Jou nal o
U ology. Accessed om Resea ch and P ac ice on 02.Ma .24
85

Sec ion A
Demog aphic Da a & Knowledge Assessmen
1.
Ma k only one o al.
18 - 25
26 - 35
O e 35
Pa icipan Ques ionnai e
In oduc ion
My name is Mada Daniel Nkhoma. I am a Be lin School o Business and Inno a ion
s uden in collabo a ion wi h Uni e si à elema ica in e nazionale Unine uno
headqua e ed in Rome, I aly. I is pa o he E asmus+ and E asmus Mundus
p og amme. I am ca ying ou esea ch on Gene a i e A icial In elligence (AI) in
heal hca e managemen . May you espond o he ques ions ha ollow a you
con enience. The esponses gi en will be anonymous and conden ial. Names will no be
asked o on his o m. I will ake abou 7 minu es o comple e.
I you would like o hea he esul s o his in es iga ion, send me a message a Email:
[email p o ec ed]
Ins uc ions
1. Answe as many ques ions as you can
2. Please do no w i e you names on any o hese o ms
3. Pe sonal da a including email add esses will no be collec ed o sa ed
4. Kindly answe he ques ions below, when necessa y, include explana ions in he spaces
p o ided
* Indica es equi ed ques ion
Please p o ide he age ange you belong o *
APPENDIX 1
2.
Check all ha apply.
3.
Ma k only one o al.
Heal h Se ices Manage
Hospi al o Clinic Adminis a o
Resea che , Heal hca e
O he , Nu sing Adminis a ion
O he , Medical Adminis a ion
O he , Public Heal h
Human Resou ces o Heal h
O he , Heal hca e Educa ion
4.
Ma k only one o al pe ow.
Sex *
Female Male Di e se
P e e
no o
say
You
answe
You
answe
Which o he ollowing i s well wi h you p o ession? Selec an op ion closes o
wha you do
*
Do you know wha Gene a i e A i icial In elligence is? *
Yes No No
Su e
You
answe
You
answe
Gene a i e
Al’s
e olu ion
Fo
an
ad anced
echnology
ha ’s
conside ed
ela i ely
new,
gene a i e
Al
is
deep- oo ed
in
his o y
and
inno a ion.
1932
Geo ges
A s ouni
in en s
a
machine
he
epo edly
called
he
“mechanical
b ain”
o
ansla e
be ween
languages
ona
mechanical
compu e
encoded
on o
punch
ca ds.
1966
MIT
p o esso
Joseph
Weizenbaum
c ea es
he
i s
cha bo ,
Eliza,
which
simula es
con e sa ions
wi h
a
psycho he apis .
1980
Michael
Toy
and
Glenn
Wichman
de elop
he
Unix-based
game
Rogue,
which
uses
p ocedu al
con en
gene a ion
o
dynamically
gene a e
new
game
le els.
1986
Michael
I win
Jo dan
lays
he
ounda ion
o
he
mode n
use
o
ecu en
neu al
ne wo ks
(RNNSs)
wi h
he
publica ion
o
“Se ial
o de :
a
pa allel
dis ibu ed
p ocessing
app oach.”
2000
Uni e si y
o
Mon eal
esea che s
publish
“A
Neu al
P obabilis ic
Language
Model,”
which sugges s
a
me hod
o
model
language
using
eed- o wa d
neu al
ne wo ks.
2011
Apple
eleases
Si i,
a
oice-powe ed
pe sonal
assis an
ha
can
gene a e
esponses
and
ake
ac ions
in
esponse
o
oice
eques s.
2013
Google
esea che
Tomas
Mikolo
and
colleagues
in oduce
wo d2 ec
o
iden i y
seman ic
ela ionships
be ween
wo ds
au oma ically.
2015
S an o d
esea che s
publish
wo k
on
di usion
models
in
he
pape
“Deep
Unsupe ised
Lea ning
using
Noneq
m
The modynamics.”
The
echnique
p o ides
a
way
o
e e se-enginee
he
p ocess
o
adding
noise
oa
inal
image.
2018
Google
esea che s
implemen
ans o me s
in o
BERT,
which
is
ained
on
mo e
han
3.3
billion
wo ds
and
can
au oma ically
lea n
he
ela ionship
be ween
wo ds
in
sen ences,
pa ag aphs
and
e en
books
o
p edic
he
meaning
o
ex .
I
has
110
million
pa ame e s.
Google
DeepMind
esea che s
de elop
AlphaFold
o
p edic ing
p o ein
s uc u es,
laying
he
ounda ion
o
gene a i e
Al
applica ions
in
medical
esea ch,
d ug
de elopmen
and
chemis y.
OpenAl
eleases
GPT
(Gene a i e
P e- ained
T ans o me ).
T ained
on
abou
40
gigaby es
o
da a
and
consis ing
o
117
million
pa ame e s,
GPT
pa es
he
way
o
subsequen
LLMs
in
con en
gene a ion,
cha bo s
and
language
ansla ion.
SYNTACTIC
SiRUCTURE
2023
1957
Linguis
Noam
Chomsky
publishes
Syn ac ic
S uc u es,
which
desc ibes
g amma ical
ules
o
pa sing
and
gene a ing
na u al
language
sen ences.
1968
Compu e
science
p o esso
Te y
Winog ad
c ea es
SHRDLU,
he
i s
mul imodal
Al
ha
can
manipula e
and
eason
ou
a
wo ld
o
blocks
acco ding
o
ins uc ions
oma
use .
1985
Compu e
scien is
and
philosophe
Judea
Pea l
in oduces
Bayesian
ne wo ks
causal
analysis,
which
p o ides
s a is ical
echniques
o
ep esen ing
unce ain y
ha
leads
o
me hods
o
gene a ing
con en
in
a
speci ic
s yle,
one
o
leng h.
1989
Yann
LeCun,
Yoshua
Bengio
and
Pa ick
Ha ne
demons a e
how
con olu ional
neu al
ne wo ks
(CNNs)
can
be
used
o
ecognize
images.
2006
Da a
scien is
Fei-Fei
Li
se s
up
he
ImageNe
da abase,
which
p o ides
he
ounda ion
o
isual
objec
ecogni ion.
2012
Alex
K izhe sky
designs
he
AlexNe
CNN
a chi ec u e,
pionee ing
a
new
way
o
au oma ically
aining
neu al
ne wo ks
ha
ake
ad an age
o
ecen
GPU
ad ances.
2014
Resea ch
scien is
lan
Good ellow
de elops
gene a i e
ad e sa ial
ne wo ks
(GANs),
which
pi
wo
neu al
ne wo ks
agains
each
o he
o
gene a e
inc easingly
ealis ic
con en .
Diede ik
Kingma
and
Max
Welling
in oduce
a ia ional
au oencode s
o
gene a e
images,
ideos
and
ex .
2017
Google
esea che s
de elop
he
concep
o
ans o me s
in
he
seminal
pape
“a en ion
is
all
you
need,”
inspi ing
subsequen
esea ch
in o
ools ha
could
au oma ically
pa se
unlabeled
ex
in o
la ge
language
models
(LLMs).
2021
Openal
in oduces
Dale,
which
can
gene a e
images
om
ex
p omp s.
The
name
is
a
combina ion
o
WALL-E,
he
name
o
a
ic ional
obo ,
and
he
a is
Sal ado
Dali.
2022
Resea che s
om
Runway
Resea ch,
S abili y
Al
and
CompVis
LMU
elease
S able
Di usion
as
open
sou ce code
ha
can
au oma ically
gene a e
image
con en
om
a
ex
p omp .
Opendl
eleases
Cha GPT
in
No embe
o
p o ide
a
cha -based
in e ace
o
i s
GPT
3.5
LLM.
I
a ac s
o e
100
million
use s
wi hin
wo
mon hs,
ep esen ing
he
as es
e e
consume
adop ion
o
a
se ice.
Ge y
Images
and
a
g oup
o
a is s
sepa a ely
sue
se e al
companies
ha
implemen ed
S able
Di usion
o
copy igh
in ingemen .
Mic oso
in eg a es
a
e sion
o
Cha GPT
in o
i s
Bing
sea ch
engine.
Google
quickly
ollows
wi h
plans
o
elease
he
Ba d
cha
se ice
based
on
i s
Lamda
engine.
And
he
con o e sy
o e
de ec ing
Al-gene a ed
con en
hea s
up.
com
e a oc
Aun
ncen en
TechTa ge
APPENDIX 3