Geog aphies o T us : AI, Biomedicine, and he Nex
E a o Fede a ed and Visi ing Da a Models
Au ho s: Pa icia Buendia, Seonyoung Kim, Na alie Meye s, F ancis P. C awley, Ga in Fa ell,
Roni Pu ian, RDA A i icial In elligence Da a Visi a ion WG
Published: 21 Augus 2025
Ve sion: 1
INTRODUCTION
Biomedical esea ch inc easingly elies on as , sensi i e da ase s— om genomic sequences
o clinical eco ds. Ye he e hical, legal, and logis ical challenges o mo ing o copying his da a
o cen alized eposi o ies a e immense. In esponse, a di e se ecosys em o pla o ms has
eme ged, each o e ing a unique app oach o da a p o ec ion, dis ibu ed analysis, and
p i acy-p ese ing compu a ion.
This epo ca ego izes hese pla o ms in o h ee p ima y models: Da a Visi a ion (DV),
Cen alized DV, and Fede a ed. The e is also a Hyb id Cen alized/Fede a ed model o
conside . This epo examines how each model add esses he challenge o enabling esea ch
while main aining da a secu i y.
Beyond cu en implemen a ions, he epo also explo es he e ol ing landscape o AI
echnologies—pa icula ly hose in ol ing dis ibu ed machine lea ning, syn he ic da a
gene a ion, and p i acy-p ese ing compu a ion—and assesses hei po en ial o eshape
biomedical and heal h esea ch da a access and analysis amewo ks. By engaging wi h
eme ging ends, he epo o e s a o wa d-looking pe spec i e on how hese models mus
adap o ensu e e hical, secu e, and scalable esea ch in inc easingly complex da a
ecosys ems.
COMPARATIVE MODELS OF DATA PROTECTION
Th ee Models (and a Hyb id Model)
This sec ion p o ides a isual compa ison o decision guide o help choose he mos
app op ia e echnology o he sha ing o access- es ic ed da a.
● Da a Visi a ion (DV): A decen alized model whe e ools isi locally go e ned da a
wi hou equi ing a cen al eposi o y o uni o m a chi ec u e.
● Fede a ed: A dis ibu ed app oach whe e da a emains local bu equi es a uni o m
a chi ec u e and ini ial se up o enable ool-based access ac oss si es.
● Cen alized Da a Visi ing: A cen alized model whe e da a is pooled in o a cen al
eposi o y, and esea che s access i di ec ly, necessi a ing uni o m a chi ec u e and
se up.
Table 1 ou lines he key dis inguishing ea u es o he h ee models, while Table 2 ca ego izes
hei addi ional pe o mance- ela ed a ibu es in o low, medium, and high ie s. A de ailed
desc ip ion o each model, accompanied by examples, ollows in he nex sec ion. A
comp ehensi e compa a i e able o pla o ms is being main ained and cu a ed he e: Table
Compa ing Fede a ed Da a Pla o ms and Da a Visi ing Technologies and is discussed in
Pla o m Compa ison Table.
Table 1: Main Fea u es o Da a-In-Place Analy ics Pla o ms
Fea u e
DV
Fede a ed
Cen alized DV
Da a s ays in place wi h local
go e nance
Resea che s isi he da a
Uni o m a chi ec u e equi ed
Cen al eposi o y
Tool/model isi s da a
Ini ial da a se up is equi ed
be o e DV
Table 2: Ca ego iza ion o Pe o mance Fea u es
Fea u e
DV
Fede a ed
Cen alized DV
Pe o mance/la ency
Medium (depends
on ne wo k) [1,2]
Low o Medium
(depends on
a chi ec u e) [1,2]
High (local
p ocessing) [3]
Secu i y/p i acy isks
Low (da a s ays
local) [4,5]
Medium (some
exposu e h ough
ede a ion) [4.5]
High (da a
cen alized) [6]
Compliance wi h ju isdic ional
laws
High (da a ne e
lea es ju isdic ion)
[7,8]
Medium o High
(depends on node
compliance) [7,8]
Low o Medium
(cen al eposi o y
may c oss bo de s)
[7,8]
In as uc u e cos dis ibu ion
Dis ibu ed (each
si e co e s i s own)
[9,10]
Sha ed/Dis ibu ed
[9,10]
Cen alized (hos
bea s he cos )
[9,10]
In e ope abili y
Medium ( equi es
compa ible ools)
[11,12,13]
High (uni o m
a chi ec u e)
[11,12,13]
High (cen alized
con ol) [11,12,13]
Da a eshness
High ( eal- ime
possible) [14,15]
Medium o High
(depends on sync)
[14, 15]
High (da a
cen alized and
upda ed cen ally
[14,15]
Cen alized DV
De ini ion: Da a and ools a e main ained in a cen al loca ion. Resea che s access he da a by
isi ing he secu e en i onmen whe e analysis ools a e p o ided.
Key Fea u es:
● Da a emains in a secu e encla e.
● Resea che s access ools like Jupy e , RS udio, o cus om dashboa ds.
● S ong go e nance and audi ails.
Examples:
● All o Us Resea ch Hub: A cen alized, secu e cloud-based esea ch pla o m hos ing
genomic, EHR, su ey, and physical measu emen da a om di e se U.S. pa icipan s.
● UK Biobank: O e s cen alized access o genomic and pheno ypic da a h ough secu e
po als. As o 2024, UK Biobank no longe allows esea che s o download indi idual-
le el da a. The da a is isi ed and no mo ed.
● NCATS N3C by Palan i : COVID-19 da a encla e wi h ha monized da ase s and buil -in
analy ics.
● HDRUK TREs: T us ed en i onmen s o UK heal h da a, wi h s ic access con ols.
● FAIR Squa e: Cen alized FAIRness assessmen ools and me ada a cu a ion.
S eng hs:
● High con ol and secu i y wi hin a single en i onmen .
● Rich oolse s a ailable in one place.
● High pe o mance and low la ency due o local p ocessing.
● S ong in e ope abili y h ough cen alized a chi ec u e.
● Clea , s anda dized onboa ding p ocess o new da a sou ces once pipeline is
es ablished.
Limi a ions:
● Requi es da a mo emen o he cen al loca ion.
● May ace c oss-bo de da a sha ing and localiza ion es ic ions.
● Highe secu i y/p i acy isks i he cen al en i onmen is comp omised.
● Cen alized in as uc u e cos bu den on he hos o ganiza ion.
● Scalabili y is limi ed by he capaci y o he cen al in as uc u e.
● Ini ial da a se up and ha moniza ion equi ed be o e in eg a ion.
Da a Visi a ion
De ini ion: Da a can eside anywhe e and be o any ype. The echnology o ML model isi s
he da a, o en wi hou equi ing i o be pa o a o mal ne wo k.
Key Fea u es:
● Flexible a chi ec u e.
● Ideal o he e ogeneous o sensi i e da ase s.
● O en used in ea ly-s age pilo s o ad hoc collabo a ions.
Examples:
● FAIR Da a T ain: Enables me ada a-d i en access and isi a ion ac oss di e se
da ase s.
● FAIRlyz: Uses LLMs and AI ools o cu a e and analyze da a emo ely.
● OSSDIP: Focuses on secu e access p o ocols and me ada a egis ies.
● O e u e: Me ada a o ches a ion ools ha suppo decen alized da a disco e y.
S eng hs:
● Minimal da a mo emen – da a emains in i s o iginal loca ion
● Adap able o a ious da a ypes, o ma s, and go e nance models.
● High da a eshness due o eal- ime o nea eal- ime access.
● Low sec ui y/p i acy isk since da a does no lea e i s hos en i onmen .
● Dis ibu ed in as uc u e cos ac oss pa icipa ing si es.
● No manda o y ini ial se up—can access da a wi hou p io ha moniza ion.
Limi a ions:
● May lack s anda diza ion o ne wo k cohesion.
● Tool in e ope abili y can be challenging ac oss he e ogeneous si es.
● Va iable pe o mance depending on ne wo k quali y.
● In e ope abili y depends on compa ible ools a each si e.
● Compliance managemen may be inconsis en ac oss si es.
● Scalabili y o many new sou ces can be challenging wi hou p io ha moniza ion.
Fede a ed
De ini ion: Da a is s anda dized and s o ed in nodes wi hin a ne wo k. Each node ollows
uni o m a chi ec u e and SOPs. Technology o ML models isi he da a, no he o he way
a ound.
Key Fea u es:
● S ong s anda diza ion (e.g., OMOP, FHIR).
● Enables ede a ed lea ning and analy ics.
● O en used in la ge-scale conso ia.
Examples:
● Da aSHIELD: R-based s a is ical analysis ac oss dis ibu ed nodes.
● T iNe X/i2b2: Real- ime coho disco e y ac oss hospi al ne wo ks.
● Swa m Lea ning Ne wo ks: Fede a ed AI model aining ac oss ins i u ions.
● ATLAS by OHDSI: OMOP-based analy ics ac oss global nodes.
● EUCAIM: Imaging da a ede a ion wi h AI model aining.
● Genomic Da a In as uc u e (GDI): Fede a ed genomic analysis using GA4GH
s anda ds.
● ELIXIR on Cloud: C oss-si e ede a ion o li e science da a.
● Fede a ed EGA: disco e y and access o sensi i e human omics and associa ed da a
consen ed o seconda y use.
S eng hs:
● High scalabili y - easily add new s anda dized nodes once s anda ds a e in place.
● A consis en onboa ding p ocess o new nodes ensu es quali y and compliance.
● S ong s anda diza ion (e.g., OMOP, FHIR) enables in e ope abili y.
● Da a so e eign y p ese ed a local si es.
● Enables ede a ed lea ning and la ge-scale analy ics.
● High po en ial o consis en compliance ac oss nodes.
● Ideal o mul i-ins i u ional esea ch.
Limi a ions:
● Requi es signi ican up on in es men in ha moniza ion.
● Go e nance complexi y ac oss mul iple o ganiza ions and nodes.
● Pe o mance may be impac ed by in e -node la ency.
● Medicum secu i y/p i acy isk due o ne wo ked access.
● In as uc u e cos s a e sha ed bu equi e sus ained in es men .
● Ini ial da a se up and alignmen o ne wo k s anda ds equi ed be o e pa icipa ion.
Hyb id Cen alized/Fede a ed
De ini ion: Da a s o age is cen alized, bu go e nance and access p o ocols a e ede a ed. I ’s
a hyb id model—no ue ede a ion in a chi ec u e, bu ede a ed in policy.
Key Fea u es:
● Cen alized in as uc u e wi h ede a ed access logic.
● O en used in na ional o mul i-agency ini ia i es.
Examples:
● NCPI AnVIL: Cen alized genomic da a wi h ede a ed go e nance ac oss NIH ins i u es.
● Li ebi and DARE UK: Fede a ed TREs wi h compu e en i onmen s
S eng hs:
● Balances con ol wi h collabo a ion.
● Easie o manage in as uc u e.
Limi a ions:
● S ill in ol es cen alized s o age.
● May no mee s ic da a localiza ion equi emen s.
PLATFORM COMPARISON TABLE
This dynamic compa a i e able ca alogs pla o ms ha suppo DV, Fede a ed, Hyb id
Cen alized/Fede a ed, and Cen alized DV models in biomedical and heal h esea ch. I is
ac i ely main ained o e lec he la es de elopmen s and pla o m capabili ies. The la es
e sion can be accessed he e: Table Compa ing Fede a ed Da a Pla o ms and Da a Visi ing
Technologies.
These pla o ms ep esen a di e se and e ol ing ecosys em designed o enable secu e,
p i acy-p ese ing analysis o sensi i e da a ac oss ins i u ions, wi hou unnecessa y duplica ion
o cen aliza ion.
Each en y includes key a ibu es such as:
● Da a isi a ion ype and a chi ec u e
● So wa e a ailabili y and ools
● Da a models and s anda ds used (e.g., OMOP, FHIR, GA4GH)
● S akeholde s se ed (e.g., esea che s, hospi als, pha ma)
● Ou pu s gene a ed (e.g., s a is ical models, coho insigh s)
● Whe he ools/models a e mo ed o he da a o expec s anda dized inpu s
This able is in ended o suppo s a egic planning, pla o m selec ion, and policy de elopmen
o o ganiza ions seeking o collabo a e ac oss da a bounda ies while main aining compliance
and us .
NEW AI TECHNOLOGIES
The e’s a wa e o eme ging echnologies eshaping how ede a ed pla o ms and da a isi a ion
sys ems ope a e. These inno a ions go beyond GenAI and LLMs, hough many in e sec wi h
hem. Below is a cu a ed lis o cu ing-edge echnologies ha a e ei he being ac i ely explo ed
o show s ong po en ial o impac hese domains.
Eme ging Technologies Impac ing Da a Visi a ion & Fede a ed
Pla o ms
1. Agen ic AI
● Combines LLMs wi h au onomous planning and execu ion capabili ies.
● Enables “ i ual cowo ke s” ha can isi da a sou ces, pe o m mul i-s ep asks, and
e u n esul s.
● Use ul o ede a ed wo k lows whe e human-in- he-loop is limi ed bu agen s a e
human-supe ised.
Key Resou ces
1. Agen ic A i icial In elligence o Suppo Au onomous Medical Ope a ions
NASA’s Human Resea ch P og am (HRP) is le e aging he powe and e iciency o
a i icial in elligence (AI) ools o educe human sys em isk in space medicine
ope a ions.[16]
2. Agen ic AI in Biomedical Resea ch: A New E a o In elligen Collabo a ion
ICA.ai o e s AI solu ion o biomedical esea ch and public heal h. [17]
3. Sample Agen s o Heal hca e and Li e Sciences on AWS
One-click deploymen o in as uc u e and In e ac i e in e ace o human-agen cha
wi h p e-buil agen s.[18]
2. Secu e Mul ipa y Compu a ion (SMPC)
● SMPC allows mul iple pa ies o join ly compu e a unc ion o e hei inpu s while
keeping hose inpu s p i a e.
● Ideal o ede a ed analy ics in inance, genomics, and heal hca e.
● O en pai ed wi h ede a ed lea ning o da a isi a ion models.
Key Resou ces
1. Secu e Mul i-Pa y Compu a ion – Sp inge Link
Use cases discussed: secu e machine lea ning and p i acy-p ese ing ne wo k
moni o ing. [19]
2. SMPC o P i acy-P ese ing Da a Analysis – IJCRT
Explo es SMPC’s ole in heal hca e and c oss-ins i u ional esea ch. [20]
3. Causal AI
● Goes beyond co ela ion-based models o in e cause-e ec ela ionships.
● Valuable in ede a ed heal h esea ch whe e unde s anding causali y is key bu da a
sha ing is es ic ed.
Key Resou ces
1. Fede a ed causal disco e y wi h missing da a in a mul icen ic s udy on endome ial
cance
A no el ede a ed causal disco e y algo i hm capable o pooling in o ma ion om
mul iple sou ces wi h he e ogeneous missing da a o lea n a g aph ep esen ing cause-
e ec ela ionships. [21]
2. Causal disco e y om obse a ional and in e en ional da a
Dis ibu ed and ede a ed causal disco e y app oaches o decen alized scena ios in
heal hca e ha ha e p i acy and egula o y da a access cons ain s. [22]
4. G aph Da a Science (GDS)
● Applies g aph heo y o model ela ionships be ween dis ibu ed da ase s.
● Enhances ede a ed disco e y and linkage o da a ac oss ins i u ions.
Key Resou ces
1. Fede a ed biomedical knowledge g aph-based ques ion-answe ing sys em
Biomedical disco e y h ough an in o ma ics pla o m ha enables explo a ion and
easoning o e an open-sou ce, ede a ed KG-based ecosys em. [23]
2. How Fede a ed Knowledge G aphs Suppo AI Au oma ion
AI au oma ion elies no only on da a, bu also on unde s anding. A ede a ed knowledge
g aph p o ides ha unde s anding in se e al ways. [24]
6. Applica ion-Speci ic Semiconduc o s
● Cus om chips designed o ede a ed AI wo kloads (e.g., p i acy-p ese ing aining).
● Imp o es e iciency and scalabili y o ede a ed pla o ms.
Key Resou ces
1. NVIDIA Cus om GPU A chi ec u es o Fede a ed Lea ning
NVIDIA FLARE is buil on NVIDIA Cla a T ain, which uns on NVIDIA GPUs designed o
high- h oughpu AI wo kloads enabling dis ibu ed aining ac oss hospi als and esea ch
cen e s wi hou sha ing pa ien da a. [25]
2. Azu e TEE is pa o Azu e Con iden ial Compu ing, which enables p i acy-p ese ing AI
and ede a ed lea ning
A T us ed Execu ion En i onmen is a seg ega ed a ea o memo y and CPU ha 's
p o ec ed om he es o he CPU by using enc yp ion. [26]
7. Syn he ic Da a Gene a ion o Twin Da ase s
● C ea es ealis ic bu a i icial da ase s o aining and alida ion.
● Can be used in ede a ed se ings o simula e da a en i onmen s wi hou exposing eal
da a.
Key Resou ces
1. “Digi al win” da ase s om complex EHR and wea able-de ice eco ds
Syn he ic clinical da a gene a ion maximizes biomedical esou ce u iliza ion and
minimizes pa icipan e-iden i ica ion isks. [27]
2. Immune digi al wins o complex human pa hologies
A bioin o ma ics ecosys em is p oposed o da a analysis, in eg a ion and modelling in
Immune digi al wins implemen a ions. [28]
8. Neu o-Symbolic AI
● Combines symbolic AI easoning wi h neu al ne wo ks.
● Use ul o ede a ed sys ems ha equi e explainabili y and ule-based go e nance.
Key Resou ces
1. Neu osymbolic AI can be le e aged in medical diagnos ics
Applied o Men al heal h diagnosis and con e sa ional assis ance. [29]
2. Neu osymbolic AI o easoning on biomedical knowledge g aphs (KG)
KG comple ion (KGC) can help esea che s make p edic ions o in o m asks like d ug
eposi ioning wi h hyb id app oaches based on neu osymbolic a i icial in elligence
becoming mo e popula . [30]
9. Rein o cemen Lea ning in Fede a ed Se ings
● Enables adap i e lea ning ac oss dis ibu ed nodes.
● Can op imize esou ce alloca ion, model upda es, and p i acy ade-o s.
Key Resou ces
1. Fede a ed Lea ning in Sma Heal hca e
A Comp ehensi e Re iew on P i acy, Secu i y, and P edic i e Analy ics wi h IoT
In eg a ion. [31]
2. Fede a ed lea ning applica ions o biomedical da a
Collabo a i ely aining machine lea ning models wi hou sha ing aw da a ia ‘ ede a ed
lea ning’ enables in es iga o s o ain a model locally on hei own da a, and sha e he
pa ame e s o he model wi h o he s o gene a e a cen al model. [32]
10. P i acy-P ese ing MLOps
● In eg a es p i acy ools (e.g., di e en ial p i acy, homomo phic enc yp ion) in o model
deploymen pipelines.
● Ensu es end- o-end compliance in ede a ed AI wo k lows
