Structuring electronic health records of breast cancer with Natural Language Processing

P ocesamien o del Lenguaje Na u al pa a la es uc u ación de his o iales clínicos de cánce de mama S uc u ing elec onic heal h eco ds o b eas cance wi h Na u al Language P ocessing T abajo de Fin de G ado Cu so 2022–2023 Au o Ál a o Ga cía Ba agán Di ec o Víc o Robles Fo cada Tu o UCM: José Ignacio Hidalgo Pé ez Colabo ado E nes ina Menasal as G ado en Ingenie ía In o má ica Facul ad de In o má ica Uni e sidad Complu ense de Mad id P ocesamien o del Lenguaje Na u al pa a la es uc u ación de his o iales clínicos de cánce de mama S uc u ing elec onic heal h eco ds o b eas cance wi h Na u al Language P ocessing T abajo de Fin de G ado en Ingenie ía In o má ica Au o Ál a o Ga cía Ba agán Di ec o Víc o Robles Fo cada Tu o UCM: José Ignacio Hidalgo Pé ez Colabo ado E nes ina Menasal as Con oca o ia: Junio 2023 G ado en Ingenie ía In o má ica Facul ad de In o má ica Uni e sidad Complu ense de Mad id 29 de mayo de 2023 Dedica o ia A mi mad e, po sabe de p ime a mano que es el cánce de mama y supe a lo Ag adecimien os A odo el g upo de in es igación del labo a o io MEDAL, comenzando po Víc o y E nes, quienes desde el p ime día me han hecho sen i como en casa, y e minando con odos los es udian es con quienes he enido la opo unidad de abaja y c ece an o p o esionalmen e como pe sonalmen e. ii Resumen P ocesamien o del Lenguaje Na u al pa a la es uc- u ación de his o iales clínicos de cánce de mama Los his o iales clínicos es án esc i os en lenguaje na u al y, po an o, es in- o mación no es uc u ada. El obje i o del p oyec o es es uc u a la in o mación de his o iales clínicos de pacien es con cánce de mama de un hospi al público de Mad id con el in de consegui in o mación u il pa a los medicos. De es a o ma, se p opone ealiza la es uc u ación a a és del uso de edes neu onales p o undas pa a la clasi icacion de las en idades: NER (Named En i y Recogni ion), en conjun o con o as écnicas de NLP. Finalmen e se gene a á una base de da os semies uc- u ada en o ma o json con los his o iales, que pod á se p ocesada pos e io men e con di e en es in enciones. Palab as cla e P ocesamien o del Lenguaje Na u al (PNL), Ap endizaje P o undo, T ans o - me s, Reconocimien o de en idades nomb adas, Ex acción de in o mación, Ex ac- ción del diagnós ico del cánce , Cánce de mama ix Lis o igu es 2.1. Spacy okeniza ion [1] . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.2. Wo d embbedings ep esen ed in 3D [2] . . . . . . . . . . . . . . . . . 11 2.3. Rep esen a ion o an RNN . . . . . . . . . . . . . . . . . . . . . . . 12 2.4. LSTM cell a chi ec u e [3] . . . . . . . . . . . . . . . . . . . . . . . . 13 2.5. BILSTMlaye o NER.......................... 14 2.6. La ge Language Models: A New Moo e’s Law ? [4] . . . . . . . . . . 15 2.7. Compa ing CNN (padding okens a e omi ed), RNN, and sel -a en ion a chi ec u es[5].............................. 16 2.8. BERT s.GPT[6] ............................ 17 2.9. Ex ac ing cance concep s using BERT. . . . . . . . . . . . . . . . . 17 2.10.BERT ine uning[6] ........................... 18 2.11.SpaCypipeline[1]............................. 21 2.12. Ins ance o spaCy syn ac ic ee . . . . . . . . . . . . . . . . . . . . . 22 2.13. Example o IOB o ma . . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.14. Example o CONLL o ma . . . . . . . . . . . . . . . . . . . . . . . 26 2.15. 5- old c oss alida ion . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 4.1. T ans o ming Clinical No es in o S uc u ed JSON . . . . . . . . . . 33 4.2. Co puscollec ion ............................. 34 4.3. Modelsc ea ion.............................. 34 4.4. NCP: NLP Cance Pipeline . . . . . . . . . . . . . . . . . . . . . . . 35 4.5. Technology pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 4.6. Au oma ed pipeline in ba ches . . . . . . . . . . . . . . . . . . . . . . 37 4.7. P e-p ocessing o he no es . . . . . . . . . . . . . . . . . . . . . . . . 38 4.8. Example o an anno a ed clinical no e . . . . . . . . . . . . . . . . . . 39 4.9. Nega ion and Unce ain y Anno a ions . . . . . . . . . . . . . . . . . 41 4.10. JSON o ma o s uc u ing b eas cance in o ma ion . . . . . . . . 42 x ii 5.1. G aphical ep esen a ion o Neg Unce Co pus coun ing . . . . . . . 45 5.2. G aphical ep esen a ion o Clinical co pus coun ing . . . . . . . . . . 46 Lis o ables 2.1. Numbe o no es pe se ice and ype . . . . . . . . . . . . . . . . . . 8 2.2. P e ained models de ails [7] . . . . . . . . . . . . . . . . . . . . . . . 19 2.3. spaCyNLPPipeline ........................... 22 2.4. spaCy Command Line In e ace . . . . . . . . . . . . . . . . . . . . . 23 4.1. De ini ion o he en i ies . . . . . . . . . . . . . . . . . . . . . . . . . 40 4.2. No maliza ion............................... 42 5.1. Model hype -pa ame e s . . . . . . . . . . . . . . . . . . . . . . . . . 44 5.2. In o ma ion abou NER co pus . . . . . . . . . . . . . . . . . . . . . 44 5.3. Resul s om Nega ion and Unce ainly Co pus pe en i y ype . . . . 47 5.4. Gene al esul s om Clinical Co pus . . . . . . . . . . . . . . . . . . 47 5.5. Resul s om Clinical co pus pe en i y ype . . . . . . . . . . . . . . 47 5.6. Discussion o Clinical NER model . . . . . . . . . . . . . . . . . . . . 49 xix Chap e 1 In oduc ion “Science is no only a discipline o eason, bu also o omance and passion” — S ephen Hawking Cance emains one o he main public heal h p oblems, anked as he leading cause o dea h globally [8]. Acco ding o he Wo ld Heal h O ganiza ion (WHO)1, cance caused nea ly 10 million dea hs wo ldwide in 2020. In pa icula , b eas cance is cu en ly he mos common cance globally, accoun ing o 12.5% o all new annual cance cases wo ldwide2. In 2020, he e we e 2.3 million women diagnosed wi h b eas cance and 685,000 dea hs a ound he wo ld. The p ocess o diagnosing and ea ing cance pa ien s gene a es a huge amoun o in o ma ion ha desc ibes symp oms, he cance diagnosis, amily his o y, ea - men s, and he e olu ion o he pa ien a he ime. Physicians egis e his in o - ma ion in Elec onic Heal h Reco ds (EHR) using clinical no es w i en in na a i e o m. Ex ac ing and mining his in o ma ion is c ucial o suppo oncology esea ch, design ea men plans, and imp o e pa ien ou comes [9]. Howe e , ex ac ing in o ma ion om clinical na a i es is a challenge due o he complexi y o na u al language [10]. Mo eo e , clinical ex s a e w i en by highly skilled physicians and nu ses using domain-speci ic e ms, unde ime p essu e, wi h ich and complex ja gon, which makes hese ex s di e om hose o o he domains. In ecen yea s he use o Na u al Language P ocessing (NLP) in he biomedical domain has inc eased he possibili y o au oma ically ex ac ing in o ma ion om clinical na a i es. NLP is a sub- ield o Compu e Science (CS), and A i icial In- elligence (AI) ha enables compu e s o unde s and he meaning o human na u al language [11]. The applica ion o NLP and AI echniques o p ocessing medical eco ds plays an inc easingly signi ican ole in ad ancing clinical decision suppo . The use o EHR o pe o m s udies in he cance ield has also inc eased in he las ew yea s. 1h ps://www.who.in /news- oom/ ac -shee s/de ail/cance 2h ps://www.b eas cance .o g/ ac s-s a is ics 1 2Chap e 1. In oduc ion Cu en ly, we a e expe iencing a signi ican boom in he ield o AI, wi h new and e olu iona y sys ems eme ging on a egula basis. One a ea ha is pa icula ly g owing is NLP, hanks o he elease o he esea ch pape "A en ion is all you need" [12], which in oduced he inno a i e T ans o me s a chi ec u e. This has led o he c ea ion o popula language models such as Cha GPT [13], which a e causing a buzz in he indus y. Tha ’s why his a chi ec u e has been chosen o he ask desc ibed in his p ojec . This p ojec is hoped o b ing some imp o emen s o he e iciency and quali y o heal hca e, as well as making some p og ess in medical esea ch and public heal h policy planning ela ed o b eas cance . 1.1. Mo i a ion Medical eco ds ha e exis ed o a long ime, bu he o maliza ion o elec onic heal h eco ds began in he ea ly 20 h cen u y. P io o his, doc o s eco ded no es in pape books and physical iles, bu hese eco ds we e o en incomple e, diso ganized, and di icul o access. The o maliza ion o medical eco ds began in 1928 when he Ame ican Associa- ion o Medical Reco d Lib a ians and Heal h In o ma ion Managemen (AHIMA)3 es ablished a commi ee o de elop a uni o m sys em o medical eco ds. This com- mi ee de eloped he i s edi ion o he In e na ional Lis o Medical Te ms in 1932, which was designed o be a guide o consis en and uni o m medical eco d keeping. In he ollowing yea s, medical eco ds became inc easingly impo an o heal h- ca e. In he 1950s, in o ma ion echnology began o be used in heal hca e, and he use o compu e ized sys ems o medical eco d keeping became mo e common. In he 1970s, au oma ed medical eco d keeping sys ems we e in oduced, which al- lowed doc o s o inpu da a di ec ly in o a compu e . In he 1990s, in e ne echnology began o be used o medical eco d keeping. This allowed doc o s o sha e in o ma ion mo e easily and quickly, and also allowed pa ien s o access hei medical eco ds online. EHRs ha e e ol ed in di e en gene a ions as echnology has ad anced. The di e en gene a ions a e as ollows [14]: Fi s gene a ion: This gene a ion eme ged in he 1970s and was based on au oma ed medical eco d sys ems. These sys ems we e limi ed in hei unc- ionali y and we e designed p ima ily o s o e basic pa ien in o ma ion, such as diagnoses, ea men s, and es esul s. Second gene a ion: In he 1980s, EHRs second-gene a ion we e de eloped, which we e based on compu e ized medical eco d sys ems and o e ed mo e unc ionali y. These sys ems allowed doc o s o inpu de ailed in o ma ion abou pa ien heal hca e, including eco ds o pas isi s, p esc ibed medica- ions, and labo a o y es esul s. 3h ps://www.ahima.o g/ 1.1. Mo i a ion 3 Thi d gene a ion: The hi d gene a ion was de eloped in he 1990s and o- cused on in e connec i i y and da a exchange be ween di e en heal hca e sys ems. These sys ems allowed doc o s o access medical in o ma ion om pa ien s o o he heal hca e p o ide s and we e designed o imp o e heal hca e coo dina ion. Fou h gene a ion: EHRs ou h-gene a ion eme ged in he ea ly 2000s and we e based on web echnology. These sys ems allowed doc o s o access pa ien medical eco ds online and also allowed pa ien s o access hei own medical in o ma ion and communica e wi h hei doc o s ia he web. Fi h gene a ion4: Cu en ly, i h-gene a ion is being de eloped. This gene - a ion will use ad anced echnologies such as machine lea ning and a i icial in elligence o imp o e diagnos ic accu acy and clinical decision-making. In his way, an EHR speci ically e e s o he elec onic eco d o a pa ien ’s med- ical in o ma ion. The e o e, an EHR is a compu e sys em ha allows heal hca e p o essionals o cap u e, s o e, and access a pa ien ’s medical in o ma ion elec on- ically. This includes he pa ien ’s medical and heal h in o ma ion, such as medical his o y, diagnoses, ea men s, es esul s, and p esc ibed medica ions. On he o he hand, Heal h Ca e In o ma ion Sys em (HCIS) e e o clinical in- o ma ion sys ems used in hospi als o manage pa ien in o ma ion, s a , and hospi- al esou ces. These sys ems ypically include modules o appoin men scheduling, pa ien admission, clinical da a eco ding, pa ien acking, pha macy managemen , billing managemen , and human esou ces managemen . The e o e, an HCIS usually includes a module o EHR egis a ion. The Mad id Heal h Se ice (SERMAS) has a unique and cen alized Elec onic Heal h Reco d sys em in he p ima y ca e sys em. Howe e , in he specialized ca e sys em (hospi als), he e a e di e en HCIS, which a e speci ic o each hospi al cen e . Some examples o hese a e: HCIS include Heal hca e P o ide - Heal h Ca e In o ma ion Sys em (HP-HCIS), Elec onic Medical Reco d Sys em (SELENE) and Medical and Heal h In o ma ion Exchange (IMDH). In HCIS, in o ma ion can be classi ied in o wo ca ego ies, s uc u ed and un- s uc u ed: S uc u ed in o ma ion e e s o in o ma ion ha is s o ed in da a ields wi h de ined o ma s, making i easie o compu e sys ems o p ocess and analyze. Examples o s uc u ed in o ma ion include pa ien demog aphics, diagnoses, es esul s, p esc ibed medica ions, among o he s. Uns uc u ed in o ma ion e e s o in o ma ion ha does no ollow a de ined o ma and can be mo e di icul o p ocess and analyze au oma ically. Ex- amples o uns uc u ed in o ma ion include medical epo s, p og ess no es, clinical obse a ions, medical images, among o he s. 4h ps://www2.deloi e.com/us/en/insigh s/indus y/heal h-ca e/eh -sys ems- he- u u e-o - elec onic-heal h- eco ds.h ml 4Chap e 1. In oduc ion The s a ing hypo hesis is ha i is possible o ex ac use ul in o ma ion om uns uc u ed da a in HCIS and con e i in o s uc u ed in o ma ion ha can be u ilized by compu e sys ems. Thus, he main mo i a ion behind his wo k is he de elopmen o NLP ools o ex ac use ul in o ma ion om uns uc u ed da a in HCIS. This uns uc u ed da a may include medical epo s, p og ess no es and clinical obse a ions. 1.2. Objec i es The main objec i es o his wo k a e: To de elop and e alua e NLP ools o ex ac ing use ul in o ma ion om uns uc u ed da a in HCIS in spanish. To con e he ex ac ed in o ma ion in o s uc u ed da a ha can be used by compu e sys ems in HCIS. De elop an au oma ed NLP pipeline ha can ex ac he mos impo an in o ma ion om a gi en clinical no e. These objec i es en ail hese implica ions: To imp o e he e iciency and quali y o heal hca e by making de ailed and accu a e pa ien in o ma ion eadily a ailable o heal hca e p o essionals. To enhance he abili y o compu e sys ems o analyze and gene a e accu a e and de ailed epo s on he heal h and well-being o pa ien s. To con ibu e o medical esea ch and public heal h policy planning by p o- iding accu a e and de ailed pa ien in o ma ion. To ca y ou his ask o de eloping NLP ools o ex ac use ul in o ma ion om uns uc u ed da a in HCIS, b eas cance da a om a enowned hospi al will be used. These da a will include medical epo s, p og ess no es, and clinical obse a ions ela ed o b eas cance . Addi ionally, we will wo k in collabo a ion wi h heal hca e p o essionals and b eas cance specialis s o ensu e he ele ance and use ulness o he ex ac ed and con e ed s uc u ed in o ma ion o use in HCIS. 1.3. Wo k Plan In o de o ul ill he goals o he p ojec he ollowing asks ha e been de ined: Iden i y he speci ic ypes o uns uc u ed da a ha a e mos ele an o he p ojec and he ypes o in o ma ion ha need o be ex ac ed. 1.3. Wo k Plan 5 Conduc a ho ough li e a u e e iew on NLP ools and echniques o ex- ac ing in o ma ion om uns uc u ed da a in HCIS. Ga he and p ep ocess he uns uc u ed da a om he HCIS, cleaning and s anda dizing i o u he analysis. De elop and ain NLP models o ex ac he desi ed in o ma ion om he uns uc u ed da a, such as named en i y ecogni ion o ela ion ex ac ion. E alua e he pe o mance o he NLP models using app op ia e me ics and adjus he models as needed. 12 Chap e 2. Ma e ials and Techniques uns uc u ed ex da a, and o label hem wi h hei co esponding ca ego ies. This is a challenging ask, as named en i ies can be exp essed in many di e en o ms and can appea in a ious con ex s wi hin he ex . Fu he mo e, people o en make spelling e o s, which can u he complica e his ask. Techniques o NER include: egex, RNNs, LSTM and BiLSTM ne wo ks. LMs like BERT ha e achie ed s a e-o - he-a pe o mance on NER asks, hanks o he T ans o me s a chi ec u e. Models a e p e ained on la ge da ase s in an unsu- pe ised manne , and hen ine- uned on smalle , ask-speci ic da ase s. All hese echniques, o ganized by he quali y o hei ou comes, a e de ailed in his sec ion. 2.4.1. REGEX Regula exp ession, also known as egex o egexp, is a pa e n made up o a sequence o cha ac e s ha is used o sea ch o o ma ch speci ic pa e ns o ex . I is a powe ul ool ha allows you o speci y a se o ules o ma ching and manipula ing ex . Regula exp essions a e widely used in p og amming, da a p ocessing, and ex edi ing applica ions o sea ch, eplace, and ex ac ex based on speci ic pa e ns. In he ield o NLP, i is common o use egex o asks such as ex cleaning, no maliza ion o NER. These echniques a e widely used o iden i y and ex ac impo an in o ma ion om ex , as well as o p epa e he ex o u he analysis o machine lea ning asks. 2.4.2. RNN and LSTM Recu en Neu al Ne wo ks (RNNs) ha e been widely used in NLP o NER. RNNs a e a ype o neu al ne wo k a chi ec u e ha can handle sequen ial da a, making hem well-sui ed o NER asks. In NER, RNNs p ocess ex inpu s one oken a a ime, and hey use he con ex o he p e ious okens o classi y each oken as a named en i y o no . The key ad an age o RNNs is hei abili y o cap u e dependencies be ween he cu en oken and p e ious okens, which can be essen ial o accu a e NER. Figu e 2.3: Rep esen a ion o an RNN 2.4. NER 13 The anishing g adien p oblem is a common issue ha a ises when aining Recu en Neu al Ne wo ks (RNNs) on long sequences o da a. The p oblem a ises when g adien s, which a e used o upda e he ne wo k weigh s du ing aining, become ex emely small as hey a e backp opaga ed h ough he ne wo k om he ou pu o he inpu laye . When his happens, he ne wo k may no be able o lea n long- e m dependencies e ec i ely and may pe o m poo ly on long sequences. The anishing g adien p oblem is pa icula ly p e alen in RNNs because hey p ocess inpu sequences one elemen a a ime, and he g adien s mus low back h ough he same se o weigh s o each elemen in he sequence. As a esul , e o s can accumula e and cause g adien s o become e y small o e en anish o e ime. Unlike adi ional RNNs, which su e om he anishing g adien p oblem and ha e di icul y lea ning long- e m dependencies, LSTMs a e designed o cap u e long- e m dependencies by using a mo e complex ga ing mechanism ha selec i ely emembe s o o ge s in o ma ion o e ime. Long Sho -Te m Memo y (LSTM) was in oduced in 1997 by Hoch ei e and Schmidhube [21] and ha e since become one o he mos popula and e ec i e a chi ec u es o sequen ial da a p ocessing. They consis o memo y cells ha s o e in o ma ion o e ime, inpu ga es ha con ol he low o in o ma ion in o he cells, ou pu ga es ha con ol he low o in o ma ion ou o he cells, and o ge ga es ha selec i ely disca d in o ma ion om he cells. Figu e 2.4: LSTM cell a chi ec u e [3] O e all, LSTMs a e a powe ul and lexible ool o modeling sequen ial da a and ha e enabled signi ican ad ances in a wide ange o applica ions. 2.4.3. BiLSTM Bidi ec ional LSTM [22] (BiLSTM) is a a ian o he Long Sho -Te m Memo y (LSTM) a chi ec u e ha has he abili y o p ocess inpu sequences in bo h o wa d and backwa d di ec ions. This allows he ne wo k o cap u e in o ma ion om bo h pas and u u e con ex s, making i pa icula ly use ul o sequence labeling asks such as NER. In con as , s anda d LSTMs p ocess inpu sequences in only one di ec ion, yp- ically om pas o u u e. While LSTMs a e also e ec i e a cap u ing long- e m 14 Chap e 2. Ma e ials and Techniques Figu e 2.5: BILSTM laye o NER dependencies in sequences, hey may no be as e ec i e as BiLSTMs o asks ha equi e bidi ec ional con ex . In p ac ice, BiLSTMs o en ou pe o m LSTMs in asks ha equi e bidi ec ional con ex , pa icula ly when dealing wi h longe sequences o da a. Howe e , hey can also be mo e compu a ionally expensi e and equi e mo e memo y han LSTMs, due o he need o s o e and p ocess in o ma ion in bo h o wa d and backwa d di ec ions. 2.4.4. LMs A Language Model (LM) is a s a is ical model ha is ained on a co pus o ex da a o p edic he likelihood o a sequence o wo ds occu ing in a language. The model assigns a p obabili y sco e o each possible sequence o wo ds in he language, based on he equency o occu ence o hose wo ds in he aining da a. In o he wo ds, an LM is a machine lea ning algo i hm ha is capable o gen- e a ing ex ha is simila o human language by modeling he s a is ical pa e ns and s uc u es o language. The goal o an LM is o gene a e ex ha is cohe en , g amma ical, and meaning ul, gi en a sequence o inpu wo ds o cha ac e s. Language Models (LMs) ha e a wide ange o applica ions in na u al language p ocessing, including bu no limi ed o he ollowing: Tex Comple ion: LMs can be used o p edic he nex wo d o ph ase in a gi en ex , which is use ul o applica ions such as au o-comple ion in ex 2.4. NER 15 edi o s o sea ch engines. Machine T ansla ion: LMs can be used o ansla e ex om one language o ano he by modeling he p obabili y dis ibu ion o wo ds in bo h languages. Speech Recogni ion: LMs can be used o con e spoken language o ex by modeling he p obabili y dis ibu ion o wo ds in he spoken language. Sen imen Analysis: LMs can be used o analyze he sen imen o a piece o ex , such as de e mining whe he a e iew is posi i e o nega i e. Ques ion Answe ing (Q&A): LMs can be used o answe ques ions posed in na u al language by gene a ing an answe based on he inpu ex and he con ex o he ques ion. Named En i y Recogni ion(NER): LMs can be used o lea n associa ions be- ween wo ds and speci ic named en i ies, like o ganiza ions o loca ions. Some examples o LMs used o NER include BERT (see subsec ion 2.4.6) [6] and Gene a i e P e- ained T ans o me (GPT) [23] models. These models ha e achie ed s a e-o - he-a pe o mance on NER asks and ha e been used in a ious na u al language p ocessing applica ions, such as in o ma ion ex ac ion and ex classi ica ion. Figu e 2.6: La ge Language Models: A New Moo e’s Law ? [4] 16 Chap e 2. Ma e ials and Techniques 2.4.5. T ans o me s T ans o me s is a new a chi ec u e de elop by google enginee s in 2017 [12], which has become he s a e o he a hanks o wo concep s: Posi ional encoding is a echnique used o inco po a e in o ma ion abou he posi ion o he okens in a sequence in o he inpu ep esen a ion. In con as o RNNs, which use he o de o he wo ds implici ly h ough hei sequen ial p ocessing, ans o me s a e no designed o p ocess inpu s sequen- ially. Ins ead, hey p ocess all he wo ds in a sequence in pa allel. To inco - po a e posi ion in o ma ion in o he inpu , ans o me s use a ixed unc ion ha maps each posi ion in he sequence o a unique ec o ep esen a ion. This ec o ep esen a ion is added o he embedding o each wo d, allowing he ans o me o di e en ia e be ween wo ds based on hei posi ion in he sequence. Sel a en ion mechanism is used o cap u e he ela ionships be ween wo ds in a sequence, allows he model o ocus on di e en pa s o he inpu du ing p ocessing. The a en ion ma ix de e mines how much a en ion o gi e o each inpu ec o when compu ing he weigh ed sum. By paying a en ion o he mos impo an pa s o he inpu , he sel -a en ion mechanism helps he model o be e cap u e he ela ionships and dependencies be ween di e en pa s o he ex , which can ul ima ely lead o mo e accu a e p edic ions. Figu e 2.7: Compa ing CNN (padding okens a e omi ed), RNN, and sel -a en ion a chi ec u es [5] Fo mo e in o ma ion abou his sec ion, please e e o my p esen a ion a ailable a : Link p esen a ion. I ha e c ea ed his p esen a ion o unde s and he echnology behind ans o me s. 2.4. NER 17 2.4.6. BERT Bidi ec ional Encode Rep esen a ions om T ans o me s (BERT) is based on a deep neu al ne wo k composed o se e al laye s o bidi ec ional ans o me en- code s. The encode s used a e simila o hose ound in he o iginal ans o me a chi ec u e, bu di e in hei bidi ec ional app oach. Ins ead o p ocessing ex om le o igh o igh o le , i uses in o ma ion om bo h di ec ions o gene a e con ex ual ep esen a ions o wo ds. On he o he hand, GPT T ans o me [24] uses sel -a en ion ha limi s each symbol’s a en ion o only he con ex o i s le . Figu e 2.8: BERT s. GPT [6] The igu e depic ed in Figu e 2.9 illus a es he u iliza ion o BERT o he pu pose o ex ac ing named en i ies om clinical eco ds. Figu e 2.9: Ex ac ing cance concep s using BERT. 18 Chap e 2. Ma e ials and Techniques 2.4.7. P e aining and Fine Tuning P e aining e e s o he p ocess o aining a model on a la ge da ase in an unsupe ised manne . The main objec i es a e: Lea ning con ex ualized wo d embeddings: lea n wo d embeddings ha cap- u e he meaning o wo ds based on hei su ounding con ex . T ans e lea ning: model can lea n gene al linguis ic ea u es ha can be use ul o a wide ange o asks. E iciency: a model can lea n use ul ea u es ha can educe he amoun o da a needed o ine- uning, and can also speed up aining and in e ence imes. BERT has been p e- ained in wo unsupe ised asks: Masked Language Modeling (MLM): ce ain okens in a sen ence a e andomly masked and he model is ained o p edic he o iginal wo ds. The pe cen age o okens ha a e masked in BERT du ing aining is ypically se o 15% o he inpu okens. Nex sen ence p edic ion (NSP): The aim o NSP is o each he model o unde s and he ela ionship be ween wo consecu i e sen ences in a ex . Fine- unning is he p ocess o adjus ing a p e- ained LM o i on a smalle labeled da ase o a speci ic ask h ough a supe ised aining p ocess. Due o p e aining, he da a se equi ed is small and leads o equi es less lea ning ime. Fi ing o a speci ic ask, o en equi es adding a laye a he end o he model. As depic ed in Figu e 2.10, BERT is capable o pe o ming asks such as SQuAD5, NER and MNLI6. Figu e 2.10: BERT ine uning [6] 5S an o d Ques ion Answe ing Da ase [25] 6Mul i-Gen e Na u al Language In e ence [26] 2.4. NER 19 2.4.8. BERT p e ained se ups BERT models ha e been p e ained wi h di e en s con igu a ions ha a e a ail- able in Hugging Face Hub7. These o e a ange o choices wi h di e en sizes and p e- aining cha ac e is ics enabling esea che s o selec he mos app op ia e model o hei speci ic NLP asks and language equi emen s. The able p o ided in Table 2.2 showcases a ious p e- ained models wi h di e en con igu a ions. I is impo an o no e ha he able ep esen s only a subse o he ull ange o a ailable models. These speci ic con igu a ions a e p ima ily designed o p ocessing Spanish o English ex s. He e a e some conside a ions o selec ing a BERT model: Model Size and Capaci y: Models base a e smalle and ha e ewe pa ame e s, making hem sui able when compu a ional esou ces a e limi ed o o smalle - scale asks. On he o he hand, models la ge a e la ge models wi h mo e capaci y, which can be bene icial o complex asks ha equi e cap u ing ine-g ained language pa e ns. Case Sensi i i y: I p ese ing he o iginal casing o wo ds is impo an o you ask, you should choose he cased a ian s. These models e ain he o iginal casing in he p e- aining phase and can be use ul when he dis inc ion be ween uppe case and lowe case is seman ically signi ican . Language: I you NLP ask in ol es mul iple languages, you can conside us- ing he mul ilingual models. These models a e ained on ex om mul iple languages and can handle a ious languages simul aneously, making hem sui - able o mul ilingual applica ions. Howe e , i you ask speci ically in ol es he Spanish language, he be o models a e a good op ion. Table 2.2: P e ained models de ails [7] Model Laye s Hidden Heads Pa ams T aining Da a be -base-uncased 12 768 12 110M LC English ex be -la ge-uncased 24 1024 16 340M LC English ex be -base-cased 12 768 12 110M Cased English ex be -la ge-cased 24 1024 16 340M Cased English ex be -base-mul ilingual-uncased 12 768 12 110M LC ex in op 102 languages wi h la ges Wikipedias be -base-mul ilingual-cased 12 768 12 110M Cased ex in op 104 languages wi h la ges Wikipedias be o-base-uncased 12 1024 16 110M LC Spanish ex be o-base-cased 12 1024 16 110M LC Spanish ex *LC : Lowe -Cased 7h ps://hugging ace.co/ 20 Chap e 2. Ma e ials and Techniques 2.5. NLP ool: spaCy Al hough he e a e nume ous NLP ools a ailable such as NLTK8, Spa kNLP9, and S an o d Co eNLP10, we op ed o spaCy due o i is de eloped wi h Py hon, i s in ui i e and use - iendly in e ace, and i s wide ange o powe ul ea u es. Thus, SpaCy11 is an open-sou ce Py hon lib a y ha is widely used o NLP asks. I has been de eloped by Explosion AI a company ha specializes in NLP echnologies. I has a numbe o impo an ea u es. Some o he mos ou s anding ones a e: E iciency: is a e y as and e icien ool o p ocessing la ge amoun s o ex , because i is p og ammed in Cy hon12. Ease o use: is easy o ins all and use, e en o hose who ha e no expe ience in na u al language p ocessing. Accu acy: uses machine lea ning models o analyse he ex , which allows i o ha e a high accu acy in he analysis. Cus omisa ion: allows you o cus omise machine lea ning models o sui spe- ci ic needs and imp o e pa sing accu acy. Mul ilingual suppo : suppo s mul iple languages, making i a use ul ool o analysing ex in di e en languages. In eg a ion: in eg a es easily wi h o he na u al language p ocessing ools and amewo ks as Tenso low 13, making i a popula choice in he de elopmen o ex analysis applica ions and se ices. Suppo : Suppo o mul iple NLP asks, including NER, POS agging, de- pendency pa sing, and ex classi ica ion Fu he mo e, can con e wo d ec o s om popula ools such as Fas Tex 14 and Gensim15, o load hem in o any p e- ained ans o me model. I is also possible o pe o m one’s own p e- aining o he language. SpaCy’s p ocessing pipeline in ol es h ee main componen s: language p ocess- ing, linguis ic analysis, and machine lea ning. The ool also comes wi h a CLI (see subsec ion 2.5.2) o simpli ied model de elopmen . SpaCy employs an "ea ly s opping" echnique du ing aining. In his sec ion, hese ea u es a e explained. 8h ps://www.nl k.o g/ 9h ps://spa knlp.o g 10h ps://s an o dnlp.gi hub.io/Co eNLP/ 11h ps://spacy.io/ 12h ps://cy hon.o g/ 13h ps://www. enso low.o g/ 14h ps:// as ex .cc/ 15h ps:// adim ehu ek.com/gensim/ 2.5. NLP ool: spaCy 21 2.5.1. P ocessing pipelines The spaCy pipeline is a sequence o componen s ha a e sequen ially applied o a ex o pe o m di e en NLP asks as shown in he Figu e 2.11. In spaCy, he pipeline can be cus omized o include only he componen s needed o a speci ic NLP ask. The componen s in he pipeline a e execu ed in a p ede ined o de , bu i is possible o modi y he o de o add new componen s as needed. The componen s o he spaCy pipeline can be di ided in o h ee main ca ego ies: Language p ocessing componen s: pe o m p ep ocessing asks, such as o- keniza ion and lemma iza ion, which con e ex in o a mo e manageable s uc u e o downs eam componen s. Linguis ic analysis componen s: pe o m mo e complex asks, such as pa sing and named en i y iden i ica ion, o ex ac seman ic and s uc u al in o ma ion om he ex . Machine lea ning componen s: use machine lea ning models o pe o m NLP asks, such as ex classi ica ion o POS labeling. These models a e p e- ained on la ge da ase s o lea n linguis ic pa e ns, and hen uned o speci ic asks using smalle da ase s. Table 2.3 shows he main componen s o spacy. Howe e , new componen s can be c ea ed and added acco ding o he needs o he applica ion. Figu e 2.11: SpaCy pipeline [1] 28 Chap e 2. Ma e ials and Techniques is used as he es se while he emaining olds a e used o aining. This p ocess helps p o ide a mo e obus es ima e o he model’s pe o mance by using all he da a o bo h aining and es ing. The a e age pe o mance ac oss all i e a ions is calcula ed o assess he model’s gene aliza ion abili y and de ec any issues such as o e i ing o unde i ing. Common alues o k include 5 and 10. k- old CV =1 k k X i=1 e alua e(Mi)(2.5) Figu e 2.15: 5- old c oss alida ion 2.8. Wo ds simila i y In NLP i is no uncommon o wo ds o con ain spelling e o s. Howe e , when i comes o s uc u ed da abases, i is impe a i e ha he wo ds a e in hei co ec o m. To add ess his issue, i is essen ial o ha e a s uc u ed da abase ha includes a comp ehensi e ocabula y. This allows o a compa ison o each wo d o de e mine i s alidi y. By employing his app oach, spelling and ocabula y e o s can be iden i ied and co ec ed e ec i ely. 2.8. Wo ds simila i y 29 The e a e mul iple ways o compa e wo ds in a compu e , and i ’s impo an o dis inguish be ween compa ing embeddings, which is used when he wo ds a e co ec ly spelled. The p ima y me hods include: Le ensh ein Dis ance [28]: used o calcula e he minimum numbe o single- cha ac e edi s (inse ions, dele ions, o subs i u ions) equi ed o ans o m one s ing in o ano he . I was named a e he So ie ma hema ician Vladimi Le ensh ein, who in oduced i in 1965. le (s 1, s 2) =                    leng h(s 2) i leng h(s 1) = 0, leng h(s 1) i leng h(s 2) = 0, le ( ail(s 1), ail(s 2)) i s 1[0] = s 2[0], 1 + min      le ( ail(s 1), s 2) le (s 1, ail(s 2)) le ( ail(s 1), ail(s 2))      o he wise (2.6) whe e he ail o a s ing, deno ed as ail(x), e e s o he subs ing ob ained by excluding he i s cha ac e om he s ing x. Ja o-Winkle Dis ance [29]: This measu e calcula es he simila i y be ween wo wo ds by aking in o accoun cha ac e ansposi ions and p e ix simila i y. To compa e embeddings, he Cosine dis ance [30] is commonly used. I measu es he simila i y be ween wo nume ical ec o s. cosine_simila i y(A,B) = A·B ∥A∥∥B∥(2.7) In o de o lea n how o calcula e he simila i y be ween wo s ings in Py hon, you can e e o my Gi Hub eposi o y a ailable a : h ps://gi hub.com/Al a o8gb/ S ings-simila i y Chap e 3 S a e o he A In ecen yea s he use o NLP in he biomedical domain has inc eased he possibili y o au oma ically ex ac ing in o ma ion om clinical na a i es [31; 32; 33]. The i s challenge o be add essed when ex ac ing in o ma ion om clinical ex s is he iden i ica ion o medical-named en i ies. Ex ac ing named en i ies is one o he mos impo an asks in he medical domain since pe o ming clinical s udies commonly equi es de ailed pa ien in o ma ion eco ded in clinical no es [34; 10]. Clinical Named En i y Recogni ion (Clinical NER) is he ask ha aims o iden i y medical concep s om clinical ex [32; 35]. Recen ly, deep lea ning-based app oaches ha e shown impo an ad ances and imp o emen s in ex ac ing in o ma ion in he biomedical domain [36; 37; 38; 39]. Howe e , mos o hese p oposals ha e ocused on he English language [36; 40]. In ac , in o ma ion ex ac ion in he medical domain ep esen s i s own challenges in languages o he han English [41; 42]. The main ad an age o deep lea ning ap- p oaches is he abili y o au oma ically lea n high-le el ea u es om ex s, educing he ime in he hand-c a ed ea u e enginee ing p ocess. The use o deep-lea ning me hods has also encou aged he ex ac ion o mo e de ailed in o ma ion ela ed o cance . Fo ins ance, in [43], he au ho s desc ibed a deep-lea ning app oach o ex ac b eas cance concep s using BERT. The goal o his p oposal is o ex ac a comp ehensi e se o b eas cance concep s om clinical no es w i en in Chinese. The au ho s demons a e ha he BERT-based model pe o ms be e han adi ional machine lea ning algo i hms a ex ac ing named en i ies in he cance ield. In [44], he au ho s desc ibe a BiLSTM-based model o clinical concep ex ac ion om oncological clinical no es w i en in Ge man. This model suppo s ex ac ing se e al concep s such as diagnosis, ea men s, and medica ions. Al hough deep lea ning-based app oaches ha e imp o ed he abili y o ex ac medical concep s in he cance medical ield, mos o hese p oposals ha e ocused on he English language [36; 40] and mos ecen ly, on Chinese [43; 45]. In he Spanish language case, in [46], he au ho s p opose Can emis , an anno- a ed co pus o suppo umo mo phology ex ac ion. Se e al s udies [47; 48; 49] ha e used his co pus o pe o m mo phology ex ac ion. Howe e , he main lim- i a ion o hese p oposals is hey only suppo iden i ying one en i y ype ( umo 31 32 Chap e 3. S a e o he A mo phology). Cance is a complex and specialized medical ield equi ing a comp e- hensi e se o medical concep s o unde s anding i s e olu ion om clinical na a- i es [43]. In conclusion, ex ac ing named en i ies om oncology clinical ex s w i en in Spanish has no been explo ed deeply ye . The e is a lack o co po a o suppo in o ma ion ex ac ion in he b eas cance domain in his language. Chap e 4 NLP Cance Pipeline 4.1. P oblem s a emen In Heal h Ca e In o ma ion Sys em, da a is p ima ily classi ied as s uc u ed o uns uc u ed. S uc u ed da a, wi h i s de ined o ma s, is eadily p ocessable by compu e sys ems. I encompasses pa ien demog aphics, diagnoses, es esul s, and p esc ibed medica ions. On he o he hand, uns uc u ed da a lacks a de ined o ma , complica ing au oma ic p ocessing and analysis. I ypically includes med- ical epo s, p og ess no es, clinical obse a ions, and medical images. This p ojec is based on he hypo hesis ha uns uc u ed da a in HCIS holds aluable in o ma ion ha , i ex ac ed and ans o med in o s uc u ed da a, could signi ican ly bene i he heal h sys ems. The cen al mo i a ion o his esea ch is o c ea e NLP pipeline ha can e ec- i ely mine uns uc u ed da a in HCIS. The o e all p ocess is shown in Figu e 4.1. Figu e 4.1: T ans o ming Clinical No es in o S uc u ed JSON 33 34 Chap e 4. NLP Cance Pipeline 4.2. S uc u ing b eas cance in o ma ion To ob ain ob ain a module ha s uc u e a Clinical no e, we ha e o do o he s eps ea lie . Fi s o all, as show in Figu e 4.2, we acqui ed wo anno a ed co pus, clinical and Nega ion & Unce ainly. Figu e 4.2: Co pus collec ion Subsequen ly, we p oceeded o ain he models using he co pus we had acqui ed ea lie and e alua ed hei pe o mance, as illus a ed in 4.3. This speci ic s ep is de ailed in chap e 5, whe e we demons a e he de elopmen o wo NER models, namely he Clinical and Nega ion & Unce ain y models. Figu e 4.3: Models c ea ion Finally, bo h models ha e been in eg a ed in o a comp ehensi e pipeline known as NCP, which se es as he inal implemen a ion. This module has he capabili y o ake a clinical no e in aw ex o ma and gene a e a JSON ile ha cap u es he mos signi ican concep s wi hin he no e. 4.3. NCP S ages 35 Figu e 4.4: NCP: NLP Cance Pipeline 4.3. NCP S ages In NLP, a s anda d wo k low in ol es mul iple s eps, each dedica ed o a unique unc ion. The speci ics o hese s ages can a y based on he p oblem and ap- plica ion a hand. Ou me hodology o s uc u ing clinical no es is depic ed in Figu e 4.4 (showcasing he inal module) and Figu e 4.5 (highligh ing he echnolo- gies employed). A specialized pipeline, e med NCP (NLP Cance Pipeline) has been de eloped. Each clinical no e a e ses h ough e e y s age o he pipeline, e en ually eme ging in a s uc u ed o ma : Clinical No es: This s age in ol es collec ing and compiling he clinical no es om a SQL Da abase. Clinical no es e e o he eco ds o pa ien s’ medical his o ies, including hei symp oms, diagnoses, ea men s and ou comes. P ep ocessing (Regex): In his s age, he clinical no es a e p ep ocessed o p epa e hem o u he analysis. The p ep ocessing s ep in ol ed a ious s eps, which a e p ima ily accomplished h ough he u iliza ion o egula ex- p essions. The ollowing s eps a e in ol ed in his s age and desc ibed in de ail in sec ion 4.4. Co pus Anno a ion (P odigy): In his s age, he clinical no es a e anno a ed wi h ele an medical en i ies desc ibed in Appendix A. I is impo an o no e ha he anno a ed no es used in he anno a ion p ocess a e a subse o he no es desc ibed in sec ion 2.2. This p ocess in desc ibed in sec ion 4.5. T ans o me s Ne wo ks (BERT): This s age in ol es he use o ans o me s 36 Chap e 4. NLP Cance Pipeline ne wo ks wi h mul ilingual uncased BERT1 o p ocess he anno a ed co pus. This ne wo k use deep lea ning echniques o lea n om he anno a ed da a and gene a e accu a e p edic ions. Pos -p ocessing (Wo ds simila i y): In his s age, he p edic ions gene a ed by he ans o me s ne wo ks a e pos -p ocessed o imp o e hei accu acy and no malize he esul s. P ima ily, he employs wo d simila i y echniques o accomplish his ask. S uc u ed Clinical No es (mongoDB): In his inal s age, he p ocessed clinical no es a e con e ed in o s uc u ed clinical no es. These no es a e o ganized in o a s anda dized o ma ha can be easily sea ched and analyzed. S uc- u ed clinical no es a e used o se e al applica ions such as clinical decision suppo , popula ion heal h managemen , and quali y imp o emen . Figu e 4.5: Technology pipeline We de eloped a pipeline o e icien ly p ocess la ge olumes o no es by ope - a ing on ba ches, as shown in Figu e 4.6. This app oach op imizes compu a ional e iciency and educes o e head associa ed wi h deep lea ning models. A p esen , he pipeline is only capable o s uc u ing no es in cases whe e he diagnoses wi hin a gi en no e a e no in g amma ical ag eemen . As an example, he clinical judgmen , "p esen a dos ca cinomas uno duc al y o o lobulilla " would no be p ope ly ex ac ed by he pipeline due o he g amma ical ag eemen be ween 1h ps://hugging ace.co/be -base-mul ilingual-uncased 4.3. NCP S ages 37 he e ms "duc al" and "lobulilla ." Al hough he no e indica es ha he pa ien has wo cance s, he pipeline only ecognizes one. In ou assessmen , each diagnosis ep esen s a dis inc loca ion o he cance . Diagnoses ha iden i y mul iple si es a e he e o e conside ed as sepa a e diagnoses. As e idenced by he ollowing no e, he pipeline will ecognize he p esence o wo dis inc diagnoses. 2 ocos: - ca cinoma mucinoso in il an e de 17 mm, g1, e nega i o, p neha i o, he 2 posi i o (3+), ki67 5% - ca cinoma mic opapila in il an e de 28mm, n+, g1, e 90%, p 90%, he 2 posi i o (3+), ki 67 10% In o de o gain a comp ehensi e unde s anding o how he pipeline unc ions, we ha e p o ided a de ailed desc ip ion in he ollowing sec ions. 1 2de ba ch(no es, no es_index): 3 4p ep ocess(no es, no es_index[" ex "]) 5 6classi ica o (no es, no es_index[" ex "]) 7 8s uc _no es =pos p ocess(no es, no es_index) 9 10 dump2mongo_db(s uc _no es) # o JSON db 11 12 i __name__ == "__main__" : 13 14 no es, no es_index =load_db() # om SQL db 15 16 ba chs =[ no es[i:i +ba ch_size] 17 o iin ange(0,len(no es), ba ch_size)] 18 19 h eads =[ Th ead( a ge =ba ch, 20 a gs=(chunk, no es_index)) 21 o chunk in ba chs ] 22 23 o in h eads: 24 .s a () 25 26 o in h eads: 27 .join() Figu e 4.6: Au oma ed pipeline in ba ches 44 Chap e 5. Expe imen s and Resul s Table 5.1: Model hype -pa ame e s Pa ame e Value Seed 8 Accumula e g adien s 3 D opou 0.1 Op imize Adam GPU alloca o Py o ch Ba ch size 2000 Ba ch size bu e 256 Ba ch size disca d o e size T ue Lea n a e wa mup-linea Wa m up s eps 250 Ini ial a e 0.00005 To al s eps 20000 Table 5.2: In o ma ion abou NER co pus Co pus Nega ion and Unce ainly Clinical NER Numbe o okens 324116 38378 Numbe o en i y okens 21091 10280 Pe cen age label okens 6.51% 26.79% dispa i y be ween he numbe o anno a ions and he unique wo ds ( ocabula y size) in igge labels (NegCue, Uce Cue) is mo e p onounced, which p o es ad- an ageous o he model’s lea ning p ocess. This di e gence ensu es an abundan supply o samples, allowing he model o e ec i ely comp ehend he ocabula y. On he o he hand, he e is minimal di e ence obse ed in scope labels (NegScope, Uce Scope), which is expec ed since nega ion and unce ain y can be exp essed in ela ion o any wo d in na u al language. This p esen s a challenging ask o he model, as i needs o accu a ely iden i y and scope wo ds ha i hasn’ encoun e ed in he majo i y o cases. In he Clinical co pus, we obse e signi ican a ia ion in he numbe o anno a ed wo ds among di e en en i ies. The en i ies wi h he highes numbe o anno a ed okens a e Molecula Ma ke and Cance Loca ion, bo h exceeding 1500. In con as , he e a e labels such as Cance In a ype o T ea men F equency, which ha e ewe han 250 anno a ed okens. In addi ion, he Implici Da e label in he clinical co pus is pa icula ly challenging o p edic , alongside he scope labels o Nega ion and Unce ain y. The challenge s ems om he wide ange o g amma ical exp essions u ilized o communica e empo al in o ma ion, o en inco po a ing a ious o ms o empo al ad e bs. These ad e bs can be exp essed in nume ous ways, adding o he complexi y o p edic ing he Implici Da e label in he clinical co pus. Fo example, he occu ence o an e en ha happened yes e day can be exp essed in a ious ways, such as "Aye " (Yes e day), "El día an e io " (The p e ious day), "Un día a ás" (One day ago), "Hace un día" (A day ago) o "En la íspe a" (On he e e). 5.3. Expe imen s 45 Figu e 5.1: G aphical ep esen a ion o Neg Unce Co pus coun ing 5.3. Expe imen s To assess he pe o mance o he p oposed app oach, we employed he well- es ablished s anda d me ics, P ecision, Recall, and F-Sco e ( e e o subsec ion 2.7.1). These me ics we e selec ed due o he i al impo ance o Recall in NER. Inco ec ly iden i ying an en i y can esul in he p opaga ion o alse in o ma ion. The e o e, i is p e e able o e ain om anno a ing an en i y when in doub , a he han anno a ing i inco ec ly, i.e. a high Recall is be e han a high P ecision. To assess he e ec i eness o he Clinical model, 10- old c oss alida ion (see subsec ion 2.7.2) has been selec ed, due o he low numbe o samples om some en i ies. The pe o mance was calcula ed as he a e age o all en olds execu ed by he c oss- alida ion s a egy. To de e mine he ep esen a ion o he media, we calcula e he s anda d de ia ion. The s anda d de ia ion is a measu e o he dis- pe sion o a iabili y in a da a se . A lowe s anda d de ia ion indica es ha he media is mo e e enly ep esen ed h oughou he da a se , sugges ing a balanced dis ibu ion. On he o he hand, a highe s anda d de ia ion implies ha he occu - ence o media ins ances is mo e a iable, indica ing po en ial biases o imbalances in ep esen a ion. In o de o e alua e he pe o mance o he nega ion and unce ain y model, we employed an 80-20 spli . Due o he e enough ep esen a ion o each en i y. This means ha he da a se was di ided in o wo pa s: 80% o aining and 20% o es ing. I is impo an o no e ha spli s a e made in a ie ed manne o ensu e su icien ep esen a ion o each en i y. 46 Chap e 5. Expe imen s and Resul s Figu e 5.2: G aphical ep esen a ion o Clinical co pus coun ing 5.4. Resul s This sec ion p esen s he esul s o he expe imen s desc ibed in he p e ious sec ion. The esul s o he Nega ion and Unce ain y NER model, which ocuses on he iden i ica ion o ou en i ies, a e shown in Table 5.3. Besides, we p esen he esul s ob ained om he Clinical NER model ained on he b eas cance co pus, which includes 21 en i ies. Table 5.4 displays he o e all esul s achie ed by Clinical model. Following ha , we will p o ide he pe o mance me ics o each indi idual en i y, as shown in Table 5.5. 5.4. Resul s 47 Table 5.3: Resul s om Nega ion and Unce ainly Co pus pe en i y ype En i y Type P ecision Recall F-Sco e Nega ion Concep 0.956 0.956 0.956 Nega ion Scope 0.836 0.823 0.829 Unce ainly Concep 0.874 0.846 0.860 Unce ainly Scope 0.748 0.708 0.728 O e all 0.873 0.859 0.866 Table 5.4: Gene al esul s om Clinical Co pus Me ic Mean ±De ia ion P ecision 0.9346 ±0.0123 Recall 0.9366 ±0.0083 F-sco e 0.9356 ±0.0099 Table 5.5: Resul s om Clinical co pus pe en i y ype En i y Type P ecision Recall F-Sco e Cance Concep 0.9806 ±0.0171 0.9732 ±0.0143 0.9767 ±0.0076 Cance Expansion 0.9801 ±0.0173 0.9796 ±0.0222 0.9798 ±0.0178 Cance G ade 0.9325 ±0.0309 0.9373 ±0.0281 0.9347 ±0.0261 Cance In a ype 0.963 ±0.1048 0.9798 ±0.0571 0.9667 ±0.0667 Cance Loca ion 0.9117 ±0.0324 0.9174 ±0.0268 0.9143 ±0.0255 Cance Me as asis 0.9055 ±0.0284 0.9014 ±0.0375 0.903 ±0.0275 Cance Recu ence 1.0±0.0 0.9815 ±0.0524 0.9899 ±0.0286 Cance S age 0.9576 ±0.0341 0.9442 ±0.0412 0.9501 ±0.0265 Cance Sub ype 0.7773 ±0.1789 0.7956 ±0.138 0.7675 ±0.1092 Cance Type 0.9789 ±0.0133 0.9822 ±0.0144 0.9804 ±0.0092 Da e 0.9596 ±0.029 0.969 ±0.0271 0.9641 ±0.0254 Implici Da e 0.7435 ±0.1084 0.7368 ±0.1123 0.7392 ±0.1064 Molec Ma ke 0.9343 ±0.0204 0.9506 ±0.0138 0.9423 ±0.016 Su ge y 0.9296 ±0.0551 0.9328 ±0.0386 0.931 ±0.0455 TNM 0.9019 ±0.0499 0.8916 ±0.0526 0.8965 ±0.0494 T ea men 0.9321 ±0.0336 0.9235 ±0.0547 0.9268 ±0.0356 T ea men D ug 0.9289 ±0.045 0.9397 ±0.0398 0.9335 ±0.0327 T ea men F equency 0.9841 ±0.0449 0.8889 ±0.1757 0.9221 ±0.1109 T ea men In e al 0.9348 ±0.0458 0.898 ±0.0481 0.9148 ±0.0319 T ea men Quan i y 0.9008 ±0.1648 0.8581 ±0.1692 0.8673 ±0.1463 T ea men Schema 0.868 ±0.1098 0.9196 ±0.0823 0.8886 ±0.0749 48 Chap e 5. Expe imen s and Resul s 5.5. Discussion The esul s p esen ed in Table 5.3 o he Nega ion and Unce ain y Co pus p o ide insigh s in o he pe o mance o he p oposed app oach o each en i y ype: Fo Nega ion Concep , he model achie ed a p ecision o 0.956, ecall o 0.956, and an F-sco e o 0.956. This indica es a high le el o accu acy and comple e- ness in iden i ying nega ion concep s. In e ms o Nega ion Scope, he p ecision was measu ed a 0.836, ecall a 0.823, and F-sco e a 0.829. Al hough sligh ly lowe han he p ecision and ecall o Nega ion Concep , he model s ill exhibi ed easonably good pe o - mance in ecognizing he scope o nega ion. Mo ing on o Unce ain y Concep , he model ob ained a p ecision o 0.874, ecall o 0.846, and an F-sco e o 0.860. These esul s sugges ha he e a e ins ances whe e he model does no make accu a e p edic ions. In he case o Unce ain y Scope, he p ecision sco e was 0.748, ecall was 0.708, and he F-sco e was 0.728. The p ecision and ecall alues o his en i y ype a e he lowes among he o he s, causing equen con usion o he model. Conside ing he o e all pe o mance, he p oposed app oach alls sho wi h an a e age p ecision o 0.873, ecall o 0.859, and an F-sco e o 0.866. This indica es ha he e is oom o imp o emen in accu a ely iden i ying nega ion and unce ain y concep s, as well as hei espec i e scopes, ac oss all en i y ypes in he Nega ion and Unce ain y Co pus. The unde -pe o mance could po en ially be a ibu ed o he model being ained on di e en ypes o no es. These esul s sugges ha he model’s pe o mance in his a ea needs o be enhanced o ensu e a mo e e ec i e implemen a ion o he p oposed app oach. The esul s ob ained om he Clinical Co pus in Table 5.4 indica e a high le el o pe o mance o he p oposed app oach wi h p ecision, ecall, and F-sco e sco es consis en ly abo e 0.93. This sugges s ha he model is capable o accu a ely iden- i ying en i ies while cap u ing a subs an ial po ion o he ele an en i ies p esen in he da a. The esul s ob ained om he Clinical Co pus, as displayed in Table 5.5, p o ide aluable insigh s in o he pe o mance o he p oposed app oach o each en i y ype: Fo he Cance Concep en i y ype, he model achie ed a high p ecision o 0.9806, ecall o 0.9732, and an F-sco e o 0.9767. These esul s sugges ha he model demons a es excellen accu acy and comple eness in iden i ying cance concep s wi hin he clinical da a. Simila ly, o he Cance Expansion en i y ype, he model achie ed a p ecision o 0.9801, ecall o 0.9796, and an F-sco e o 0.9798. These sco es indica e a high le el o pe o mance in ecognizing expanded cance e minologies. 5.5. Discussion 49 Table 5.6: Discussion o Clinical NER model En i y F-sco e IV Vocabula y size Cance Concep 0.98 1.00 32 Cance Expansion 0.98 1.00 25 Cance G ade 0.93 0.99 49 Cance In a ype 0.97 0.98 6 Cance Loca ion 0.91 0.98 249 Cance Me as asis 0.90 0.99 31 Cance Recu ence 0.99 0.98 6 Cance S age 0.95 0.99 35 Cance Sub ype 0.77 0.88 26 Cance Type 0.98 1.00 24 Da e 0.96 0.45 540 Implici Da e 0.74 0.33 123 Molecula Ma ke 0.94 0.87 620 Su ge y 0.93 0.99 88 TNM 0.90 0.38 467 T ea men 0.93 0.97 58 T ea men D ug 0.93 0.95 110 T ea men F equency 0.92 0.92 12 T ea men In e al 0.91 0.88 81 T ea men Quan i y 0.87 0.59 36 T ea men Schema 0.89 0.96 19 The Cance G ade en i y ype achie ed a p ecision o 0.9325, ecall o 0.9373, and an F-sco e o 0.9347. These esul s indica e sa is ac o y pe o mance in iden i ying cance g ades wi hin he clinical da a. Rega ding he Cance In a ype, Cance Loca ion, Cance Me as asis, Cance Recu ence, Cance S age, Cance Sub ype, and Cance Type en i y ypes, he model exhibi ed a ying le els o p ecision, ecall, and F-sco e. O e all, he model achie ed ela i ely high sco es, sugges ing i s capabili y o cap u e and iden i y hese cance - ela ed en i ies e ec i ely. Fo en i ies such as Da e, Implici Da e, Molec Ma ke , Su ge y, TNM, T ea - men , T ea men D ug, T ea men F equency, T ea men In e al, T ea men Quan i y, and T ea men Schema, he model demons a ed easonable pe o - mance, as indica ed by he p ecision, ecall, and F-sco e sco es. Examining he pe o mance o di e en en i ies, we ound ha ags wi h low sup- po , such as T ea men Schema, T ea men Quan i y, T ea men In e al, T ea - men F equency, and Cance Sub ype, exhibi ed poo e esul s. Howe e , ags wi h a limi ed ocabula y, such as T ea men o Cance Recu ence, demons a ed highe accu acy, likely due o he model’s abili y o easily lea n hem. We p esen he co - ela ion be ween he IV (Index o Va iabili y) and ocabula y size a iables wi h he 50 Chap e 5. Expe imen s and Resul s F-sco e in Table 5.6. The esul s show ha , in mos cases, a highe IV co esponds o be e classi ica ion pe o mance. Fu he mo e, we no ed ha he Da e label consis en ly ou pe o med he Implici Da e label. The Da e label, being sho e and mo e s uc u ed, ypically included speci ic da es such as "23/02" o "24/03/2023," while he Implici Da e label o en con ained ph ases like " omo ow in he mo ning". In p ac ice, egula exp essions a e commonly employed o ex ac en i ies like Da e o TNM, le e aging hei sha ed s uc u al pa e ns. In o de o ackle he challenges and enhance he o e all pe o mance o he models and hei unc ionali y wi hin he pipeline, we p opose he ollowing ecom- menda ions: Conduc addi ional anno a ion o en i ies ha exhibi signi ican de ia ions. This will in ol e ca e ully iden i ying and labeling ins ances whe e he models s uggle o p oduce inaccu a e esul s. Anno a e he Nega ion and Uce anly labels using he b eas cance co pus. By le e aging a specialized da ase speci ic o b eas cance , we can ine- une he model o be e unde s and and handle nega ion and unce ain y ela ed o his pa icula domain. This a ge ed e aining will lead o imp o ed pe - o mance and accu acy when dealing wi h b eas cance - ela ed in o ma ion. Explo e he inco po a ion o egula exp ession o imp o e en i y ex ac ion in s uc u ed pa e ns such as Da e o TNM, in conjunc ion wi h deep lea ning models. By implemen ing hese ecommenda ions, we can add ess he challenges aced by he models, imp o e hei pe o mance, and enhance he o e all unc ionali y o he pipeline. To es ablish he o de o p io i ies, i is ad isable o ollow he Amdahl’s Law [51], which emphasizes imp o ing modules wi h b oade global impac . In ou case, he Nega ion and Unce ain y model should ake p ecedence since i s e oneous de ec ion leads o he ejec ion o clinical e ms p edic ed by he Clinical NER model. Chap e 6 Conclusions and Fu u e Wo k This p ojec in oduces a deep lea ning-cen ic me hodology o s uc u ing he ex ac ion o b eas cance in o ma ion. Ou p oposal is an au oma ed pipeline, called NCP -NLP Cance Pipeline-, capable o p ocessing clinical no es in Spanish and p oducing a s uc u ed JSON ile. The esul s o his app oach unde sco e he necessi y o clinical NER and he de ec ion o nega ion and unce ain y as he main asks in in o ma ion s uc u ing wi hin he clinical domain. Le e aging deep lea ning-based me hods o he ex ac ion o medical concep s yielded p omising ou comes in p ocessing Spanish b eas cance clinical na a i es. We p oposed an anno a ion scheme Table 4.1, which aims o ex ac de ailed and speci ic concep s o cance diagnosis and ea men s. Thus, his anno a ion schema is aimed a ex ac ing medical concep s and s uc u ing hem. The deep lea ning- based model ained wi h mul ilingual BERT has also shown p omising esul s in de ec ing nega ion and unce ain y in Spanish clinical ex s. Speci ically, we used he scope o nega ion and unce ain y o ex ac ing medical concep s a ec ed by hese linguis ic phenomena. 6.1. Deep Lea ning s. REGEX Machine lea ning models a e conside ed be e o NER han egex, because hey can lea n mo e complex pa e ns and ela ionships in ex da a, and can be e adap o di e en con ex s and a ia ions in he da a. In con as , egula exp essions ely on p ede ined ex pa e ns and a e less lexible in e ms o adap ing o di e en a ia ions in he da a. Howe e , i is impo an o no e ha egula exp essions a e mo e in e p e able1 han deep lea ning ne wo ks o his eason egex is s ill widely used. The p oblem wi h egula exp essions is ha i can be ime-consuming and di icul o de ine pa e ns ha accu a ely ma ch all possible a ia ions o named en i ies. Addi ionally, i may no be able o cap u e named en i ies ha do no 1A model is said o be in e p e able when i is easy o humans o unde s and and explain how i makes p edic ions o decisions based on he inpu da a. 51 52 Chap e 6. Conclusions and Fu u e Wo k ollow a speci ic pa e n o s uc u e. This p esen s a p oblem ha I e e o as " he egex middle poin ". This issue a ises when he ex you’ e ying o ex ac has nume ous a ia ions, and you need a egex ha migh ex ac wo ds you don’ wan . This p oblem e lec s he di icul y o making a egex ha i s exac ly wha you wan . I ’s ecommended o conside he sou ce o he ex be o e deciding on an ex- ac ion me hod. I he ex is machine-gene a ed, egula exp essions ( egex) migh o en be adequa e. Howe e , o mo e complex o na u al language ex , aining a NER model based on deep lea ning could be mo e e ec i e. Fo u he in o - ma ion and guidance on when o employ egex o a deep lea ning app oach, e- e o he ollowing guide: h ps://gi hub.com/explosion/asse s/blob/main/ P odigy/P odigy_NER_ lowcha _ 2_0_0_ligh .pd . 6.2. Fu u e wo k In u u e esea ch s udies, he e a e se e al main a enues o u he explo a ion and in es iga ion in o a ious aspec s o he p oblem. These include: Conside in es iga ing al e na i e language models designed o p ocessing Spanish clinical ex , such as RoBERTa2o RoBERTa Clinical3. Ha nessing he powe o models ha ha e a be e adap a ion o Spanish clinical no es could po en ially enhance he accu acy and e ec i eness o he esul s. Conside cons uc ing ou own language model. Le e aging a p e-exis ing Spanish model as a ounda ion, we can engage in u he p e- aining u ilizing a specialized medical co pus wi h an emphasis on oncology-speci ic language, including con en om scien i ic esea ch jou nals wi hin his domain. This model, once comple ed, could be eleased o he public, ma king a signi ican miles one as he i s Spanish language model wi h a specializa ion in medical cance ocabula y. Enhancing he pe o mance o models ha handle nega ion and unce ain y is c ucial in he ealm o clinical na a i e analysis in spanish. The cu en pipeline is designed o s uc u e no es ha con ain a single diag- nosis. To e ec i ely comp ehend and p ocess no es wi h mul iple diagnoses, an e ec i e app oach could be o u ilize ano he Language Model (LM). This LM could ini ially p ep ocess he no es, conduc ing a p ima y cleanup o he da a. I mul iple diagnoses a e de ec ed, he LM could hen segmen he ex in o dis inc pa ag aphs co esponding o each diagnosis. 2h ps://hugging ace.co/BSC-LT/ obe a-base-bne 3h ps://hugging ace.co/PlanTL-GOB-ES/bsc-bio-eh -es Bibliog aphy Imagina ion is mo e impo an han knowledge. Knowledge is limi ed, while imagina ion is no . Albe Eins ein [1] B. S ini asa-Desikan, Na u al Language P ocessing and Compu a ional Linguis- ics: A p ac ical guide o ex analysis wi h Py hon, Gensim, spaCy, and Ke as. Pack Publishing L d, 2018. [2] C. Allen and T. Hospedales, “Analogies explained: Towa ds unde s anding wo d embeddings,” in In e na ional Con e ence on Machine Lea ning. PMLR, 2019, pp. 223–231. [3] C. Olah. (2015) Unde s anding ls m ne wo ks. [Online]. A ailable: h ps: //colah.gi hub.io/pos s/2015-08-Unde s anding-LSTMs/ [4] Hugging Face, “Hugging ace blog: La ge language model,” h ps://hugging ace. co/blog/la ge-language-models, 2021. [5] A. Zhang, Z. C. Lip on, M. Li, and A. J. Smola, “Di e in o deep lea ning,” a Xi p ep in a Xi :2106.11342, 2021. [6] J. De lin, M. W. Chang, K. Lee, and K. Tou ano a, “BERT: P e- aining o deep bidi ec ional ans o me s o language unde s anding,” NAACL HLT 2019 - 2019 Con e ence o he No h Ame ican Chap e o he Associa ion o Compu a ional Linguis ics: Human Language Technologies - P oceedings o he Con e ence, ol. 1, no. Mlm, pp. 4171–4186, 2019. [7] J. De lin, M.-W. Chang, K. Lee, and K. Tou ano a, “BERT: p e- aining o deep bidi ec ional ans o me s o language unde s anding,” CoRR, ol. abs/1810.04805, 2018. [Online]. A ailable: h p://a xi .o g/abs/1810.04805 [8] R. L. Siegel, K. D. Mille , H. E. Fuchs, and A. Jemal, “Cance s a is ics, 2022,” CA: a cance jou nal o clinicians, 2022. 53 60 Appendix A. Co pus Anno a ion Guide 2. Implici Da e This label is used o ep esen da es ha a e exp essed indi ec ly in he no es h ough ime exp essions Ing esado a u gencias hace 5 días. T a ado con una ci ugía hace 3 años. a amien o neodyud an e has a agos o 2015. ing esada desde 17/08/2022 al 4/09/2020. 3. Cance Concep This label is used o ep esen he di e en ways o e e ing o he e m cance in clinical no es. Mainly e e s o he classi ica ion acco ding o he ype o cell ha becomes cance ous. Mos commons examples: Ca cinoma de mama de echa mul i ocal. Diagnos icada de adenoca cinoma de pulmón de echo. Neoplasia de mama . p esen a nec osis umo al. O he s e ms: ib oma, ib o ecoma, ib oadenoma, mioma, lin oma, sa - coma, leucemia. 4. Cance Expansion ca cinoma de mama de echa mul i ocal. ca cinoma duc al in il an e mama de echa. ca cinoma de mama lobulilla in si u. O he s e ms: in asi o, mic oin il an e, mul icén ico, bi ocal o in il- an e mul i ocal. 5. Cance Type This label ep esen s he subca ego ies o cance acco ding o hei dis inc i e ea u es and cha ac e is ics. Each sub ype may ha e di e en g ow h pa e ns, clinical beha iou , esponse o ea men and p ognosis. The e a e he mos impo an ypes o doc o s. ca cinoma de mama lobulilla in si u. ca cinoma de mama in aduc al. ca cinoma in il an e ipo his ologico no especi ico. O he s e ms: lobula , duc al, apoc ino, neo oendoc ino, endome ioide, adenoide, inespecí ico, en e medad de page , cis oadenoma, iloides. 6. Cance Sub ype This label ep esen s he o he subca ego ies o cance . 61 ca cinoma duc al con encional in il an e mama de echa. ca cinoma mucinoso de mama de echa. O he s e ms: clásico/con encional, medula , papila , ubula , mucosi- no/coloide/mucinoso/mucoso/mucinosa, comedoniano, sólido, c ibi o me, mic opapila , plano, pleomó ico/a, olicula . 7. Cance G ade This label e e s o he g ade o he cance . I may be nuclea o No ingham g ade. In addi ion, ei he he g ade numbe o he exp ession. ca cinoma in il an e inespecí ico g ado his ológico 1 de no ingham po pun uación de 5 (2+2+1). ca cinoma de mama lobulilla in il an e, g2, e 100%. con componen e in aduc al asociado de al o g ado nuclea . 8. Cance In a ype The e a e 4 molecula sub ypes o b eas cance , de ined by he molecula ma ke s. luminal A luminal B HER2 sob eexp esado T iple nega i o 9. Cance Loca ion This label exp esses he o gan whe e he cance mani es s i sel . Ca cinoma lobulilla in il an e en mama izquie da. Ca de mama izquie da cuad an e supe io . ca lobulilla o igen mama io. ca cinoma lobulilla bila e al. mas ec omía de mama de echa. cánce de pulmón. cánce de colon. 10. Cance Me as asis This label includes e ms e e ing o he sp ead o cance o o he o gans. ganglios lin a icos con me as asis. p og ession nodula . in asión pe ineu al. 62 Appendix A. Co pus Anno a ion Guide in il ación lin o ascula . 11. Cance Recu ence This label indica es whe he he pa ien has elapsed a e e adica ion o he cance . Recaída local de echa de ca cinoma lobulilla en no iemb e de 2016. Recidi a local de ca cinoma duc al in il an e de mama de echa. 12. Cance S age This label is used o ep esen he s age o he cance . Mos cance s ha e ou s ages: s ages 0, I (1) o IV (4). S age I ep esen s he ea ly s age and s age IV ep esen s he mos ad anced s age o he umou . These s ages can also be combined wi h he le e s A, B o C as shown below. Adenoca cinoma de mama es adio IV. Ca cinoma escamoso es adio I-B. Cánce de pulmón no mic ocí ico es adio IA. Ca cinoma de mama al menos es adio iii. 13. TNM This label is used o ep esen he s age o he cance using he malignan u- mou classi ica ion: TNM no a ion. This no a ion ep esen s a cance concep using h ee alphanume ic codes: Tumou (T) desc ibes he size o he umou . The le e (N) desc ibes he lymph nodes ha a e a ec ed. The le e (M) is used o ep esen whe he me as ases a e p esen . TNMs a e neoadju an ea men s a wi h y. Pacien e con cánce de pulmón cT3cN3cM1. Ca cinoma de mama T3 N2 M0. 14. Molec Ma ke This label e e s o molecula ma ke s ha a e measu ed in pa ien s, and whose esul can be posi i e, nega i e o a pe cen age alue. Thei s udy p o ides in o ma ion abou he cance and he ea men o be ollowed. This will include: Ho mone ecep o s: oes ogen ( e o e ), p oges e one ( p o p ) o bo h ho mone ecep o s ( h o hh). ki67. Shall be accompanied by a pe cen age. he 2/Neu/E B2/Ce B2. Accompanied by he sign/pending/inconclusi e. Examples: 63 ca cinoma duc al in il an e mama de echa e y p posi i os. adenoca cinoma e 90%, p 95%, he 2 -, ki-67 1%. 15. T ea men This label includes all ypes o ea men ha can be gi en o he pa ien , wi hou going in o any de ail, i.e. only gene ic ea men wo ds. Inicia a amien o con quimio e apia. T a amien o conse ado . T a amien o con ho mona. T a amien o adio e ápico. Le han p opues o quimio e apia que la en e ma ha ehusado. Tiene p e- is o ho mono e apia as la i adiación. a ada con adio e apia adyu an e. 16. T ea men D ug This label is used o iden i y speci ic d ug names used o ea cance pa ien s. Pacien e diagnos icado con ca cinoma, a ado con cispla ino en julio de 2017. Ho mono e apia con anas ozol. En a amien o con he cep ín. Inicia ensayo clínico con axol. a amien o con inhibido es de la a oma asa. 17. T ea men F equency This label e e s o he equency wi h which he ea men is applied. Pacien e que ecibe QT CON Cispla ino 75 mg/m2 cada 21 días. La dosis o al adminis ada ha sido de 50 gy día. 18. T ea men In e al This label e e s o he o al in e al o e which a gi en ea men is admin- is e ed. iene pla ini icado adio e apia y gose elina, amoxi eno du an e 5 anos. po 4 ciclos. xol x4 semanal. 19. T ea men Quan i y I is used o ep esen he dose o a medicine adminis e ed o a pa ien a a gi en ime. 64 Appendix A. Co pus Anno a ion Guide Pacien e que ecibe q con cispla ino 75 mg/m2 cada 21 días. Recibe RT adical de 50 Gy el 15/12/2014. Radio e apia de 60 gy. 20. T ea men Schema This label e e s o egimens o schedules used in chemo he apy o he ea - men o cance . Possible examples include: EP (e oposide and cispla in) FEC ( Fluo ou acil, epi ubicin hyd ochlo ide and cyclophosphamide) AC ( Ad iamycin and cyclophosphamide) ET (e oposide and ca bopla in) ICE (i os amide, ca bopla in and e oposide) CAV (cyclophosphamide, doxo ubicin o ad iamycin and inc is ine). CMF (cyclophosphamide, me ho exa e, luo ou acil). FOLFOX •FOL ( olinic acid o leuco o in) •F ( luo ou acil) •OX (oxalipla in) Examples: a ada con quimio e apia cm . 21. Su ge y This label is used o label cance - ela ed su gical p ocedu es. Examples: Lin adenec omía supe io izquie da. Mas ec omía simple en mama de echa. Realizada umo ec omía en mama de echa. Recons ucción con do sal ancho y p o esis. ci ugía conse ado a con lin adenec omía. mas ec omía de econs ucción de la mama de echa. Appendix B Lis o medical ac onyms This is he lis o ac onyms used in he p e-p ocessing o clinical no es. I is impo an o no e ha hese ac onyms a e no uni e sal, bu a he ha a e speci ic o he doc o s wo king in he hospi al. The lis is he ollowing: ca cance adio e apia ec o oquimio e apia ex aco po ea q quimio e apia h ho mono e apia cse cuad an e supe io ex e no csi cuad an e supe io in e no cie cuad an e in e io ex e no cii cuad an e in e io in e no lsi lobulo supe io izquie do lsd lobulo supe io de echo dcha de echa de de echa izq izquie da ccee cuad an es ex e nos ccss cuad an es supe io es cis ca cinoma duc al in si u 65 66 Appendix B. Lis o medical ac onyms cdi ca cinoma duc al in si u cdis ca cinoma duc al in si u nos inespeci ico ns inespeci ico bcg biopsia del ganglio cen inela bsgc biopsia selec i a del ganglio cen inela bgc biopsia ganglio cen inela bscg biopsia selec i a del ganglio cen inela il in il acion lin o ascula ipn in asion pe ineu al ihq inmunohis oquimica op ope acion gc ganglio cen inela o a amien o neoady neoadyu an e ady adyu an e ap ana omia pa ologica mi econ uccion mama ia in e na x adiog a ia hh ecep o es ho monales d x doce axel p e pendien e Lis o ac onyms EHR: Elec onic Heal h Reco ds, His o ias Clínicas Elec ónicas NLP: Na u al Language P ocessing, P ocesamien o del Lenguaje Na u al NER: Named En i y Recogni ion, Reconocimien o de En idades Nomb adas Clinical NER: Clinical Named En i y Recogni ion, Reconocimien o de En i- dades Clínicas POS: Pa e de la o ación, Pa -o -Speech CLI: Command Line In e ace, In e az de Línea de Comandos BRAT: B a Rapid Anno a ion Tool, He amien a de Ano ación Rápida B a CONLL: Con e ence on Compu a ional Na u al Language Lea ning, Con e - encia sob e Ap endizaje Compu acional del Lenguaje Na u al WHO: Wo ld Heal h O ganiza ion, O ganización Mundial de la Salud AHIMA: Ame ican Associa ion o Medical Reco d Lib a ians and Heal h In- o ma ion Managemen , Asociación Ame icana de Biblio eca ios de Regis os Médicos y Ges ión de In o mación de Salud SERMAS: Se icio Mad ileño de Salud, Mad id Heal h Se ice HCIS: Heal h Ca e In o ma ion Sys em, Sis ema de in o mación de a ención médica HP-HCIS: Heal hca e P o ide - Heal h Ca e In o ma ion Sys em, P o eedo de se icios sani a ios - Sis ema de in o mación sani a ia SELENE: Elec onic Medical Reco d Sys em, Sis ema Elec ónico de Expe- dien e y Regis os Médicos IMDH: IMDH, In e cambio de In o mación Médica y de Salud LSTM: Long Sho -Te m Memo y, Memo ia a Co o Plazo BiLSTM: Bidi ec ional Long Sho -Te m Memo y, Memo ia a Co o Plazo Bidi eccional 67 BERT: Bidi ec ional Encode Rep esen a ions om T ans o me s, Rep esen a- ciones de Codi icado Bidi eccional de T ans o mado es BETO: BERT o Spanish, BERT pa a el Español RoBERTa: Robus ly op imized BERT app oach, BERT Robus amen e Op- imizado GPT: Gene a i e P e- ained T ans o me , T ans o mado P e-en enado Gen- e a i o MLM: Masked Language Modeling, Modelado lingüís ico con másca a NSP: Nex sen ence p edic ion, P edicción de la siguien e o ación LM: Language Model, Modelo del Lenguaje CS: Compu e Science, Ciencia de los Da os AI: A i icial In elligence, In eligencia A i cial CSV: Comma-Sepa a ed Values, Valo es sepa ados po comas JSON: Ja aSc ip Objec No a ion, No ación de obje os de Ja aSc ip Neu al ne wo ks a e a powe ul ool ha can help ans o m heal h ca e and imp o e he quali y o li e o people a ound he wo ld Jensen Huang, ounde and CEO o NVIDIA