A Pipeline for Automated Phenotype Extraction from Medical Reports using Large Language Models

✦ 1 / 4 ✦ C l i n ica l R e port P he notyp e E xtr ac t i on P i p e l i n e T hi s r e pos i tory c ont ai ns a s e t o f s c r i pts to b u i l d a n e n d - to - e n d p i p e l i n e f or pro ce ss i n g c l i n ica l r e ports , e xtr ac t i n g stru c tur ed i n f orm a t i on w i t h LLM s , a n d m a pp i n g p he notyp e s to t he H um a n P he notyp e O ntolo g y ( HPO ). T he m ai n ca p abi l i t ie s a r e : OCR + tr a nsl a t i on o f PDF c l i n ica l r e ports . T e xt fi lt e r i n g to kee p only p a t ie nt - r e l e v a nt i n f orm a t i on . P he notyp e e xtr ac t i on us i n g a n LLM ( v ia O ll a m a ). M a pp i n g e xtr ac t ed p he notyp e s to HPO t e rms us i n g e m beddi n g s a n d f uzzy m a t chi n g . E xtr ac t i on o f diag nos i s c o de s f rom r e ports us i n g OMIM . A highe r - l e v e l p i p e l i n e s c r i pt to pro ce ss .docx r e ports e n d to e n d . R e pos i tory S tru c tur e . ├── Pipeline_Script.sh ├── start_ollama.sh ├── DataExtractorPipeline.py ├── PDFTranslator.py ├── PDFTextExtractor.py ├── PhenotypeExtractor.py ├── HPOMapper.py ├── WordTextExtractor.py ├── InfoExtractor.py ├── DiagnosisExtractor.py ├── utils.py ├── concepts/ ├── prompts_and_config/ └── OutputFiles/ R e qu i r e m e nts P yt h on 3.9+ O ll a m a i nst a ll ed a n d runn i n g T e ss e r ac t OCR P yt h on p ackage s : deep_translator==1.11.4 docx==0.2.4 Flask==3.1.2 fuzzywuzzy==0.18.0 InquirerPy==0.3.4 ✦ 2 / 4 ✦ json_repair==0.48.0 matplotlib==3.10.8 more_itertools==10.8.0 numpy==2.3.5 ollama==0.6.1 pandas==2.3.3 pdf2image==1.17.0 pronto==2.7.0 psycopg2==2.9.11 PyJWT==2.10.1 pytesseract==0.3.13 python_Levenshtein==0.27.3 Requests==2.32.5 scikit_learn==1.8.0 torch==2.8.0 transformers==4.55.0 S t a rt i n g t he LLM B acke n d ./start_ollama.sh # or ollama serve I n our e xp e r i m e nts , w e us e t he gpt-oss:120b mo de l , w hich i s a v ai l ab l e on O ll a m a H u b a n d , ba s ed on our pr e l i m i n a ry t e sts , p e r f orms w e ll f or p he notyp e e xtr ac t i on t a s k s . M ai n P yt h on mo d ul e s I n t he f ollow i n g , w e r e port t he m ai n mo d ul e s t ha t ca n be us ed i n de p e n de ntly or a s p a rt o f t he e n d - to - e n d p i p e l i n e , a n d w e r e port t hei r m ai n i nputs a n d outputs . PDFTranslator.py : T r a nsl a t e s a n d e xtr ac ts t e xt f rom PDF c l i n ica l r e ports us i n g OCR . M ai n i nputs : PDF fi l e p a t h P rompt fi l e f or fi l e tr a nsl a t i on ( b y defa ult prompts_and_config/translation_prompt.txt ) M ai n outputs : T r a nsl a t ed t e xt fi l e ( b y defa ult i n OutputFiles/ClinicalReport_TRANSLATED.txt ) PDFTextExtractor.py : F i lt e rs t e xt f rom PDF fi l e s b y r e mov i n g non - p a t ie nt - r e l a t ed i n f orm a t i on . M ai n i nputs : T r a nsl a t ed t e xt fi l e ( output o f PDFTranslator.py ) M ai n outputs : F i lt e r ed t e xt fi l e ( b y defa ult i n OutputFiles/ClinicalReport_FILTERED.txt ) PhenotypeExtractor.py : E xtr ac ts p he notyp e s f rom t e xt us i n g a n LLM . M ai n i nputs : ✦ 3 / 4 ✦ T e xt fi l e w i t h c l i n ica l r e port ( b y defa ult i n OutputFiles/ClinicalReport_FILTERED.txt ) P rompt fi l e f or p he notyp e e xtr ac t i on ( b y defa ult prompts_and_config/phenotypes_prompt.txt ) M ai n outputs : JSON fi l e w i t h e xtr ac t ed p he notyp e s ( b y defa ult i n OutputFiles/phenotypes.json ) HPOMapper.py : M a ps e xtr ac t ed p he notyp e s to HPO t e rms us i n g e m beddi n g s a n d f uzzy m a t chi n g . M ai n i nputs : JSON fi l e w i t h e xtr ac t ed p he notyp e s ( output o f PhenotypeExtractor.py ). I f not prov ided , t he i nput fi l e i s a ssum ed to be OutputFiles/phenotypes.json M ai n outputs : JSON fi l e w i t h m a pp ed HPO t e rms ( b y defa ult i n OutputFiles/phenotypes_with_score.json ) DataExtractorPipeline.py : H igh - l e v e l p i p e l i n e to pro ce ss .docx r e ports e n d to e n d . M ai n i nputs : t he p a t h to a .docx fi l e or f ol de r c ont ai n i n g mult i pl e .docx fi l e s M ai n outputs : JSON fi l e s w i t h e xtr ac t ed p he notyp e s a n d m a pp ed HPO t e rms ( b y defa ult i n OutputFiles/ ), w i t h n a m e < ca s e _ id >_ p he notyp e s . j son T yp ica l W or kf lows 1. PDF → P he notyp e s → HPO T hi s s c r i pt pro ce ss e s a c l i n ica l r e port i n PDF f orm a t , e xtr ac ts p he notyp e s us i n g a n LLM , a n d m a ps t he m to HPO t e rms . I t r e qu i r e s two a r g um e nts : t he p a t h to t he c l i n ica l r e port PDF a n d t he n a m e o f t he O ll a m a mo de l to us e . ./Pipeline_Script.sh <clinical_report.pdf> <ollama_model> 2. DOC X E n d - to - E n d P i p e l i n e T hi s p i p e l i n e pro ce ss e s ge n e t ic r e ports i n .docx f orm a t e n d to e n d . I t i n c lu de s t e xt e xtr ac t i on , p he notyp e e xtr ac t i on , a n d HPO m a pp i n g . T o m ake i t wor k , i t i s n eeded to c orr ec tly fi ll t he prompts_and_config/word_sections.json fi l e w i t h t he s ec t i ons pr e s e nt i n t he r e ports . python DataExtractorPipeline.py -i <file_or_folder> -m <ollama_model> 3. T e xt → P he notyp e s → HPO ✦ 4 / 4 ✦ T hi s p i p e l i n e ca n be us ed f or t e st i n g purpos e s , to e xtr ac t p he notyp e s f rom pl ai n t e xt fi l e s a n d m a p t he m to HPO t e rms . I t r e qu i r e s t e xt to be a lr ead y i n E n g l i s h . python PhenotypeExtractor.py -f <txt_file_path> -m <ollama_model> -p <phenotype_prompt_file> python HPOMapper.py -f <json_file_path> N ot e t ha t t he p he notyp e prompt fi l e ca n be t he on e w e prov ide i n prompts_and_config/phenotypes_prompt.txt or a c ustom on e . T he JSON fi l e f or HPO m a pp i n g s h oul d be t he output o f t he p he notyp e e xtr ac t i on st e p . 4. I n di v id u a l S c r i pts E ach s c r i pt ca n be run i n de p e n de ntly ( PDF T r a nsl a tor , PDF T e xt E xtr ac tor , P he notyp e E xtr ac tor , HPOM a pp e r , e t c .) C on fig ur a t i on F i l e s prompts _ a n d _ c on fig / p he notyp e s _ prompt . txt prompts _ a n d _ c on fig / wor d _ s ec t i ons . j son c on ce pts / h po _ c on ce pts _ fi lt e r ed . j son c on ce pts / m i m T i tl e s . txt L o ggi n g a n d O utput A ll i nt e rm edia t e a n d fi n a l outputs a r e wr i tt e n to OutputFiles/ , if no diffe r e nt p a t h i s gi v e n .