Fey's Project

Objectives

Clinical event extractions over patient’s timeline

Staging - whether metastasis

e.g. T1[a,b,c], T3[a,b], metastatic CRPC

Gleason score - cell differentiation

e.g. gleason 3+3=6, gl 3+4=7

PSA level

e.g. PSA is 14, PSA 3

Medication

e.g. Hormone therapy -> Leuprolide, Lupron

e.g. Chemotheropy -> Docetaxel, Novantrone

Pathway grouping for pattern exploration

At word level

At sentence level

Challenges

Highly unstructured notes

Very low efficient to extract at note level

Sectionize data using a binary classification method

Direct keyword extraction is difficult

Regular expressions fail to capture keywords along with context

- e.g. Rule ‘PSA [digit][digit].[digit]’ can capture `PSA 10.8’ but cannot capture sentences such as ‘PSA on 12/2015 was 9.0’

Use part-of-speech tagging to extract keywords in different patterns

Lack of labelled data

Cannot train a classification model

Group up keywords and key sentences using clustering methods

Pos Tagging Method

Parts Of Speech Tagging

Extension to tag words such as medications, dates, indexes as tags

e.g. replace Lupron’s tag NOUN with MEDICINE, CD with DATE

- Sentence: started Lupron on 7/8/14

- POS Sentence: ('started', 'VBD'), ('lupron', ‘NN'), ('on', 'IN'), ('7/8/14', 'CD')

- Modified POS Sentence: ('started', 'VBD'), ('lupron', 'MEDICINE'), ('on', 'IN'), ('7/8/14', 'DATE')

POS tag chunking rules for each clinical event

ON_DATE: extract events occurring on a given date

- ('started', 'VBD'), ('lupron', 'MEDICINE'), ('on', 'IN'), ('7/8/14', 'DATE')

- ('PSA', 'PSA'), ('on', 'IN'), ('02/12/2008', 'DATE'), ('was', 'VBD'), ('10', 'CD')

PRESENT: extract events occurring at present – specific to date of note

- ('continues', 'VBZ'), ('on', 'IN'), ('xofigo', 'MEDICINE')

- ('PSA', 'PSA'), ('is', 'VBZ'), ('10', 'CD')

Pos Tagging Results

System Design

Patient Note Timeline

Keyword Extraction

Patient Pathway

Medications on Admission/Discharge

Word Similarity

Time to Development of Metastatic Disease

- Important for physicians to understand when metastasis happened

Embedding

- Numerical representation for words or sentences

- Allows performing arithmetic operations on words/sentences and similarity calculations

- Similar words/sentences are closer in space

Sentence Similarity

Sentences similar to ‘revealed metastatic disease’

Clinical Pathway Extraction and Clustering for Prostate Cancer Patients