This was presented at the 2014 IEEE Conference on Big Data
For details:
http://kavita-ganesan.com/content/general-supervised-approach-segmentation-clinical-texts
Citation:
Ganesan, Kavita, and Michael Subotin. "A General Supervised Approach to Segmentation of Clinical Texts."
Kavita Ganesan & Michael Subotin Present Segmentation of Unstructured Clinical Notes
1. Kavita Ganesan & Michael Subotin
Presented at: 2014 Conference on IEEE Big Data
2. All sorts of notes types!
Admit notes
◦ documenting why patient is being admitted
◦ baseline status, etc.
Progress notes
◦ progress during course of hospitalization
Discharge notes
◦ conclusion of a hospital stay or series of treatments
Others
◦ Operative notes
◦ Procedure notes
◦ Delivery notes
◦ Emergency Department notes, etc
3. PRIMARY CARE PHYSICIAN:
Dr. XXXXX XXXXXXXX.
CHIEF COMPLAINT:
Injured right little toe.
HISTORY OF PRESENT ILLNESS:
This is a 63-year-old male with a past medical history of multiple
myeloma who presents today after hitting his fifth toe of the right foot
on a wood panel yesterday……
Review of Systems:
CONSTITUTIONAL: No fever, chills, or weight loss.
RESPIRATORY: No cough, shortness of breath, or wheezing.
CARDIOVASCULAR: No chest pain, chest pressure, or palpitations.
...............
PAST MEDICAL HISTORY
Multiple myeloma, peripheral neuropathy, hypertension..
PAST SURGICAL HISTORY:-
Stem cell transplant.
SOCIAL HISTORY
The patient formerly smoked tobacco; however, quit within the last 10
years.
FAMILY HISTORY:
Hypertension.
ALLERGIES:
ASPIRIN.
………
Purpose of visit
Patient’s current
condition in
narrative form
Ongoing issues,
issues in the past
Information on
allergies
4. PRIMARY CARE PHYSICIAN:
Dr. XXXXX XXXXXXXX.
CHIEF COMPLAINT:
Injured right little toe.
HISTORY OF PRESENT ILLNESS:
This is a 63-year-old male with a past medical history of multiple
myeloma who presents today after hitting his fifth toe of the right foot
on a wood panel yesterday……
Review of Systems:
CONSTITUTIONAL: No fever, chills, or weight loss.
RESPIRATORY: No cough, shortness of breath, or wheezing.
CARDIOVASCULAR: No chest pain, chest pressure, or palpitations.
...............
PAST MEDICAL HISTORY
Multiple myeloma, peripheral neuropathy, hypertension..
PAST SURGICAL HISTORY:-
Stem cell transplant.
SOCIAL HISTORY
The patient formerly smoked tobacco; however, quit within the last 10
years.
This is how most notes look:
• some longer, some shorter
• different set of headers, etc
FAMILY HISTORY:
Hypertension.
ALLERGIES:
ASPIRIN.
………
Purpose of visit
Patient’s current
condition in
narrative form
Ongoing issues,
issues in the past
Information on
allergies
5. PRIMARY CARE PHYSICIAN:
Dr. XXXXX XXXXXXXX.
PRIMARY CARE PHYSICIAN:
Dr. XXXXX XXXXXXXX.
PRIMARY CARE PHYSICIAN:
Dr. XXXXX XXXXXXXX.
CHIEF COMPLAIN:
Injured right little toe.
CHIEF COMPLAIN:
Injured right little toe.
CHIEF COMPLAINT:
Injured right little toe.
HISTORY OF PRESENT ILLNESS:
This is a 63-year-old male with
a past medical history of…
HISTORY OF PRESENT ILLNESS:
This is a 63-year-old male with
a past medical history of…
HISTORY OF PRESENT ILLNESS:
This is a 63-year-old male with
a past medical history of…
Review of Systems:
CONSTITUTIONAL: No fever,
chills, or weight loss.
CARDIOVASCULAR: No chest pain,
chest pressure, or palpitations.
...............
Review of Systems:
CONSTITUTIONAL: No fever,
chills, or weight loss.
CARDIOVASCULAR: No chest pain,
chest pressure, or palpitations.
...............
………
Review of Systems:
CONSTITUTIONAL: No fever,
chills, or weight loss.
CARDIOVASCULAR: No chest pain,
chest pressure, or palpitations.
...............
………
………
Very unstructured
◦ formatting cues inconsistent
◦ varies: across physicians, notes,
hospitals
Hard to analyze specific sections
◦ E.g. analyze allergies patient population
◦ Need to segment notes to extract
all allergy info.
6. ◦ Information collected vary from note types to note types
Ex. info on progress notes vs. admit note
◦ Contents & formatting can vary from hospital to hospital
Even within the same organization – E.g. Kaiser
◦ Contents & formatting vary between physicians
Different styles, speed of typing, etc.
7. If you are looking at a single note type, from a single
hospital - then maybe
Not suitable as a general segmentation approach:
Can easily break:
◦ on unseen note types and minor format variations
◦ Example:
regex based on all caps
regex based on seen headers only
8. Several works have explored supervised methods to
segmenting clinical notes
[Cho et al. 2003, tepper et al. 2012, apostolva et al. 2009]
Problem: methods not general!
◦ Cho et al. 2003: One model for each type of note
20 note types 20 models!
Not practical maintain each model
◦ Tepper et al. 2012: Model had low adaptability to unseen
documents
features used, training data used, etc.
9. General segmentation approach for clinical texts
Requirements:
◦ Single model/approach for most note types
◦ Discount extreme non-standard formatting
e.g. tabular format
Segment:
◦ Header
◦ Top level sections
◦ Footer
10. PRIMARY CARE PHYSICIAN:
Dr. XXXXX XXXXXXXX.
CHIEF COMPLAINT:
Injured right little toe.
HISTORY OF PRESENT ILLNESS:
This is a 63-year-old male with a past medical history of multiple
myeloma who presents today after hitting his fifth toe of the right foot
on a wood panel yesterday……
Review of Systems:
CONSTITUTIONAL: No fever, chills, or weight loss.
RESPIRATORY: No cough, shortness of breath, or wheezing.
CARDIOVASCULAR: No chest pain, chest pressure, or palpitations.
...............
PAST MEDICAL HISTORY
Multiple myeloma, peripheral neuropathy, hypertension..
PAST SURGICAL HISTORY:-
Stem cell transplant.
SOCIAL HISTORY
The patient formerly smoked tobacco; however, quit within the last 10
years.
FAMILY HISTORY:
Hypertension.
ALLERGIES:
ASPIRIN.
………
Header
Top-level section
Top-level section
Top-level section
Top-level section
Top-level section
Top-level section
Top-level section
11. Supervised approach using L1-Logistic Regression with a
constraint combination approach
Idea: scan each line in a clinical document and label as:
◦ BeginHeader
◦ ContHeader
◦ BeginSection
◦ ContSection
◦ Footer
Labels are predicted with certain confidence
But, problem using line-wise predictions as is:
◦ Label sequences may not make sense
◦ E.g. There maybe a BeginHeader after a BeginSection
incorrect
12. Post-processing: enforce sequence combination rules:
◦ First line of document: BeginHeader or BeginSection
◦ BeginHeader cannot come right after BeginHeader or ContHeader
◦ ContHeader must come after BeginHeader or ContHeader
◦ ContSection must come after BeginSection or ContSection
◦ Footer cannot come right after BeginHeader or ContHeader
Rules applied after all lines in document labeled
◦ Applied to consecutive label pairs
◦ Computed efficiently: Viterbi algorithm
13. Inpatient Outpatient
• Notes from 12 different enterprises
• Some large enterprises
• All sorts of note types
• Some noisy sectioning, some clean
• 100 radiology notes
• Fairly clean sections
• One hospital
• All sorts of note types
• Fairly well sectioned
• 35, 000 notes in total
• 2000 randomly sampled notes
(inpatient)
• 100 radiology notes
• Fairly clean sections
14. Emphasis on training data
Variation in training data
◦ Use different note types for training
◦ Intuition: help model generalize well
Sample training data:
◦ Instead of using all training data from 2100 notes
◦ Generated subsets of training data with varying size and
cross-validate on test sets
◦ Intuition: allows to pick the best model
Best model only used < 700 notes (out of 2100)
15. 5 test sets
◦ 4/5 test set from hospitals not in train set
true estimate of accuracy
◦ Covers both inpatient and outpatient notes
◦ Covers different note types
◦ ~12,500 test notes
Primary evaluation metric: line-wise accuracy
◦ percentage of correctly predicted line labels
16. 1st model: limited variety
(hp + discharge)
Train set
3-folded cross
validation
Unseen test
accuracy
Inp1HospB (300 - limited) 96.70% 67.00%
Inp3HospD (300 - varied) 96.58% 88.23%
2nd model: variety
(11 types - hp, ds, pn…)
Model with variety:
higher accuracy on
unseen test set
3-folded cross-validation
accuracy: high in both
Important to have variety in training notes in
building general segmentation model
17. Accuracy consistently
> 90% across enterprises
Client/Data In/Outpatient # Test Docs Accuracy
1. Inp1HospB In 300 92.58%
2. Inp2HospC In 1000 93.29%
3. Inp3HospD In 300 95.81%
4. Rad1MixedHosps Out 9000 92.45%
5. Rad2HospA Out 1902 93.67%
Average 93.56%
• Average accuracy: 93.56%
• Covers inpatient/outpatient
Single model: But, performs well across enterprises
18. Document Type Accuracy
1. History and Physical 95.70%
2. Physician Clinicals 93.10%
3. Discharge Summary 94.00%
4. Consult Note 94.60%
5. Short Stay Summary 94.60%
6. Operative Note 92.20%
7. Progress Note 87.80%
8. Cardiac Cath Report 85.40%
9. Procedure Note 83.60%
• Model performs well across note types
• Lowest performance: procedure notes
low recall on segmenting “technique” sections
Performs
very well
> 90%
Reasonable..
> 80%
Accuracy Breakdown for Inp2HospC
19. 94.00%
93.00%
92.00%
91.00%
90.00%
89.00%
88.00%
87.00%
86.00%
# Notes vs. Accuracy
No benefit with more notes
0 500 1000 1500 2000
Accuracy
# Training Notes
Avg. accurracy peaks @500
notes on all test sets
No benefit with more notes
No need for big data for a general model.
We need good data from all that big data!
20. Unigrams – of each line (LineUnigram)
Relative position of line in document (PosInDoc)
◦ Top, Middle, Bottom
Known Header features (KnownHeader)
◦ Find potential headers using repository of seen headers
◦ Seen headers can have canonical type
E.g. Past Medical History, Previous Med History “PAST_MEDICAL_HISTORY”
◦ If potential headers found, we include features:
Canonical type
Unigram & Char n-gram of potential header
Caps/colon info – mixed case, all caps, lowercase
Length of potential header
22. Explored:
◦ Supervised approach to building a very general segmentation
model for clinical texts
Evaluation showed:
◦ Model works well on notes across enterprises
◦ Model works across note types
Key to effectiveness:
◦ Variation in training data –all sorts of note types
◦ Training data selection strategy – sample and cross-validate
◦ Feature set – not explored in existing works