Modeling human comprehension of Swedish medical records for - - PowerPoint PPT Presentation

Modeling human comprehension

• f Swedish medical records

for intelligent access and summarization systems – Future vision, a physician’s perspective

Maria Kvist Maria Skeppstedt Sumithra Velupillai Hercules Dalianis Karolinska Universitetssjukhuset Inst för data- och systemvetenskap(DSV), Stockholms universitet

Clinical documentation

Mia Kvist

Mia Kvist

Clinical documentation

Mia Kvist

We need: Overview Search tool Text summarization Knowledge extraction

Mia Kvist

Din journal på nätet electronic access on Internet to my own patient record

Mia Kvist

Din journal på nätet electronic access on Internet to my own patient record Patients need: Overview Search tool Summarization tool Translation tool

Mia Kvist

Patients need: Overview Search tool Summarization tool Translation tool

Mia Kvist

We need: Overview Search tool Text summarization Knowledge extraction

Talented group of people

• Inst. för data- och systemvetenskap

Hercules Dalianis Martin Hassel Sumithra Velupillai Maria Skeppstedt Aron Henriksson Gunnar Nilsson, KI

Mia Kvist

Text summarization

• Domain specific language
• How do physicians read, comprehend,

summarize and reach an understanding for a patients situation and needs?

Mia Kvist

Text summarization

• Domain specific language

Need better understanding

• How do physicians read, comprehend,

summarize and reach an understanding for a patients situation and needs? Need models to spot what is important Need grading of the certainty of knowledge

Mia Kvist

Stockholm EPR Corpus

• Karolinska University Hospital
• 2 milj deidentified patient records
• 2006-2010
• Various medical, surgical specialities

+ childrens hospital

• No psychiatry

Mia Kvist

Finding important facts

• Entity recognition

Mia Kvist

Patient’s twin probably had a myocardial infarction two years ago

Finding important facts

• Entity recognition
• Certainty

Mia Kvist

Patient’s twin probably had a myocardial infarction two years ago

Finding important facts

• Entity recognition
• Certainty
• Temporality

Mia Kvist

Patient’s twin probably had a myocardial infarction two years ago

Finding important facts

• Entity recognition
• Certainty
• Temporality
• Subject identification

Mia Kvist

Patient’s twin probably had a myocardial infarction two years ago

Finding important facts

• Entity recognition
• Certainty
• Temporality
• Subject identification

Mia Kvist

Patient’s twin probably had a myocardial infarction two years ago P’s twin prob had MI 2y ago

• Unstructured free text

incompleate sentences, few subjects, passive verbs

• Medical terminology
• Medical jargong
• Abbreviations and acronyms
• Latin, Greek and English

Mia Kvist

Electronic patient records

• language

Mia Kvist

MIE 2011:

Using SNOMED CT for High Precision Entity Recognition in Swedish Clinical Text

Maria Skeppstedt Hercules Dalianis

Spot all clinical findings in free text. Traditional way: Match to list of diagnoses, symptoms. (ICD-10, SNOMED CT etc) But do we express diagnoses in that way in free text?

Mia Kvist

Entity recognition

• Clinical findings

Dialysbehandlad kärlsjuk diabetiker med huvudvärksproblematik. Dialysistreated vascular disease-sick diabetic with headache problems.

Mia Kvist

Entity recognition

• Implicit disorders

Entity recognition

• annotation of clinical findings

Methods Annotation Machine learning systems

• Train on the annotated data
• Automatically detect symptoms and

disorders through contextual or other markers

Mia Kvist

Qualitative results The same clinical expression can be a finding or a disorder depending on context. Ex) tachycardia Some clinical findings are not recognized as the same. Different expressions, abbreviations, misspellings. Ex) Troponin-T normal

Mia Kvist

Entity recognition

• annotation of clinical findings

Mia Kvist

Swedish is rich in compound words

Chestpain symptom Pain in the chest symptom and body part Lungclinic a place or institution, no body part

MIE 2011:

Factuality Levels of Diagnoses in Swedish Clinical Text

Sumithra Velupillai (presenter) Hercules Dalianis Maria Kvist

Mia Kvist

Certainty of diagnosis

• annotation

Machine learning process Model subtelties expressed in natural language Learn to determine the certainty of a diagnosis by

• cue words
• patterns inherent in the diagnosis expression

Certainly Positive Probably Positive Possibly Positive Certainly Negative Probably Negative Possibly Negative

+

• Mia Kvist

Patient has Parkinsons disease. Physical examination strongly suggests Parkinson. Patient possibly has Parkinson. Parkinson cannot yet be outruled. No support for Parkinson. Parkinsson can be excluded.

0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%

Annotation classes for 15 diagnoses Certainly Positive Probably Positive Possibly Positive Certainly Negative Possibly Negative Probably Negative

28

Mia Kvist

Certainty of diagnosis

• annotation

Qualitative results There is more to it than just cue phrases.

• Overt findings - high certainty
• Diagnoses that lack negative classes
• Some diagnoses do not need to be certain
• Comlementary diagnoses

10 20 30 40 50

ischemia heart attack angina pectoris Annotations Certainly Pos Probably Pos Possibly Pos Certainly Neg Probably Neg Possibly Neg

Conclusions

It may be possible to, bit by bit, put together system that can

• find important facts
• determine their certainty
• summarize

Many obstacles are yet to be overcome

Mia Kvist