A text mining system must go way beyond indexing and search to appear truly intelligent. First, it should understand language beyond keyword matching (for example, distinguishing between “Jane has the flu,” “Jane may have the flu,” “Jane is concerned about the flu," “Jane’s sister has the flu, but she doesn’t,” or “Jane had the flu when she was 9” is of critical importance). This is a natural language processing problem. Second, it should “read between the lines” and make likely inferences even if they’re not explicitly written (for example, if Jane has had a fever, a headache, fatigue, and a runny nose for three days, not as part of an ongoing condition, then she likely has the flu). This is a semi-supervised machine learning problem. And third, it should automatically learn the right contextual inferences to make (for example, learning on its own that fatigue is (sometimes) a flu symptom—only because it appears in many diagnosed patients—without a human ever explicitly stating that rule). This is an association-mining problem, which can be tackled via deep learning or via more guided machine-learning techniques. This is a live demo of an end-to-end system that makes nontrivial clinical inferences from free-text patient records and provides real-time inferencing at scale. The architecture is built out of open source big data components: Kafka and Spark Streaming for real-time data ingestion and processing, Spark for modeling, and Titan and Elasticsearch for enabling low-latency access to results. The data science components include a UIMA pipeline with custom annotators, machine-learning models for implicit inferences, and dynamic ontologies based on deep learning with Word2Vec for representing and learning new relationships between concepts. Source code is publicly available to enable you to hack away on your own.

1. David Talby
@davidtalby
CTO, Atigeo
SEMANTIC NATURAL LANGUAGE UNDERSTANDING
WITH SPARK, UIMA & MACHINE-LEARNED ONTOLOGIES
Claudiu Branzan
@melcutz
Principal Lead, Atigeo

2. 2
2
THE PROBLEM
Who needs to
be vaccinated?
Who fits this
clinical trial?
Who is at risk
for sepsis?
Who is getting
meds they’re
allergic to?
Who on this protocol
did not have this
side effect?

3. 3
AT THE BEGINNING, THERE WAS SEARCH
Scalable & robust Indexing pipeline
Tokenizers & analyzers
Synonyms, spellers & Auto-suggest
File formats & header boosting
Rankers, link & reputation boosting

4. 4
THEN THERE WAS SEMANTIC SEARCH
“cheap red prom dresses”
“laptops under $500”
“italian restaurants near me that deliver”
“captain america civil war tonight”
“nba scores”
Dictionary Based Attribute Extraction
Dell - XPS 15.6 4K Ultra HD Touch-Screen
Laptop - Intel Core i5 - 8GB Memory -
256GB Solid State Drive - Silver
Machine Learned Attribute Extraction
If you go for the ambience, you'll be
disappointed. If you go for good,
inexpensive and authentic Mexican food,
then you're in the right place.

5. 5
AND THEN, YOU NEED TO UNDERSTAND LANGUAGE
Prescribing sick days due to diagnosis of influenza. Positive
Jane complains about flu-like symptoms. Speculative
Jane may be experiencing some sort of flu episode. Possible
Jane’s RIDT came back negative for influenza. Negative
Jane is at high risk for flu if she’s not vaccinated. Conditional
Jane’s older brother had the flu last month. Family history
Jane had a severe case of flu last year. Patient history

6. 6
LANGUAGE GETS COMPLEX & DOMAIN SPECIFIC
Joe expressed concerns about the risks of bird flu. Nothing
Joe shows no signs of stroke, except for numbness. Double Negative
Nausea, vomiting and ankle swelling negative. Compound
(it gets worse – in reality a lot of text isn’t valid English)
Patient denies alcohol abuse. Speculative
Allergies: Penicillin, Dust, Sneezing. Compound

7. 7
7
LET’S BUILD THIS!
The input
(patient records)
The processing
framework
The output The query engines

8. 8
8
SENTENCE DETECTION
SECTION DETECTION
TOKENIZER LEMMATIZER
STOPWORD REMOVAL
NEGATION DETECTION
CONDITIONAL SCOPE
SPECULATIVE SCOPE
DATE NUMBER UNIT QUANITITY
CONCEPT EXTRACTION

9. 9
9
First Demo: Annotators & Assertions

10. 1 0
10
MACHINE LEARNED ANNOTATORS
Grammatical Patterns
If … then …
Direct Inferences
Age < 18 ==> Child
Lookups
RIDT (lab test)
Under-diagnosed conditions
Flu Depression
Implied by Context
relevant labs normal
Sometimes, it’s easier to just code an annotation’s business logic
But sometimes it’s easier to learn it from examples:

11. 1 1
11
Second Demo: Machine Learned Annotator

12. 1 2

13. 1 3
13
WHAT ABOUT EXPANDING & UPDATING ONTOLOGIES?
Word2Vec

14. 1 4
14
LET’S BUILD THIS TOO!

15. 1 5
15
Third Demo: Ontology Enrichment

16. 1 6
16
SUMMARY & APPLICATIONS
Who needs to
be vaccinated?
Who fits this
clinical trial?
Who is at risk
for sepsis?
Who is getting
meds they’re
allergic to?
Who on this protocol
did not have this
side effect?

17. 1 7
17
@Atigeo
@melcutz
@davidtalby

18. © 2015 Atigeo, Corporation. All rights reserved. Atigeo and the xPatterns logo are trademarks of Atigeo. The information herein is for informational purposes only and represents the current view of Atigeo as of the date of this presentation. Because Atigeo
must respond to changing market conditions, it should not be interpreted to be a commitment on the part of Atigeo, and Atigeo cannot guarantee the accuracy of any information provided after the date of this presentation. ATIGEO MAKES NO WARRANTIES,
EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

19. APPENDIX
In case the live demo gets cold feet on stage
1 9

20. 2 0

21. 2 1

22. 2 2

23. 2 3

24. 2 4

25. 2 5

26. 2 6

27. 2 7

28. 2 8