Watson – From Jeopardy to healthcare. Presented by Annette Hicks, IBM Australia at HINZ 2014, 11 November 2014, 11.15am, Marlborough Room

1. IBM Watson - From Jeopardy to Healthcare
Annette Hicks
Healthcare & Social Services Industry Leader
A/NZ IBM

2. Taking on Jeopardy! – Real Language is Real Hard
• Jeopardy! poses a grand challenge for a
computing system:
– Broad range of subject matter
– Speed of accurate responses and confidence
– Requires analysing subtle meaning, irony, riddles,
and other complexities – tasks that humans excel
at but which computers traditionally do not
– Ability to understand human language
• Chess
– A finite, mathematically well-defined search space
– Large but limited number of moves and states
– Everything explicit, unambiguous mathematical
rules
The Grand Challenge

3. A new era of computing will transform our future
Tabulating
Systems Era
Programmable
Systems Era
Cognitive
Systems Era
© 2014 International Business Machines Corporation
3

4. Cognitive systems expand the problems we can
address
4
Programmatic Systems Cognitive Systems
• Leverage traditional data sources
• Follow pre-defined rules (programs)
• Provide the same output to all users
• Are taught, not programmed.
• Learn and improve based on experience
• Interpret sensory and non-traditional
data
• Relate to each of us as individuals
• Allow us to expand and scale our own
thinking

5. Brief History of IBM Watson
R&D
Demonstration
Commercialization
Cross-industry
Applications
IBM
Research
Project
(2006 – )
Jeopardy!
Grand
Challenge
(Feb 2011)
Watson
for
Healthcare
(Aug 2011 –)
Watson
Industry
Solutions
(2012 – )
Watson
for Financial
Services
(Mar 2012 – )
Expansion

6. Why Watson for Healthcare?
Complexity
Evidence-based Medicine  Shifts in Models of
Care & Funding
© 2014 International Business Machines Corporation
6
Personalized Medicine
 Focus on Wellness
and Prevention
 Universal coverage
 Costs are 18% of US
GDP
 34% of $2.3T US spend
is waste
 Costs can vary up to
10x
 Diagnosis and treatment
errors
 Shortage of MDs
 Demand for remote
medicine
 Medical data
doubles every 5 years
 Detailed patient
biomedical markers
 Targeted therapies
System
Changes
Costs
Info Overload
Over 80% of stored patient health information is unstructured

7. Watson can read these medical records in six seconds!
© 2014 International Business Machines Corporation
7

8. DeepQA: Massively Parallel Probabilistic Evidence-Based
Question
100s Possible
Answers
1000’s of
Pieces of Evidence
Multiple
Interpretations
100,000’s scores from many
100s sources simultaneous Text Analysis Algorithms
Hypothesis
Generation
. . .
Hypothesis and
Evidence Scoring
Final Confidence
Merging &
Ranking
Synthesis
Question &
Topic
Analysis
Question
Decomposition
Hypothesis
Generation
Hypothesis and Evidence
Scoring
Answer &
Confidence
Architecture

9. Putting the pieces together at point of impact can be game
changing…and life changing
Symptoms
Renal Failure
UTI
Diabetes
Influenza
Hypokalemia
fever
dry mouth
thirst
anorexia
frequent urination
no abdominal pain
no back pain
no cough
no diarrhea
Oral cancer
Bladder cancer
Hemochromatosis
Purpura
(Thyroid Autoimmune)
Esophagitis
cutaneous lupus
osteoporosis
hyperlipidemia
Alendronate
Findings Medications History
pravastatin
levothyroxine
hydroxychloroquine
Diagnosis Models
frequent UTI
hypothyroidism
Confidence
difficulty swallowing
dizziness
Family
History
Graves’ Disease
Patient
urine dipstick:
leukocyte esterase
supine 120/80 mm HG
heart rate: 88 bpm
urine culture: E. Coli
Symptoms
Family
History
Her family history included oral and
bladder cancer in her mother,
Graves' disease in two sisters,
hemochromatosis in one sister, and
idiopathic thrombocytopenic
purpura in one sister
MFeidnidciantgiosns
A urine dipstick was positive for
leukocyte esterase and nitrites. The
Patient
History
A 58-A year-58-year-old old woman woman presented complains to her
of
Her medications were levothyroxine,
primary dizziness, care physician anorexia, after dry mouth,
several days
hydroxychloroquine, pravastatin, and
of dizziness, increased patient anorexia, thirst, given and dry frequent
a
mouth,
alendronate.
increased prescription urination. thirst, She fo ciprofloxacin and had frequent also had urination.
for a fever.
a
She urinary She had reported also tract had infection. no a fever pain 3 in and days her reported later,
abdomen,
that
food patient back, would reported and “get no cough, stuck” weakness when or diarrhea.
and
she was
swallowing. dizziness. Her She supine reported blood no pain pressure
in her
abdomen, was 120/80 back, mm or Hg, flank and and pulse no was cough,
88.
shortness of breath, diarrhea, or dysuria
Her history was notable for cutaneous
lupus, hyperlipidemia, osteoporosis,
frequent urinary tract infections, a left
oophorectomy for a benign cyst, and
primary hypothyroidism, diagnosed a
year earlier
• Extract Symptoms from record
Most Confident Diagnosis: IEDUnsiTfaloIubpehntaezsgaitis
• Use paraphrasings mined from text to handle
alternate phrasings and variants
• Perform broad search for possible diagnoses
• Score Confidence in each diagnosis based on
evidence so far
•• • EExxttrraacctt Identify MPeadtiiecnatt negative iHoinsstory
Symptoms
• • Use Reason database with mined of drug relations side-effects
to explain away
• Together, symptoms multiple (thirst is diagnoses consistent may w/ best UTI)
explain
symptoms
• Extract Findings: Confirms that UTI was present
• Extract Family History
• Use Medical Taxonomies to generalize medical
conditions to the granularity used by the models
© 2014 International Business Machines Corporation
9

10. 75%
of
Total
$ Trillion
Why Oncology?
3X
rate cancer cost climbs
vs. std. health costs or
15-18% / yr.
$263.8B
overall costs of
cancer
in the US in 2010
20-44%
of cancer cases
receive the wrong
diagnosis initially

11. Watson Oncology to assist
physicians with the diagnosis and
treatment of cancer
Solution:
• Clinical support for patient assessment based on objective
evidence – patient data, medical info, research, studies,
articles, best practices, guidelines, etc.
• 600K+ pieces of evidence and 2M pages of text
from 42 publications
• Analysis of patient data against thousands of historical
cases and trained through 5000+ Memorial Sloan-
Kettering MD and analyst hours
• Initial focus on lung, breast, prostate and colorectal cancers
Goals
• Create individualised cancer
diagnostic and treatment plans
• Enhance clinical confidence with
greater access and understanding
of information
• Speed time to evidence-based
treatment
• Reduce diagnostic and
administrative errors
• Accelerate the dissemination of
practice-changing research
© 2014 International Business Machines Corporation 11

12. Medical
Terminology
Watson first
needed to learn
basic medical
terminology
Health and
Wellness
Watson will
broaden its base
of medical
knowledge and
deepen its
understanding of
data relevant to
patient health
Literature,
Guidelines and
Policies
Watson is taught
how to adjudicate
medical requests
subject to policies
and guidelines
supported by
literature
Clinical Treatment
Options
Watson begins to
learn how to
understand a
patient case and
identify candidate
treatment options
and relevant
supporting
evidence
Training Watson for Healthcare

13. Medical concept annotations
Medications
Diseases Symptoms
Modifiers

14. And Creating a Corpus of Knowledge for Cancer Care
Ingestion of NCCN guidelines for breast cancer and lung cancer:
– Roughly 500,000 unique combinations of breast cancer patient attributes.
– Roughly 50,000 unique combinations of lung cancer patient attributes.
Over 600,000 pieces of evidence ingested, from 42 different publications/publishers,
including:
– The Breast Journal, National Comprehensive Cancer Network (Clinical
Practice Guidelines, Drug and Biologics compendium, et al.), American
Journal Of Hematology, Annals Of Neurology, CA: A Cancer Journal For
Clinicians, Cancer Journal, Cochrane, EBSCO, Hematological Oncology,
Hepatology, International Journal Of Cancer, Journal Of Gene Medicine,
Journal of Clinical Oncology, Journal of Oncology Practice, Massachusetts
Medical Society Journal Watch, Massachusetts Medical Society New
England Journal Of Medicine, Merck, Nephrology, UptoDate, Clinical Lung
Cancer, Current Problems in Cancer, Cancer Treatment Reviews, Elsevier's
Monographs in Cancer (multiple), Clinical Breast Cancer, European Journal
of Cancer, Lung Cancer (the journal).
IBM Confidential

15. Watson Oncology helps medical oncologists and their care
teams address these challenges
Search a corpus of
evidence data to find
supporting evidence for
each option
Evidence
Candidate Treatment
Options
3
Patient Case • Inclusion / exclusion
61 y/o woman s/p
mastectomy is here to
discuss treatment options
for a recently diagnosed
4.2 cm grade 2 infiltrating
ductal carcinoma…
Prioritized Treatment
Options
+
Evidence Profile
criteria
• Co morbidities
• Contraindications
• FDA risk factors
• MSK preferred
treatments
• Other guidelines
• Published literature -
studies, reports, opinions
from Text Books,
Journals, Manuals, etc.
Key Case Watson Oncology
Attributes
Supporting
Evidence
Extract key attributes
from a patient’s case
1
Use those attributes to
find candidate treatment
options as determined by
consulting NCCN
Guidelines
2
Use Watson’s analytic
algorithms to prioritize
treatment options based
on best evidence.
4
Guidelines
© 2014 International Business Machines Corporation 15

16. A Cognitive solution in Oncology - differs from conventional
programmatic systems
It is taught by example, not programmed. This gives it the ability to learn and adapt
as new information is presented and outcomes and actions are taken. Ground truth
is communicated through training / test cases as well as interpretation of those
cases by physicians
It generates and evaluates a set of hypothesis appropriate to the case being
presented rather than being confined to only a set of rules, or simple decision tree
logic (e.g. if this, not that), giving it a much wider range of alternatives to consider
It delivers a set of probabilistic treatment responses, which are confidence
weighted into categories
It provides the underlying rationale for its recommendations and points to evidence
that support and refute those recommendations
For physicians, Watson Oncology provides ranked treatment options
with supporting evidence rather than just listing valid guideline choices
© 2014 International Business Machines Corporation
16

17. Benefits
• Provides evidence-based suggestions to support oncologists’
decisions
• Combines patient data with massive volumes of medical literature,
including journal articles, physicians’ notes, and NCCN guidelines and
best practices
• Continues to evolve as new oncology techniques, treatments and
evidence are developed
Future Benefits :
• Improving oncologist efficiency to meet increasing demand
• Assisting physicians in handling the complexity of prescribing and
managing new cancer treatments
• Improved value of patient care through better monitoring and control
of soaring costs
• Providing expert knowledge to facilities without specialist oncology
expertise ie rural areas
© 2014 International Business Machines Corporation 17

18. 18
Watson Oncology Advisor - Helps oncologists make better, more
personalised treatment decisions by ranking treatment plans based on
national guidelines, published literature, and expert insight
MD Anderson Oncology Expert Advisor -Democratise health care by
enabling clinicians to get a clear understanding of their patient, identify
relevant SOC and Clinical Trial treatments, and a prediction and
appropriate responses to adverse events
Clinical Trial Matching - Enables decision makers to quickly evaluate
compliance / match of an input case to defined policies
Watson Genomic Analytics - Case specific analysis identifying mutations,
gene expression, tumor heterogeneity which then can be used to identify drug
options, provide rapid evidence retrieval and patient molecular profile analysis.
© 2014 International Business Machines Corporation
18
Watson Oncology Solutions

19. A new era of cognitive systems has
just begun for Healthcare
 Transforming how healthcare professionals
think, act, and operate
 Learning through interactions
 Delivering evidence based care driving better
outcomes

20. © 2014 International Business Machines Corporation
20