Talk at PARC, Oct. 30, 2013. Abstract at: http://j.mp/PARCabs
[To see animations, you may need to download the file and use powerpoint.]
Also see related talks on Smart Data for Smart Energy and other applications: http://wiki.knoesis.org/index.php/Smart_Data
The proliferation of smartphones and sensors, the continuous monitoring of physiology and environment (personal health signals), notifications from public health sources (public health signals), and more digital access to clinical data, are resulting in massive multisensory and multimodal observational data. The technology has significant potential to improve health and well-being, through early detection, better diagnosis, effective prevention and treatment of a disease; and improved the quality of life. However, to make this personalized digital medicine a reality, it is crucial to derive actionable insights from data including heterogeneous and fine-grained observations.
At Kno.e.sis, we have collaborations with clinicians in growing number of specializations (Cardiovascular, Pulmonology, Gastroenterology) to study personalized health decision making that involve the use of real-world patient data, deep background knowledge and well targeted clinical applications. For example:
* For a patient discharged from hospital with Acute Decompensated Heart Failure, can we compute post hospital discharge risk factor to reduce 30-day readmissions?
* For children with Asthma, can we predict an impending attack to enable actions that prevent an attack reducing the need for post-attack symptomatic relief?
* For Parkinson’s Disease, can we characterize the progression to adjust medication and therapeutic changes?
The above provides the context for a research agenda around what I call Smart Data, which (a) provides value from harnessing the challenges posed by volume, velocity, variety and veracity of Big Data, in-turn providing actionable information and improve decision making, and/or (b) is focused on the actionable value achieved by human involvement in data creation, processing and consumption phases for improving the Human experience. In describing Smart Data approach to above heath applications, I will cover the following technical capabilities that adds semantics to enhance or complement traditional NLP and ML centric solutions:
* Semantic Sensor Web- including semantic computation infrastructure, ability to semi-automatically create domain specific background knowledge (ontology) from unstructured data (e.g., EMR), and automatically do semantic annotation of multimodal and multisensory data
*
Semantic perception – convert low level signals into higher level abstractions using IntellegO framework that utilizes domain knowledge and hybrid abductive/deductive reasoning
* Intelligence at Edge - perform scalable and efficient semantic computation on resource constrained devices
1. Put Knoesis Banner
Smart Data enabling Personalized Digital Health :
Deriving Value via harnessing Volume, Variety and Velocity
using semantics and Semantic Web
Amit P. Sheth
Ohio Center of Excellence in Knowledge-enabled Computing (Kno.e.sis)
Wright State University, USA
Contributions by many, but Special Thanks to:
Pramod
Anantharam
Cory
Henson
Dr. T.K.
Prasad
Sujan
Perera
Delroy
Cameron
2. A Historical Perspective on Collecting Health Observations
Doctors relied only on
external observations
Stethoscope was the
first instrument to go
beyond just external
observations
Though the stethoscope
has survived, it is only one
among many observations
in modern medicine
Laennec’s stethoscope
Imhotep
Image Credit: British Museum
2600 BC
Diseases treated only
by external observations
~1815
First peek beyond just
external observations
http://en.wikipedia.org/wiki/Timeline_of_medicine_and_medical_technology
Today
Information overload!
2
3. The Patient of the Future
MIT Technology Review, 2012
http://www.technologyreview.com/featuredstory/426968/the-patient-of-the-future/
3
4. Big Data in Medicine: Implications
“We should not make the mistake of seeing data as a technical issue.
It’s a synthesis problem. That’s because information is not the scarce
resource. Attention is.”
-- Conrad Wai, The data addiction | The Ideas Economy
http://www.davidscaduto.com/post/9048831674/we-should-not-make-the-mistake-of-seeing-data-as
4
5. Sources of Big Data in Digital Health
Variety
Veracity
Velocity
Volume
Image: http://www.dr4ward.com/dr4ward/2013/04/what-is-the-power-of-the-big-data-in-healthcare-infographic.html
5
7. Big Data in Digital Health: Can alerts work?
"According to multiple recent studies, doctors ignore between 49–96%
of all CDS alerts that EMRs give them.”1
"Clinical Decision Support systems link health observations with health
knowledge to influence health choices by clinicians for improved health care".
-- Robert Hayward, Centre for Health Evidence
1http://www.fastcodesign.com/1664763/badly-designed-electronic-medical-records-can-kill-you
7
8. Information Overload leading to Alert Fatigue
Ignoring alerts is not limited to Emergency Rooms but has also crept
into EMR alerts commonly referred to as “alert fatigue”
http://health.embs.org/editorial-blog/noise-in-hospital-intensive-care-units-icus/
8
9. Questions typically asked on Big Data
• What if your data volume gets so large and
varied you don't know how to deal with it?
• Do you store all your data?
• Do you analyze it all?
• How can you find out which data points are
really important?
• How can you use it to your best advantage?
http://www.sas.com/big-data/
9
10. Variety of Data Analytics Enablers
http://techcrunch.com/2012/10/27/big-data-right-now-five-trendy-open-source-technologies/
10
11. Illustrative Big Data Applications
• Prediction of the spread of flu in real time during H1N1 2009
– Google tested a mammoth of 450 million different mathematical
models to test the search terms, comparing their predictions against
the actual flu cases; 45 important parameters were founds
– Model was tested when H1N1 crisis struck in 2009 and gave more
meaningful and valuable real time information than any public health
official system [Big Data, Viktor Mayer-Schonberger and Kenneth Cukier, 2013]
• FareCast: predict the direction of air fares over different
routes [Big Data, Viktor Mayer-Schonberger and Kenneth Cukier, 2013]
• NY city manholes problem [ICML Discussion, 2012]
11
12. What is missing?
• Current focus mainly to serve business intelligence and targeted analytics
needs, not to serve complex individual and collective human needs (e.g.,
empower human in health, fitness and well-being; better disaster
coordination,
smart
energy
consumption)
that
is
highly
personalized/individualized/contextualized
– Incorporate real-world complexity: multi-modal and multi-sensory nature of realworld and human perception
– Need deeper understanding of data and its role to information (e.g., skew,
coverage)
– Beyond correlation -> causation :: actionable info, decisions grounded on insights
• Human involvement and guidance: Leading to actionable information,
understanding and insight right in the context of human activities
– Bottom-up & Top-down processing: Infusion of models and background knowledge
(data + knowledge + reasoning)
12
14. Human Centric Computing
EXPERIENCE
& DECISION
MAKING
Descriptive
Exploratory
Inferential
Predictive
Causal
Improved
Analytics
CREATION
PROCESSING
14
15. Smart Data
Smart data makes sense out of Big data
It provides value from harnessing the
challenges posed by volume, velocity,
variety and veracity of big data, in-turn
providing actionable information and
improve decision making.
15
16. Another perspective on Smart Data
“OF human, BY human and FOR human”
Smart data is focused on the actionable
value achieved by human involvement in
data creation, processing and consumption
phases for improving
the human experience.
16
17. Current Focus on Big Data
• Focus on verticals: advertising‚ social media‚ retail‚
financial services‚ telecom‚ and healthcare
– Aggregate data, focused on transactions, limited
integration (limited complexity), analytics to find
(simple) patterns
– Emphasis on technologies to handle volume/scale,
and to lesser extent velocity: Hadoop, NoSQL,MPP
warehouse ….
– Full faith in the power of data (no hypothesis),
bottom up analysis
17
19. „OF human‟ : Relevant Real-time Data
Streams for Human Experience
Petabytes of Physical(sensory)-Cyber-Social Data everyday!
More on PCS Computing: http://wiki.knoesis.org/index.php/PCS
19
21. „BY human‟: Involving
Crowd Intelligence in data processing workflows
Use of Prior Human-created Knowledge Models
Crowdsourcing and Domain-expert guided
Machine Learning Modeling
21
23. „FOR human‟ :
Improving Human Experience
Weather Application
Weather Application
Asthma Healthcare Application
Action in the Physical World
Personal
Public Health
Population
Level
Detection of events, such as
wheezing sound, indoor
temperature, humidity, dust, and
CO2 level
Close the window at home during
day to avoid CO2 inflow, to avoid
asthma attacks at night
23
24. Why do we care about Smart Data
rather than Big Data?
24
26. April 6, 2011
Mr. Michael Yocabet suffering from type 1 diabetes is recommended a
kidney transplant at the University of Pittsburgh Medical Center. The
organ donor is his life partner Ms. Christina Mecannic
http://www.nbcnews.com/id/44599555/#.UmMHMWRDszQ
26
27. May 6, 2011
The couple leaned about the botched kidney transplant making the
situation of Mr. Yocabet much worse! The kidney he got from his wife
has infected him with Hepatitis C aggravating his health issues.
http://www.nbcnews.com/id/44599555/#.UmMHMWRDszQ
27
28. Life Threatening Implications!
Mr. Yocabet was a disabled former truck driver and he has diabetes type 1.
Treatment for the liver may harm his kidney even cause organ failure and death!
“Because he’s on anti-rejection drugs, the hepatitis C will be a lot worse in him,” - Ms. Christina Mecannic
http://www.scientificamerican.com/article.cfm?id=2003-blackout-five-years-later
28
29. Cause of the Problem: Official Investigation
• Jan 26: Ms. Mecannic gets her blood work
positive for Hepatitis C virus.
• March 29: Second attempt to test for Hepatitis
C virus in Ms. Mecannic.
• Several meetings of the transplant team -they fail to notice the problem. (alert fatigue?)
• April 6: Transplant day!
• May 6: Couple learned about botched
transplant.
http://www.post-gazette.com/stories/local/breaking/upmc-sued-over-botched-kidney-transplant-315580/
http://www.nbcnews.com/id/44599555/#.UmMHMWRDszQ
29
30. Can we Prevent such life threatening incidents?
Over 28,000 organs of all types are transplanted every
year in United States alone
"Between 2007 and 2010, the CDC conducted 200 investigations into potential
transmission of HIV and hepatitis B and C due to organ transplants.”
http://www.nbcnews.com/id/44599555/#.UmMHMWRDszQ
30
31. How could Smart Data help?
Value: Healthcare Provider Context
31
32. Clinical Decision Making is Complex!
“Health professionals are required to make decisions with multiple foci (e.g.
diagnosis, intervention, interaction and evaluation), in dynamic contexts, using a
diverse knowledge base (including an increasing body of evidence-based
literature), with multiple variables and individuals involved.”
http://researchoutput.csu.edu.au/R/?func=dbin-jump-full&object_id=9063&local_base=GEN01-CSU01
32
33. Stakes are high for both doctors and patients!
http://researchoutput.csu.edu.au/R/?func=dbin-jump-full&object_id=9063&local_base=GEN01-CSU01
33
34. Multimodal, Multisensory, and Multi-organizational Observations
Expert opinion
Clinical research
Population
health record
Personal health
record
Clinical decision
support
What is the overall health of the person?
What are the vulnerabilities for organ transplant?
http://www.rugeleypower.com/electricity-generation/producing-electricity.php
34
35. Patient Health Score (diagnostic)
Semantic Perception and risk assessment algorithms can transform raw data (hard
to comprehend) to abstractions (e.g., Patient Health is 3 on a scale of 5) that is
intuitively understandable and valuable for decision makers.
Having health score for various patients will allow efficient utilization of
a decision maker’s precious attention
Expert opinion
Clinical research
Population
health record
Personal health
record
Clinical decision
support
Semantic
Perception
Risk assessment
model
35
36. Patient Vulnerability Score (prognostic)
The Clinical Decision Support systems such as EMR alert system in its
current state follows the high recall philosophy by reporting every
possible alert!
Doctors need actionable information and not a deluge of alerts to make
timely and important decisions. Providing a vulnerability score would
facilitate right use of Doctor’s time to investigate further on vulnerabilities.
Expert opinion
Clinical research
Population
health record
Personal health
record
Clinical decision
support
Semantic
Perception
Risk assessment
model
36
38. “Intelligence at the Edges” of Digital Health
3.4 billion people will have smartphones or tablets by 2017
-- Research2Guidance
m-health app market is predicted to reach $26 billion in 2017
-- Research2Guidance
http://www.digikey.com/us/en/techzone/energy-harvesting/resources/articles/zigbees-smart-energy-20-profile.html
38
39. Data Overload for Patients/health aficionados
Providing actionable information in a timely manner is
crucial to avoid information overload or fatigue
Personal
Schedule
Sleep data
Activity data
Personal health
records
Community data
39
40. Optimizing Cost, Benefit, and Preferences
Algorithms on the patient side should consider all the health signals and
provide actionable and timely information for informed decision making
What are the reasons for my increasing weight?
What should I consider before I get a kidney transplant?
Personal
Schedule
Sleep data
Activity data
Personal health
records
Community data
Semantic
Perception
Personalized
optimization
Img: http://marloncarvallovillae.blogspot.com/2011_02_01_archive.html
http://www.1800timeclocks.com/icon-time-systems/icon-time-upgrades/icon-time-advanced-pack-upgrade-sb100-pro/
Personalized
recommendation
40
41. 3 Primary Issues to be addressed
1
Annotation of
sensor data
Semantic
Sensor
Web
2
Interpretation of
sensor data
Semantic
Perception
3
Efficient execution on
resource-constrained devices
Intelligence
at the Edge
41
42. How are machines supposed to integrate and interpret sensor data?
RDF
OWL
Semantic Sensor Networks (SSN)
42
43. W3C Semantic Sensor Network Ontology
Lefort, L., Henson, C., Taylor, K., Barnaghi, P., Compton, M., Corcho, O., Garcia-Castro, R., Graybeal, J., Herzog, A., Janowicz, K.,
Neuhaus, H., Nikolov, A., and Page, K.: Semantic Sensor Network XG Final Report, W3C Incubator Group Report (2011).
43
44. W3C Semantic Sensor Network Ontology
Lefort, L., Henson, C., Taylor, K., Barnaghi, P., Compton, M., Corcho, O., Garcia-Castro, R., Graybeal, J., Herzog, A., Janowicz, K.,
Neuhaus, H., Nikolov, A., and Page, K.: Semantic Sensor Network XG Final Report, W3C Incubator Group Report (2011).
44
45. W3C Semantic Sensor Network Ontology
Lefort, L., Henson, C., Taylor, K., Barnaghi, P., Compton, M., Corcho, O., Garcia-Castro, R., Graybeal, J., Herzog, A., Janowicz, K.,
Neuhaus, H., Nikolov, A., and Page, K.: Semantic Sensor Network XG Final Report, W3C Incubator Group Report (2011).
45
46. Semantic Annotation of SWE
Lefort, L., Henson, C., Taylor, K., Barnaghi, P., Compton, M., Corcho, O., Garcia-Castro, R., Graybeal, J., Herzog, A., Janowicz, K.,
Neuhaus, H., Nikolov, A., and Page, K.: Semantic Sensor Network XG Final Report, W3C Incubator Group Report (2011).
46
47. Smart Data in Healthcare
To gain new insight in
patient care &
early indications of
disease
47
48. What if we could automate this
sense making ability?
… and do it efficiently and at scale
49
50. People are good at making sense of sensory input
What can we learn from cognitive models of perception?
• The key ingredient is prior knowledge
51
51. Perception Cycle*
Translating low-level signals
into high-level knowledge
1
Explanation
Perceive
Feature
Observe
Property
Prior Knowledge
Discrimination
2
* based on Neisser’s cognitive model of perception
Focusing attention on those
aspects of the environment that
provide useful information
52
52. To enable machine perception,
Semantic Web technology is used to integrate
sensor data with prior knowledge on the Web
53
53. Prior knowledge on the Web
W3C Semantic Sensor
Network (SSN) Ontology
Bi-partite Graph
54
54. Prior knowledge on the Web
W3C Semantic Sensor
Network (SSN) Ontology
Bi-partite Graph
55
55. Explanation
Explanation is the act of choosing the objects or events that best account for a
set of observations; often referred to as hypothesis building
Translating low-level signals
into high-level knowledge
Observe
Property
1
Explanation
Perceive
Feature
56
56. Explanation
Explanation is the act of choosing the objects or events that best account for a set of
observations; often referred to as hypothesis building
Inference to the best explanation
• In general, explanation is an abductive problem; and
hard to compute
Finding the sweet spot between abduction and OWL
• Simulation of Parsimonious Covering Theory in OWLDL (using the single-feature assumption*)
* An explanation must be a single feature which accounts for
all observed properties
57
57. Explanation
Explanatory Feature: a feature that explains the set of observed properties
ExplanatoryFeature ≡ ∃ssn:isPropertyOf—.{p1} ⊓ … ⊓ ∃ssn:isPropertyOf—.{pn}
Observed Property
elevated blood pressure
clammy skin
palpitations
Explanatory Feature
Hypertension
Hyperthyroidism
Pulmonary Edema
58
58. Discrimination
Discrimination is the act of finding those properties that, if observed, would help distinguish
between multiple explanatory features
Explanation
Perceive
Feature
Observe
Property
Discrimination
2
Focusing attention on those
aspects of the environment that
provide useful information
59
59. Discrimination
To determine which possible observations are most informative, find those observable properties
that can discriminate between the set of hypotheses.
Expected
Properties
Not-applicable
Properties
Discriminating
Properties
Universe of observable properties
60
60. Discrimination
Expected Property: would be explained by every explanatory feature
ExpectedProperty ≡ ∃ssn:isPropertyOf.{f1} ⊓ … ⊓ ∃ssn:isPropertyOf.{fn}
Expected Property
elevated blood pressure
clammy skin
palpitations
Explanatory Feature
Hypertension
Hyperthyroidism
Pulmonary Edema
61
61. Discrimination
Not Applicable Property: would not be explained by any explanatory feature
NotApplicableProperty ≡ ¬∃ssn:isPropertyOf.{f1} ⊓ … ⊓ ¬∃ssn:isPropertyOf.{fn}
Not Applicable Property
elevated blood pressure
clammy skin
palpitations
Explanatory Feature
Hypertension
Hyperthyroidism
Pulmonary Edema
62
63. Resource savings of abstracting sensor data
Orders of magnitude resource savings for generating and storing relevant
abstractions vs. raw observations.
Raw observations
Relevant abstractions
64
64. The Decisions are as Good as the Underlying Coded Knowledge
• How do we know whether we have all possible
relationships?
• How do we know which relationships are missing?
• How can we efficiently fill the missing relationships?
65
65. Semantics Driven Approach for Knowledge Acquisition from EMRs
Knowledge is built by abstracting real world facts, once
built it should be able to explain the real world
Sujan Perera, Cory Henson, Krishnaprasad Thirunarayan, Amit Sheth, Suhas Nair, 'Semantics Driven Approach for Knowledge Acquisition from
EMRs', Special Issue on Data Mining in Bioinformatics, Biomedicine and Healthcare Informatics, Journal of Biomedical and Health Informatics
(To Appear)
66
66. Semantics Driven Approach for Knowledge Acquisition from EMRs
D
D
D
D
Patient Notes
D
UMLS
Explanation
Module
Explained?
No
D
Hypothesis
Generation
Hypothesis
Filtering
Yes
Hypothesis
with High
Confidence
67. The Algorithm
1. Annotate the EMR documents with given knowledgebase
2. Find unexplained symptoms
3. Generate hypothesis for unexplained symptoms
1. All disorders in document becomes candidates
4. Filter out candidate disorder with high confidence
1. Get disorders which has relationship with unexplained
symptom in given knowledgebase
2. Collect the “neighborhood” of the disorders
3. Get the intersection of “neighborhood” and candidate
disorders
68
69. Evaluation
Precision = number of suggested correct relationships
Total number of suggested
= 73.09%
Recall
=
correct relationships found
all correct relationships – known correct relationships
= 66.67%
If we do not perform the semantic filtering step, the precision would be 30%.
High precision is important since it is hard to find domain experts to validate
the generated hypothesis.
70
70. kHealth
knowledge-enabled healthcare
Through physical monitoring and
analysis, our cellphones could act as
an early warning system to detect
serious health conditions, and
provide actionable information
canary in a coal mine
71
72. Risk Score: from Data to Abstraction and Actionable Information
Machine Sensors
Qualities
-High BP
-Increased Weight
Validate correlations
Personal Input
kHealth
Entities
-Hypertension
-Hypothyroidism
Comorbidity risk score
e.g., Charlson Index
- Find correlations
- Validation
- domain knowledge
- domain expert
Model Creation
Parameterize the
model
EMR/PHR
Historical observations
of each patient
Risk Assessment Model
Longitudinal studies of
cardiovascular risks
Risk Score
(Actionable Information)
Current Observations
-Physical
-Physiological
-History
73
73. Asthma
25
million
People in the U.S. are
diagnosed with asthma
(7 million are children)1.
300
million
People suffering from
asthma worldwide2.
$50
billion
Spent on asthma alone
in a year2
155,000
Hospital admissions in
20063
593,000
Emergency department
visits in 20063
1http://www.nhlbi.nih.gov/health/health-topics/topics/asthma/
2http://www.lung.org/lung-disease/asthma/resources/facts-and-figures/asthma-in-adults.html
3Akinbami
et al. (2009). Status of childhood asthma in the United States, 1980–2007. Pediatrics,123(Supplement 3), S131-S145.
77
74. WHY Big Data to Smart Data: Healthcare example
Asthma is a multifactorial disease with health signals spanning personal,
public health, and population levels.
Velocity
Variety
semantics
Value
Can we detect the asthma severity level?
Can we characterize asthma control level?
What risk factors influence asthma control?
What is the contribution of each risk factor?
Veracity
Understanding relationships between
health signals and asthma attacks
for providing actionable information
Volume
Real-time health signals from personal level (e.g., Wheezometer, NO in breath,
accelerometer, microphone), public health (e.g., CDC, Hospital EMR), and
population level (e.g., pollen level, CO2) arriving continuously in fine grained
samples potentially with missing information and uneven sampling frequencies.
78
75. Asthma: Demonstration of Value
Personal
Variety: Health signals span heterogeneous sources
Volume: Health signals are fine grained
Velocity: Real-time change in situations
Veracity: Reliability of health signals may be compromised
Value: Can I reduce my asthma attacks at night?
Population Level
Public Health
Decision support to doctors
by providing them with
deeper insights into patient
asthma care
79
76. Asthma: Actionable Information for Asthma Patients
Can I reduce my asthma attacks at night?
Actionable
Information
Closing the window at home
in the morning and taking an
alternate route to office may
lead to reduced asthma attacks
Population Level
What is the air quality indoors?
Personal
What are the triggers?
Sensordrone – for monitoring
environmental air quality
What is the propensity toward asthma?
What is the wheezing level?
Wheezometer – for monitoring
wheezing sounds
Commute to Work
What is the exposure level over a day?
Public Health
80
77. Personal, Public Health, and Population Level Signals for Monitoring Asthma
Sensors and their observations
for understanding asthma
Asthma Control =>
Asthma Control
and Actionable Information
Daily Medication
Choices for starting
therapy
Not Well Controlled
Poor Controlled
Severity Level of
Asthma
(Recommended Action)
(Recommended Action)
(Recommended Action)
Intermittent Asthma
SABA prn
-
-
Mild Persistent Asthma
Low dose ICS
Medium ICS
Medium ICS
Moderate Persistent
Asthma
Medium dose ICS alone
Or with
LABA/montelukast
Medium ICS +
LABA/Montelukast
Or High dose ICS
Severe Persistent Asthma
High dose ICS with
LABA/montelukast
Needs specialist care
Medium ICS +
LABA/Montelukast
Or High dose ICS*
Needs specialist care
ICS= inhaled corticosteroid, LABA = inhaled long-acting beta2-agonist, SABA= inhaled short-acting beta2-agonist ;
*consider referral to specialist
81
78. Asthma Early Warning Model
(EventShop*)
(kHealth**)
Personal Level Sensors
Personal
Level Signals
Societal Level
Signals
(Personalized
Societal Level Signal)
Storage
Recommended
Action
Action
Justification
Societal Level Sensors
(Personal Level Signals)
Societal Level Signals
Relevant to the
Personal Level
(Societal Level Signals)
Asthma Early Warning Model (AEWM)
Qualify
Action
Recommendation
Quantify
Verify & augment
domain knowledge
Query AEWM
What are the features influencing my asthma?
What is the contribution of each of these features?
How controlled is my asthma? (risk score)
What will be my action plan to manage asthma?
*http://www.slideshare.net/jain49/eventshop-120721, ** http://www.youtube.com/watch?v=btnRi64hJp4
82
79. Health Signal Extraction to Understanding
Physical-Cyber-Social System
Personal
Public Health
Observations
Health Signal Extraction
Health Signal Understanding
PollenLevel
Qualify
<PollenLevel, ChectTightness, Pollution,
Activity, Wheezing, RiskCategory>
Pollution
ChectTightness
<2, 1, 1,3, 1, RiskCategory>
<2, 1, 1,3, 1, RiskCategory>
Wheezing
Acceleration readings from <2, 1, 1,3, 1, RiskCategory>
Activity
on-phone sensors
<2, 1, 1,3, 1, RiskCategory>
Enrich
.
Risk Category assigned by
.
doctors
Wheeze – Yes
PollenLevel
Do you have tightness of chest? –Yes
Quantify
.
Sensor and personal
Signals from personal, personal
Pollution
observations
ChectTightness
spaces, and community spaces
Population Level
Background
Knowledge
Expert
Knowledge
<Wheezing=Yes, time, location>
<ChectTightness=Yes, time, location>
Activity
Wheezing
Outdoor pollen and pollution <PollenLevel=Medium, time, location>
<Pollution=Yes, time, location>
tweet reporting pollution level
and asthma attacks
<Activity=High, time, location>
RiskCategory
Well Controlled - continue
Not Well Controlled – contact nurse
Poor Controlled – contact doctor
83
80. Personal Health Score and Vulnerability Score
At Discharge
Health Score
Non-compliance
Poor economic
status
No living
assistance
Vulnerability
Score
Well Controlled
Low
Well Controlled
Very low
Not Well
Controlled
High
Not Well
Controlled
Medium
Poor Controlled
Very High
Poor Controlled
High
Estimation of readmission vulnerability based on the personal health score
84
81. Health Signal Extraction Challenges
Social streams has been used to extract
many near real-time events
Twitter provides access to rich signals but is noisy,
informal, uncontrolled capitalization, redundant,
and lacks context
We formalize the event extraction from tweets as
a sequence labeling problem
Now you know why you’re miserable! Very High Alert
for B-ALLERGEN Ragweed I-ALLERGEN pollen. B-FACILITY Oklahoma
I-FACILITY Allergy I-FACILITY Clinic says it’s an extreme exposure situation
How do we know the event phrases and who creates
the training set? (manual creation is ruled out)
Idea: Background knowledge used to create the training set
e.g., typing information becomes the label for a concept
85
82. Health Signal Understanding Challenges
Formalized as a problem of structure
extraction of a Bayesian Network
Find the structure that maximize the
scoring function
Huge exponential search space with n
Where Xi represents each
observation
Where k indexes over all
possible graph structures
Where n is the number of nodes
in the network
Different structures may result in the
same structure score (I-Map)
We use declarative
knowledge to choose
between Gi and Gj , and
to guide the search
Ehsan Nazerfard, Bayesian Networks: Structure Learning, Topics in Machine Learning, 2011.
86
83. How do we implement machine perception efficiently on a
resource-constrained device?
Use of OWL reasoner is resource intensive
(especially on resource-constrained devices),
in terms of both memory and time
• Runs out of resources with prior knowledge >> 15 nodes
• Asymptotic complexity: O(n3)
87
84. Approach 1: Send all sensor observations
to the cloud for processing
Approach 2: downscale semantic
processing so that each device is capable
of machine perception
intelligence at the edge
Henson et al. 'An Efficient Bit Vector Approach to Semantics-based Machine Perception in Resource-Constrained Devices,
ISWC 2012.
88
85. Efficient execution of machine perception
Use bit vector encodings and their operations to encode prior knowledge and
execute semantic reasoning
010110001101
0011110010101
1000110110110
101100011010
0111100101011
000110101100
0110100111
89
86. Evaluation on a mobile device
Efficiency Improvement
• Problem size increased from 10’s to 1000’s of nodes
• Time reduced from minutes to milliseconds
• Complexity growth reduced from polynomial to linear
O(n3) < x < O(n4)
O(n)
90
87. Semantic Perception for smarter analytics: 3 ideas to takeaway
1
Translate low-level data to high-level knowledge
Machine perception can be used to convert low-level sensory
signals into high-level knowledge useful for decision making
2
Prior knowledge is the key to perception
Using SW technologies, machine perception can be formalized and
integrated with prior knowledge on the Web
3
Intelligence at the edge
By downscaling semantic inference, machine perception can
execute efficiently on resource-constrained devices
91
88. PREDOSE: Prescription Drug abuse Online-Surveillance and Epidemiology
Bridging the gap between researcher and policy
makers
Early identification of emerging
patterns and trends in abuse
Kno.e.sis - Ohio Center of Excellence in Knowledge-enabled Computing
CITAR - Center for Interventions Treatment and Addictions Research
http://wiki.knoesis.org/index.php/PREDOSE
D. Cameron, G. A. Smith, R. Daniulaityte, A. P. Sheth, D. Dave, L. Chen, G. Anand, R. Carlson, K. Z. Watkins, R. Falck. PREDOSE: A Semantic Web
Platform for Drug Abuse Epidemiology using Social Media. Journal of Biomedical Informatics. July 2013 (in press)
92
89. PREDOSE: Prescription Drug abuse Online-Surveillance and Epidemiology
• Drug Overdose Problem in US
• 100 people die everyday from drug overdoses
• 36,000 drug overdose deaths in 2008
• Close to half were due to prescription drugs
Gil Kerlikowske
Director, ONDCP
Launched May 2011
In 2008, there were 14,800 prescription painkiller
deaths*
*http://www.cdc.gov/homeandrecreationalsafety/rxbrief/
90. PREDOSE: Bringing Epidemiologists and Computer Scientists together
Epidemiologist
Computer Scientist
Interviews
Automatic Data
Collection
Problems
Large Data Sample Sizes
Sample Biases
Automate Information
Extraction & Content Analysis
Online Surveys
Manual Effort
Not Scalable
Qualitative Coding
Group Therapy: http://www.thefix.com/content/treatment-options-prison90683
Access hard-to-reach
Populations
Early Identification and
Detection of Trends
92. Entity Identification
Drug Abuse Ontology (DAO)
subClassOf
subClassOf
+ve
83 Classes
37 Properties
Suboxone
Subutex
Sentiment Extraction
experience sucked
has_slang_term
bupey
feel great
-ve
Buprenorphine
has_slang_term
feel pretty damn good
33:1 Buprenorphine
24:1 Loperamide
bupe
didn’t do shit
bad headache
I was sent home with 5 x 2 mg Suboxones. I also got a bunch of phenobarbital (I
took all 180 mg and it didn't do shit except make me a walking zombie for 2 days).
I waited 24 hours after my last 2 mg dose of Suboxone and tried injecting 4 mg
of the bupe. It gave me a bad headache, for hours, and I almost vomited. I could
feel the bupe working but overall the experience sucked.
Of course, junkie that I am, I decided to repeat the experiment. Today, after
waiting 48 hours after my last bunk 4 mg injection, I injected 2 mg. There wasn't
really any rush to speak of, but after 5 minutes I started to feel pretty damn good.
So I injected another 1 mg. That was about half an hour ago. I feel great now.
Triples
DIVERSE DATA TYPES
ENTITIES
Codes
Triples (subject-predicate-object)
DOSAGE
PRONOUN
Suboxone used by injection, negative experience
Suboxone injection-causes-Cephalalgia
INTERVAL
Route of Admin.
Suboxone used by injection, amount
Suboxone injection-dosage amount-2mg
RELATIONSHIPS
SENTIMENTS
Suboxone used by injection, positive experience
Suboxone injection-has_side_effect-Euphoria
93. PREDOSE: Smarter Data through Shared Context and Data Integration
Ontology
Lexicon
Lexico-ontology
Rule-based Grammar
ENTITIES
TRIPLES
EMOTION
INTENSITY
PRONOUN
SENTIMENT
DRUG-FORM
ROUTE OF ADM
SIDEEFFECT
DOSAGE
FREQUENCY
INTERVAL
Suboxone, Kratom, Herion,
Suboxone-CAUSE-Cephalalgia
disgusted, amazed, irritated
more than, a, few of
I, me, mine, my
Im glad, turn out bad, weird
ointment, tablet, pill, film
smoke, inject, snort, sniff
Itching, blisters, flushing,
shaking hands, difficulty
breathing
DOSAGE: <AMT><UNIT>
(e.g. 5mg, 2-3 tabs)
FREQ: <AMT><FREQ_IND><PERIOD>
(e.g. 5 times a week)
INTERVAL: <PERIOD_IND><PERIOD>
(e.g. several years)
94. PREDOSE: Role of Semantic Web & Ontologies
Data Type
Semantic Web Technique
Entity
Ontology-driven
Identification &
Normalization
Triple
Schema-driven
Sentiment
Ontology-assisted Target
Entity Resolution
Limitations of Other Approaches
ML/NLP
IR
Requires Labeled
Unpredictable
Data
term frequencies
Difficult to
develop
language model
Requires entity
disambiguation
Inconsistent data Diverse simple &
for Parse Trees or complex slang
rules
terms & phrases
95. with it, SOME of it has to make it through? Not sure.”
“Normally around 100 milligrams of loperamide will get me out of withdrawals.”
“Loperamide alone is enough to keep me well without being miserable, IF I megadose.”
Loperamide-Withdrawal Discovery
“This loperamide has saved my life during w/ds.... and made me even more careless
Loperamide is used
with my monthly meds.”to self-medicate to from Opioid Withdrawal symptoms
“But I just wanted to tell you that loperamide WILL WORK. I take 105 mg of
methadone/day, and recently have been running out early due to a renewed interest in
IVing that shit. 200mg of lope 100 pills will make me almost 100 again. It brings the
sickness down to the level of, say, a minor flu. Sleep returns, restlessness dissipates.
dose of 16 mg per day. For example, web forum participants shared the following opinions:
Sometimes a mild opiation is felt.”
“Back in the day when I would run out of pills early I would take 8-10 Lopermide tabs and
“So you just stick with it. Don’t go and score big with your next paycheck. Overcome the
get some pretty good relief from w/d.”
need to make everything numb. Learn to live with normality for a while. It’ll all seem
worthwhile soon enough. Go for a like Get out of the house. Go grab some loperamide
“If you take a shitload of loperamidewalk.10-20 pills at once in withdrawal, you’ll get relief
from the store, the desperate junky’s methadone.”
from some of the physical symptoms. Im not sure exactly how it works, but it’s definitely
MORE than just relieving the GI symptoms. Im guessing if you just bombard your blood
The most commonly of it has to side effects of loperamide use were constipation, dehydration
with it, SOME discussed make it through? Not sure.”
and other types of gastrointestinal discomforts. Some also reported mild withdrawal symptoms
from using loperamide for anmilligrams period of time.will get me out of withdrawals.”
“Normally around 100 extended of loperamide
“Loperamide is good for a day keep but the without is on loperamide I I megadose.”
“Loperamide alone is enough toor twome well problembeing miserable, IF lose all desire to
eat OR drink, or do anything really.”
“I used to sing the praises my life during w/ds.... and made short term standby until
“This loperamide has savedof loperamide....and still do, as a me even more careless you
can score. Long term maintenance, it really wears you out. Starts to “feel” toxic though I
with my monthly meds.”
99
96. Big Data vs. Smart Data in Digital Health (Healthcare provider)
Big Data from Healthcare
Expert opinion
Clinical research
Population
health record
Personal health
record
Clinical decision
support
Smart Data for Healthcare
What is the overall health of the person?
What are the vulnerabilities for organ
transplant?
Red, yellow, and green indicate high,
medium, and low risk allowing decision
makers to focus on red & yellow variables
Ms. Mecannic’s
blood test not
yet complete
100
97. Big Data vs. Smart Data in Digital Health (Healthcare consumer)
Big Data from Healthcare
Smart Data for Healthcare
What are the reasons for my increasing weight?
What should I consider before I get a kidney transplant?
Personal
Schedule
Recommendation algorithms
will analyze data deluge with
domain knowledge
Sleep data
Activity data
Red, yellow, and green
indicating high, medium, and
low risk factors
Personal health
records
Community data
Ms. Mecannic: Your
blood work is
incomplete. Please
finish this before
organ donation!
101
98. Demos
• Real Time Feature Streams:
http://www.youtube.com/watch?v=_ews4w_eCpg
• kHealth: http://www.youtube.com/watch?v=btnRi64hJp4
• PREDOSE: https://www.youtube.com/watch?v=gCFPzMgEPQM
102
99. Take Away
• Data processing for personalized healthcare is lot more
than a Big Data processing problem
• It is all about the human – not computing, not device: help
them make better decisions, give actionable information
– Computing for human experience
• Whatever we do in Smart Data, focus on human-in-the-loop
(empowering machine computing!):
– Of Human, By Human, For Human
– But in serving human needs, there is a lot more than what
current big data analytics handle – variety, contextual,
personalized, subjective, spanning data and knowledge across PC-S dimensions
103
100. Acknowledgements
• Kno.e.sis team
• Funds: NSF, NIH, AFRL, Industry…
•
•
•
Note:
For images and sources, if not on slides, please see slide notes
Some images were taken from the Web Search results and all such images belong
to their respective owners, we are grateful to the owners for usefulness of these
images in our context.
104
101. References and Further Readings
•
•
•
•
OpenSource: http://knoesis.org/opensource
Showcase: http://knoesis.org/showcase
Vision: http://knoesis.org/vision
Publications: http://knoesis.org/library
105
102. Alan Smith
Wenbo
Wang
Vinh
Nguyen
Sujan
Perera
Hemant
Purohit
Cory Henson
Pramod Koneru
Amit Sheth’s
PHD students
Maryam Panahiazar
Kalpa
Gunaratna
Ashutosh Jadhav
Sanjaya
Wijeratne
Sarasi Lalithsena
Pramod
Anantharam
Pavan
Kapanipathi
Lu Chen
Delroy
Cameron
Kno.e.sis in 2012 = ~100 researchers (15 faculty, ~50 PhD students)
103. Smart Data
thank you, and please visit us at
http://knoesis.org/vision
http://knoesis.org/amit/hcls
Kno.e.sis – Ohio Center of Excellence in Knowledge-enabled Computing
Wright State University, Dayton, Ohio, USA
107
Hinweis der Redaktion
Starting slide Various Big data problems – Traditional examples vs what we are doing examples. Variety and Velocity than Volume. kHealth problem. People will be interested in Smart Data.Traditional ML techniques, High Performance Computing, Statistics. Human level of Abstraction is Smart data.
"2600 BC – Imhotep wrote texts on ancient Egyptian medicine describing diagnosis and treatment of 200 diseases in 3rd dynasty Egypt.”Sir William Osler, 1st Baronet, was a Canadian physician and one of the four founding professors of Johns Hopkins Hospital. He was called the father of modern medicine. Sir William Osler called Imhotep as the true father of medicine.
Larry Smarr is a professor at the University of California, San DiegoAnd he was diagnosed with Crones DiseaseWhat’s interesting about this case is that Larry diagnosed himselfHe is a pioneer in the area of Quantified-Self, which uses sensors to monitor physiological symptomsThrough this process he discovered inflammation, which led him to discovery of Crones DiseaseThis type of self-tracking is becoming more and more common
http://radhakrishna.typepad.com/rks_musings/2013/04/big-data-review.htmlGoogle predicted the spread of flu in real time - after analyzing two datasets, a.) 50 million most common terms that Americans type, b.) data on the spread of seasonal flu from public health agency- tested a mammoth of 450 million different mathematical models to test the search terms, comparing their predictions against the actual flu cases- model was tested when H1N1 crisis struck in 2009 and gave more meaningful and valuable real time information than any public health official system (Big Data, Viktor Mayer-Schonberger and Kenneth Cukier, 2013)
Better Algorithms Beat More Data — And Here’s Whyhttp://allthingsd.com/20121128/better-algorithms-beat-more-data-and-heres-why/Big Data Cannot Replace Human Judgmenthttp://www.matchcite.com/blog/blog/2012/july/big-data-cannot-replace-human-judgment.aspx**Comments about the articles
Smart data makes sense out of big data – it provides value from harnessing the challenges posed by volume, velocity, variety and veracity of big data, to provide actionable information and improve decision making.
Information is CREATED by human with the Machinery available – Wikipedia tool, sensors and social networksInformation is STORED in Man+Machine readable format, LODInformation is PROCESSED using the LOD and Human assisted Knowledge-basedHigher level abstraction on info is now consumed in many mechanistic ways (including GIS) to provide EXPERIENCE for humans
- HUMAN CENTRIC!!
All the data related to human activity, existence and experiencesMore on PCS Computing: http://wiki.knoesis.org/index.php/PCS
Information is CREATED by human with the Machinery available – Wikipedia tool, sensors and social networksInformation is STORED in Man+Machine readable format, LODInformation is PROCESSED using the LOD and Human assisted Knowledge-basedHigher level abstraction on info is now consumed in many mechanistic ways (including GIS) to provide EXPERIENCE for humans Example of a human guided modeling and improved performancehttp://research.microsoft.com/en-us/um/people/akapoor/papers/IJCAI%202011a.pdf
Also, we have weather application which performs abstraction on weather sensory observations to identify blizzard conditions (food for actions!!) :--20,000 weather stations (with ~5 sensors per station)-- Real-Time Feature Streams - live demo: http://knoesis1.wright.edu/EventStreams/ - video demo: https://skydrive.live.com/?cid=77950e284187e848&sc=photos&id=77950E284187E848%21276
Such a blackout would cause billions of dollars in lost revenue. This particular blackout resulted in 6 billion dollar loss.Not only the consumers lose revenue even the power utility companies are fined almost a million dollars a day.
Two of the problems are problems in “understanding” the system and the available data/observations. Lack of experience/training in deciphering this data had serious implications.The items 1 and 2 related to understanding real-world complexity by incorporating multi-modal and multi-sensory observations. Algorithms that can provide abstractions to decision makers for better comprehension of the situation.Providing abstractions in the context of the grid state continuously would lead to actionable information to decision makers.
Doctors – error in judgment leads to legal complications and psychological stress Patients – life changing (social, psychological, economical changes) and even losing life at extreme situations!
Core concepts of all sensing systems
- what if we could automate this sense making ability?- and what if we could do this at scale?
sense making based on human cognitive models
perception cycle contains two primary phasesexplanationtranslating low-level signals into high-level abstractions inference to the best explanationdiscriminationfocusing attention on those properties that will help distinguish between multiple possible explanationsused to intelligently task sensors and collect additional observations (rather than brute force approach of blindly collecting all observations)
perception cycle contains two primary phasesexplanationtranslating low-level signals into high-level abstractions inference to the best explanationdiscriminationfocusing attention on those properties that will help distinguish between multiple possible explanationsused to intelligently task sensors and collect additional observations (rather than brute force approach of blindly collecting all observations)
A single-feature (disease) assumption means that all the observed properties (symptoms) must be explained by a single feature.i.e., this framework is not expressive enough to model comorbidity where there may be more than one feature (disease) co-existing For example, if there are two diseases causing disjoint symptoms, and all the symptoms of both the diseases are observed, then this framework will not be able to find the coverage and returns no diseases.Parsimony criteria is single feature assumption to choose from among multiple explanationsNot true: if multiple disease account for single property…Rewrite with more relaxed parcimony criteria (complex, cannot be modeled in OWL)Make KB more intelligent: create an individual that represents the two disease which together explain a symptom
perception cycle contains two primary phasesexplanationtranslating low-level signals into high-level abstractions inference to the best explanationdiscriminationfocusing attention on those properties that will help distinguish between multiple possible explanationsused to intelligently task sensors and collect additional observations (rather than brute force approach of blindly collecting all observations)
So check galvanic skin response sensor
The problem’ addressed by the JBHI paper
Background knowledge is used to explain the patient notes.The explain means each symptom should be explained by at least one disorder in the documentsIf there is at least one symptom which is not explained, then we generate hypothesis based on this observation.Initially all the disorder in the document becomes candidatesBy we developed a filtering mechanism to filter out hypothesis with low confidenceWe generate hypothesis with high confidence
S1 is the unexplaned symptomD1 to D5 are the disorders in the documentSo initially D1 to D5 are candidates to have relationship with S1Within our background knowledge S1 linked to D8 and D12We collect the neighborhood of D8 and D12Then check the intersection with D1 to D5 and collected neighborhoodIf there are common disorders, they become the candidates with high confidenceSo D5 and D2 are the best candidates to have relationship with S1 from initial set (D1 to D5)
Precision = 125/171Recall = 44/ (109-44) = 66/44
- With this ability,many problems could be solved- For example: we could help solve health problems (before they become serious health problems) through monitoring symptoms and real-time sense making, acting as an early warning system to detect problematic health conditions
Massive amount of data will be collected by sensors and mobile devices yet patients and doctors care about “actionable” information.This data has all the four Vs of big data and we used knowledge enabled techniques to transform it into valueIn the context of PD, we analyzed massive amount of sensor data collected by sensors on a smartphones to understand detection and characterization of PD severity.
Main idea: Prior knowledge of PD was used to facilitate its detection from massive sensor data by reducing the search spaceDetails:Declarative knowledge of PD includes PD severity and their symptoms as shown in the logical rule aboveEach PD severity level is a conjunction of a set of PD symptomsEach symptom was mapped to its manifestation in sensor observationsThe availability of declarative knowledge significantly improved the analytics by aiding feature selection processThe graphs above contrasts the physical movements and voice of two control group members and two PD patients
[WM-13] Wheezometer by iSonea, Available online at: http://www.isoneamed.com/wheezometer.html (Accessed May 13, 2013).[NOS-13] Nitric Oxide Sensor, Available online at: http://nodesensors.com/product/oxa-gas-sensor-nitric-oxide-no/ (Accessed May 13, 2013).[SD-13] Sensordrone, a bluetooth enabled low-cost sensor for monitoring the environment, Available online at: http://www.kickstarter.com/projects/453951341/sensordrone-the-6th-sense-of-your-smartphoneand-be/ (Accessed May 31, 2013).[ODS-13] Optical Dust Sensor, Available online at: https://www.sparkfun.com/products/9689 (Accessed May 13, 2013).[ESP-13] Everyaware, Sensing Air Pollution, Available online at: http://www.everyaware.eu/activities/case-studies/air-quality/ (Accessed May 31, 2013).[AQ-13] Community-led sensing of AirQuality, Available online at: http://airqualityegg.com/ (Accessed May 13, 2013).[NLAF-13] National and Local Allergy Forecast, Available online at: http://www.pollen.com/allergy-weather-forecast.asp (Accessed May 13, 2013).[NABA-13] National Allergy Bureau Alerts, Available online at: http://www.aaaai.org/global/nab-pollen-counts.aspx (Accessed May 13, 2013).[AQI-13] Air Quality Index from United States Environmental Protection Agency, Available online at : http://www.epa.gov/(Accessed May 23, 2013).[CDC-13] Centers for Disease Control and Prevention, Available online at: http://www.cdc.gov/ (Accessed May 23, 2013).
Non-compliance, Poor economic status and No living assistance are good predictors for readmission
Only score based structure extraction is presented here. Other popular structure extraction techniques include constraint based approaches which finds independences between random variables X1, …, XnI-Map => different structures result in the same loglikelihood score. Thus recovering the original structure of the graph generating data using data alone is considered impossible! We go the the rescue of declarative knowledge to: (1) choose promising structures and (2) to break ties when two structure results in the same score
Intelligence distributed at the edge of the networkRequires resource-constrained devices (mobile phones, gateway notes, etc.) to be able to utilize SW technologies
Intelligence distributed at the edge of the networkRequires resource-constrained devices (mobile phones, gateway notes, etc.) to be able to utilize SW technologiesHenson et al. 'An Efficient Bit Vector Approach to Semantics-based Machine Perception in Resource-Constrained Devices, ISWC 2012.
compute machine perception inferences -- i.e., explanation and discrimination -- of high-complexity on a resource-constrained devices in milisecondsDifference between the other systems and what this system provides
Intelligence at the age. Shipping computation and domain models to the edge (Distributed)
For every 1 death from prescription drug overdose there are:10 users admitted for treatment 32 users admitted to the emergency department 130 people who are users/dependent 825 non-medical users of prescription drugsWhite House Office of National Drug Control Policy (ONDCP) launched Epidemic (May 24, 2011)
Epidemiologist’s ApproachData collection from interviews, surveysContent Analysis using CodingComputer Scientists’ Approach Automate Data Collection Multiple sources of rich data Automate Content Analysis Information Extraction Trend Analysis
Sample post from a user that was just discharged from rehab facility. Sent home with Suboxone and Phenobarbital treatment drugsPhenobarbital - an anti-anxiety and anticonvulsant barbiturate, used to treat anxiety and seizures This post contains entities, which require structured representations to resolve.We created the Drug Abuse Ontology (DAO) first ontology for prescription drug abuse.The ontology is very important because of the pervasive use of slang.In a manually created gold standard set of 601 posts the following was observed: 33:1 Buprenorphine 24:1 Loperamide
INTENSITY – more than, abnormal, in excess of, too muchDRUG-FORM – ointment, tablet, pill, filmINTERVAL – for several years
Loperamide is sold over the counter (OTC) in ImodiumYellow – positive sentimentsPink – EntitiesGreen – curious finding - indication of getting high in the processMention the practice of Megadosing!!
More at: http://wiki.knoesis.org/index.php/PCSAnd http://knoesis.org/projects/ssw/