Talk at Wearable 2016 Symposium in Lausanne.
This presentation talks about use of wearables and other sensors for quantifying lifestyle and relating it to build model of personal health.
5. Event Mining
Machine Learning
Individual Model
‘Likely to get severe heart
attack in 10 minutes – get him
help immediately.’
• Individual Model from data.
• Health, social, personal.
• Actionable Predictive use.
• Better disease models.
Healthcare 2020
Objective Self
6. Important Revolution in Health
In the mid-20th century, the primary causes of death
worldwide shifted from infections to chronic conditions.
13. Big 3: Lifestyle Factors
Physical
Activities
Popular now: Output and state
Food
Starting to happen:
Input
Environmental
Factors
Coming soon:
Surround
14. Healthcare Analytics is Data Rich Now
Past: Data is expensive and small
• Input data is mostly from clinical trials
• Models are small since data is limited
• Personal information is anecdotal
Today: Data is cheap and large
• Longitudinal data
• Heterogeneous data
• Diverse data from Electronic Health
Records and wearable/smartphone
35. Asthma Risk Factor Recognition
1) Pollution increases
suddenly followed by high
wind while temperature
increases slightly will cause
an asthma
Outbreak within 2 days.
2) Thunderstorm followed
by temperature decreases
steadily will cause an
asthma outbreak within 1
day.
36. Results
Temperature fluctuation has the most impact in the fall and winter seasons
and it is not a risk factor during spring or summer
During spring and summer, when rain suddenly increases to a very high level,
an asthma outbreak is more probable
The effect of PM2.5 is not noteworthy in
the fall and winter seasons.
37. Results (Cont.)
• When PM2.5 increases followed by temperature stay high within 3 days,
then asthma outbreak is probable.
• When wind decreases followed by PM2.5 increases within 5 days, then
asthma outbreak is probable.
• When rain increases followed by PM2.5 stay low within 4 days then an
asthma outbreak is probable.
38. • Exposure to polluted air is a risk factor
of asthma attack within X hour?
Personicle case
Environmental
factors
39. Spicy Indian food and 2 glasses of wine
result in severe acidity and sleepless nights.
Personicle
Food Stream
t1 t2 t3 t4 t5
41. Situation awareness is knowing
what’s going on around you.
41
Observe + Orient = Situational Awareness
Orient: Baselines, Goals, and Action Plans
Situation awareness is required for
every decision in life.
Act Orient
Observe
Decide
42. Dashboards Display Data and Information
42
Operators use relevant information to understand
situation to decide relevant Action.
43. 10/15/2016 43
Calendar PESi
FMB (Individual’s Feeling)
Accelerometer
Location
Fitness Data
(Nike, Fitbit) Data
Ingestion &
Aggregation
Heart Rate
Location (Move)
Food Log
FMB
(People’s Feeling, Location)
ESOzone
CO2
SO2
PM 2.5
Pollen (Tree, Grass)
Air Quality Index
Data
Ingestion &
Aggregation
Social Media
(News, Tweets)
Weather
Macro
Situation Recognition
Predictive Analytics
Personal
Situation Recognition
Persona
Asthma Allergy App Server
Data Collection
MacroSituationPersonalSituation
Need and Resources
Recommendation
44. Great Opportunity: Wearable
• Wearable building personal models.
• Using personal model to make people aware
of their situation.
• Informing relevant people about the situation.
• Helping people make
Right Decision,
Right Moment,
Right Place.
Editor's Notes
sensors used to gather climate information, posts to social media sites, digital pictures and videos, purchase transaction records, and cell phone GPS signals, human motions, etc…
Not only heterogeneous, and unstructured, but time is a very important dimension.
They all generate Events
What about frameworks that consume events for knowledge discovery?
In particular, the accommodation of time into mining techniques provides a window into the temporal arrangement of events and, thus, introduce the ability to suggest cause and effect that are overlooked when the temporal component is ignored or treated as a simple numerical attribute. Moreover, temporal data mining has the ability to mine the behavioral aspects of a system as opposed to mining rules that describe their states at a point in time. Hence, temporal data mining helps understanding why rather than merely understanding what.
We can formulate complex patterns using the pattern formulation operators.
These patterns will be translated to their corresponding automata
Fundamentally people can’t glean insights from machine data when they don’t know where to look or what to look for.
Goal: Study the problem of event-driven causality from a general and rigorous point of view.
Contributions: Design and implementation of a framework that facilitates qualitative causal inference on event-based multimedia data streams.
Define meaningful event streams.
Define a high level pattern formulation language.
Design fast algorithms to find causal patterns.
User interfaces that enable analyst to interact smoothly with data, ask better questions, and make better decisions.
Analyst can import pre-processed data (i.e. event streams) and choose visual bottom-up operators to explore data, generate a basic model and derive a preliminary insight. Then she can seed a hypothesis and grow it step by step using the top-down pattern formulation operators. A good hypothesis is not the one that is necessarily correct, but one that opens up a new path of investigation. In complex problem domains this path cannot be fully perceived in advance. So analyst must be provided with appropriate operators to carry out new analyses based on the original hypothesis.
Data Selection Panel: To select different event and event streams from database.
Data-driven Operator Panel: These operators are used for pattern mining from event streams.
Hypotheses-driven Operator Panel: These operators are used for pattern formulation and pattern query from input events.
Visualization panels: Results of pattern mining and pattern query are displayed visually in form of co-occurrence matrices and histograms.
We have meteorological and pollution data streams. Tweets related to asthma and some burst point in asthma histogram is also shown.
How can we understand risk factors from these time-series data?
We use our awesome framework to study complex risk factor patterns :D