The potential of personalized medicine based on machine learning is huge, but big challenges must be overcome to implement this technology in practice. Hidde will discuss both sides of the story, including a case study on the intensive care.
5. Pacmed combines medical expertise with machine learning to make
health care more personal and precise
Goals
How
Clinical Decision Support based on advanced data science and medical
expertise that continually improve over time
Patients only receive care that will truly improve their quality of life
Every doctor learns directly from all decisions all their colleagues make
Limited health care resources are spent most efficiently
Help hospitals use full potential machine learning
6. The Pacmed team combines machine learning knowledge with
medical expertise
7. Based on experience with several (implemented) projects, Pacmed
identified 6 areas in healthcare to focus on
Area
Oncology
(Colon-, Prostate-, Breast- and Lungcancer)
Intensive Care
Cardiology
Expensive Treatments
(Rheumatism, IBD)
Triage Emergency Care
(Risk of) Heart Diseases
(Hypertension, Diabetes, Chronic Kidney
Failure)
Rationale
Large scale, high impact decisions, high health care costs and lot of
outcome data
Lot of data, high impact, high costs
Large scale, high costs, high impact, lot of relevant data
Lot of opportunity for improvement, high costs, clear business case
Large problem, good Proof of Concept, easy to implement
High impact, large scale, a lot variation between patient, experience and
opportunities for Pacmed
Hospital Care
Primary Care
9. In classical medical research a hypothesis is formed on a certain
medical problem
Sys.Bloodpressure(mmHg)
Age (years)
5040 60 80 9070
75
50
100
150
175
125
Readmitted
Not readmitted
10. Data can be used to test such an hypothesis
Sys.Bloodpressure(mmHg)
Age (years)
5040 60 80 9070
75
50
100
150
175
125
11. Data can be used to test such an hypothesis…
Sys.Bloodpressure(mmHg)
Age (years)
5040 60 80 9070
75
50
100
150
175
125
12. ..but it can also be used to generate new knowledge
Sys.Bloodpressure(mmHg)
Age (years)
5040 60 80 9070
75
50
100
150
175
125
13. Machine learning techniques can be used to find more complex
relations in this data..
Sys.Bloodpressure(mmHg)
Age (years)
5040 60 80 9070
75
50
100
150
175
125
15. Medical knowledge is currently based on average results on a small
population of patients
Biased research on a
small unrepresentative
number of patients
Applied on a diverse and complex group in practice
Crude and
slow
guidelines
Health care costs are rising
Number of treatment options are rising
16. Expected
outcomes for
individual patient
Learn from every patient
Real-time
machine learning
Through the digitisation of the world, machine learning can be used to
present the expected outcomes for the individual patient
17. Through the digitisation of the world, machine learning can be used to
present the expected outcomes for the individual patient
Expected
outcomes for
individual patient
Learn from every patient
Real-time
machine learning
18. This allows us to move from an average treatment effect to an
individual treatment effect
Average treatment effect Individual treatment effect
20. Agenda
1 Introduction
2 The potential of machine learning in healthcare
3 Case study intensive care: practical challenges
From data to information
From performance to impact
Explainability: building the doctor’s trust
Causal inference in light of selection bias
a
b
c
d
21. The most critically ill patients of a hospital go to the intensive care,
where all vital signs are monitored
22. The aim of the intensive care is to make sure patients have recovered
enough to go to a regular department
23. At another department patients can get a complication, which can
cause a readmission to the intensive care
Readmissions to the ICU
are associated with adverse outcomes
Higher costs
Increased length of stay
Increased risk of mortality
A complication can cause a patient to be
readmitted to the ICU
24. Agenda
1 Introduction
2 The potential of machine learning in healthcare
3 Case study intensive care: practical challenges
From data to information
From performance to impact
Explainability: building the doctor’s trust
Causal inference in light of selection bias
a
b
c
d
25. A brief summary of information included in the model
Patient and admission characteristics
Clinical observationsDevice data
Lab values Medication data
• Age, gender
• Length, weight, BMI
• Department of origin
• Plannend/unplanned admission
• Length of stay
• Number of prior admissions
• Vital parameters
• Blood pressure
• Heart rate
• Respiratory rate
• Temperature
• Settings: amount of support given
• Blood counts
• Renal function
• Liver tests
• Cardiac enzymes
• Blood gases
• Heart rhythm
• Supplemental oxygen
• Respiratory problems
• Urine output
• Consciousness
• Confusion
• Pumps / injections / via tube
• Vasopressors
• Inotropes
• Sedation
26. For all vital parameters we start with a time series during the admission
27. We use extensive feature engineering to go from raw data to information…
first
minimum
last
average
slope standard deviation
maximum
{…}
counts
28. …calculated for the first day, last day and whole admission
{…} {…}
{…}
1
2
3
First, last,
average, min,
max, std and
slope over the
whole admission
Mean, standard
deviations and
slopes over the
days
First and last 24
hours are
considered
independently
29. Agenda
1 Introduction
2 The potential of machine learning in healthcare
3 Case study intensive care: practical challenges
From data to information
From performance to impact
Explainability: building the doctor’s trust
Causal inference in light of selection bias
a
b
c
d
31. Design of decision support: an overview of all patients on the ICU to
determine who can be discharged
32. Agenda
1 Introduction
2 The potential of machine learning in healthcare
3 Case study intensive care: practical challenges
From data to information
From performance to impact
Explainability: building the doctor’s trust
Causal inference in light of selection bias
a
b
c
d
33. How do you make machine learning models explainable to a doctor?
Global feature importance Local feature importance
Model specific importance:
For certain models feature importance can be directly calculated
Model-agnostic importance:
Decreases in performance when
permutating a single feature
LIME:
Local surrogate model that explains an indivdual prediction
Shapley values:
Division of predictive power over all features (based on
game theory)
35. Agenda
1 Introduction
2 The potential of machine learning in healthcare
3 Case study intensive care: practical challenges
From data to information
From performance to impact
Explainability: building the doctor’s trust
Causal inference in light of selection bias
a
b
c
d
36. The problem of causality: selection bias in practice
Ideal situation
Real situation in practice
Problem: can you trust the observed treatment effect?
37. More and more machine learning methods try to account for this
effect
38. Agenda
1 Introduction
2 The potential of machine learning in healthcare
3 Case study intensive care: practical challenges
From data to information
From performance to impact
Explainability: building the doctor’s trust
Causal inference in light of selection bias
a
b
c
d
39. Machine learning has huge potential in healthcare, but some
challenges have to be overcome to actually implement it in practice!