The document describes an AI-based framework for clinical decision support. It proposes a hybrid system that integrates machine learning and knowledge-based reasoning. The framework includes components like a patient records database, prescription protocols, and a drug interaction registry. It experimentally compares the performance of learning-based and knowledge-based systems on a sleep aid recommendation task. The learning system uses AdaBoost to handle missing patient data. The framework aims to offer accurate medical decision support by combining the strengths of machine learning and ontological reasoning approaches.

1. AI & eHealth
Noor Orfahly

2. Overview
Definition
Forms of
eHealth
The 10 e’s in
eHealth
AI-Based
MDSS

3. Definition
eHealth is a new term dating back to 1999.
eHealth Strategy Office – UBC: Using modern information and
communications technologies in the areas of health services,
health education and health research.
The World Health Organization: The transfer of health resources and
health care by electronic means. E-health provides a new method for using
health resources – such as information, money, and medicines – and in
time should help to improve efficient use of these resources.
Health Canada: An overarching term used today to describe the application
of information and communications technologies in the health sector. It
encompasses a whole range of purposes from purely administrative
through to health care delivery.

4. Forms of eHealth
Primary care:
 The use of computer systems by general practitioners and
pharmacists for patient management, medical records and electronic
prescribing.
Hospital care:
 ePatient administration systems
 Laboratory & radiology information systems
 Electronic messaging systems
 Telemedicine
Home care:
 Teleconsults
 Remote vital signs monitoring systems used for diabetes medicine,
asthma monitoring and home dialysis systems

5. The 10 e’s in eHealth
efficiency
enhancing quality
evidence based
empowerment
encouragement
education
enabling
extending
ethics
equity

6. The 10 e’s in eHealth
efficiency
enhancing quality
evidence based
empowerment
encouragement
education
enabling
extending
ethics
equity
Increase efficiency in health care:
 Decreasing costs: avoiding
duplicative or unnecessary diagnostic.
 Therapeutic interventions: through
enhanced communication possibilities
between health care establishments,
and through patient involvement.

7. The 10 e’s in eHealth
efficiency
enhancing quality
evidence based
empowerment
encouragement
education
enabling
extending
ethics
equity
E-health may enhance the quality of
health care for example by allowing
comparisons between different
providers, involving consumers as
additional power for quality assurance,
and directing patient streams to the
best quality providers.

8. The 10 e’s in eHealth
efficiency
enhancing quality
evidence based
empowerment
encouragement
education
enabling
extending
ethics
equity
eHealth interventions should be
evidence-based in a sense that their
effectiveness and efficiency should not
be assumed but proven by rigorous
scientific evaluation.

9. The 10 e’s in eHealth
efficiency
enhancing quality
evidence based
empowerment
encouragement
education
enabling
extending
ethics
equity
By making the knowledge bases of
medicine and personal electronic
records accessible to consumers over
the Internet, e-health opens new
avenues for patient-centered medicine,
and enables evidence-based patient
choice.

10. The 10 e’s in eHealth
efficiency
enhancing quality
evidence based
empowerment
encouragement
education
enabling
extending
ethics
equity
Encouragement of a new relationship
between the patient and health
professional, towards a true
partnership, where decisions are made
in a shared manner.

11. The 10 e’s in eHealth
efficiency
enhancing quality
evidence based
empowerment
encouragement
education
enabling
extending
ethics
equity
Education of physicians through online
sources (continuing medical education)
and consumers (health education,
tailored preventive information for
consumers).

12. The 10 e’s in eHealth
efficiency
enhancing quality
evidence based
empowerment
encouragement
education
enabling
extending
ethics
equity
Enabling information exchange and
communication in a standardized way
between health care establishments.

13. The 10 e’s in eHealth
efficiency
enhancing quality
evidence based
empowerment
encouragement
education
enabling
extending
ethics
equity
Extending the scope of health care
beyond its conventional boundaries.
 Geographical sense: e-health enables
consumers to easily obtain health
services online from global providers.
 Conceptual sense: These services can
range from simple advice to more
complex interventions or products such
a pharmaceuticals.

14. The 10 e’s in eHealth
efficiency
enhancing quality
evidence based
empowerment
encouragement
education
enabling
extending
ethics
equity
 Online professional practice
 Informed consent
 Privacy

15. The 10 e’s in eHealth
efficiency
enhancing quality
evidence based
empowerment
encouragement
education
enabling
extending
ethics
equity
  Equitable health care is one of the
promises of e-health.
  People, who do not have the money,
skills, and access to computers and
networks, cannot use computers effectively.
 The digital divide currently runs between
 rural vs. urban
 rich vs. poor
 young vs. old
 male vs. female
 neglected/rare vs. common diseases.

16. A FRAMEWORK FOR AI-BASED
CLINICAL DECISION SUPPORT
Example
Enable medical decision making in the presence of partial information

17. AI-Based Clinical Decision Support Framework
Introduction
Medical decision support systems (MDSS) map patient
information to promising diagnostic and treatment paths.
Knowledge-based systems can suffer a significant loss of
performance when patient data is incomplete:
 Patients omit details
 Access restrictions prevent viewing of remote medical records
The output of learning-based systems cannot be easily explained.
Hybrid System

18. AI-Based Clinical Decision Support Framework
Knowledge Base
Patient
Records
Prescription
Protocol
Drug
Interaction
Registry

19. AI-Based Clinical Decision Support Framework
Patient
Records
 Insomnia treatment.
 Patient records drawn from the Center for Disease Control (CDC).
 Dataset: Behavioral Risk Factor Surveillance System (BRFSS)
telephone survey for 2010.
 BRFSS dataset contains:
 Respondent information: age, race, sex, and geographic location.
 Common medical conditions: cancer, asthma, mental illness,
and diabetes.
 Behavioral risk factors: alcohol consumption, drug use, and sleep
deprivation.
 BRFSS: 450,000 individuals.
 Relational database.

20. AI-Based Clinical Decision Support Framework
Patient
Records

21. AI-Based Clinical Decision Support Framework
Patient
Records

22. AI-Based Clinical Decision Support Framework
Prescription
Protocol
 Identify a subset of sleep aids and apply the Mayo clinic sleep aid
prescription protocol to identify the conditions under which each drug
should be prescribed.
 Inference Rules examples:
1. drug-to-drug interaction rule:
If a patient is currently taking an existing drug D1, and D1 cannot
be given with drug D2, then the patient cannot be given drug
D2.
2. drug-to-condition interaction rule:
If a patient has some existing medical condition C, and a drug D
has contraindication to the condition C, then the patient cannot be
given drug D.
3. drug-to-disease interaction rule:
If a patient has a disease E, and a drug D has contraindication to
the disease E, then the patient cannot be given drug D.
Drug
Interaction
Registry

23. AI-Based Clinical Decision Support Framework
Imputation
Bayesian multiple imputation
Assume a particular joint probability model over the feature values.
a1, a2, …, an , P(a1 = x, a2 = y, …, an = z) = prop., etc
Draw imputed datasets from the posterior distribution of the missing data
given the observed data.
Make multiple imputed datasets, then take the average of the imputed
values.
a1 a2 … an
id1 x y zz
id2 x yyy z
id3 xx yy zzzz
a1 a2 … an
id1 x y ?
id2 ? ? z
id3 ? yy zzzz

24. AI-Based Clinical Decision Support Framework
Experimental Comparison
Patients who should be given sleep aids were labeled as ‘positive’
exemplars and those who should not as ‘negative’ exemplars.
When a system labeled a patient correctly in response to a query, a
‘true positive’ (tp) or ‘true negative’ (tn) was produced.
Otherwise, a ‘false positive’ (fp) or ‘false negative’ (fn) was
produced.
The results were evaluated in terms of:
 Sensitivity: rate of positive exemplars labeled as positive.
 Specificity: rate of negative exemplars labeled as negative.
 Balanced accuracy: simple average of specificity and sensitivity.

25. AI-Based Clinical Decision Support Framework
Experimental Comparison – Learning-based System
Evaluate the impact of missing information on the performance of the
learning-based system by removing known values from the patient
records.
Defined e as the average number of attribute values removed from a
patient’s record.
For each value of e, train an AdaBoost-based classifier using 50 sets of
5000 exemplars from the partially-missing data.

26. AI-Based Clinical Decision Support Framework
Experimental Comparison – Learning-based System
AdaBoost
AdaBoost helps you combine multiple “weak classifiers” into a single
“strong classifier”.
A weak classifier is simply a classifier that performs poorly, but performs
better than random guessing (accuracy is greater than 50%).
AdaBoost can be applied to any classification algorithm.
What does AdaBoost do for you?
1. It helps you choose the training set for each new classifier that you train based
on the results of the previous classifier.
2. It determines how much weight should be given to each classifier’s proposed
answer when combining the results.

27. AI-Based Clinical Decision Support Framework
Experimental Comparison – Learning-based System
AdaBoost
Training Set Selection:
 Each weak classifier should be trained on a random subset of the total training set.
 The subsets can overlap.
 AdaBoost assigns a “weight” to each training example, which determines the probability that
each example should appear in the training set.
 After training a classifier, AdaBoost increases the weight on the misclassified examples so
that these examples will make up a larger part of the next classifiers training set, and
hopefully the next classifier trained will perform better on them.
Classifier Output Weights:
 After each classifier is trained, the classifier’s weight is calculated based on its accuracy.
 A classifier with 50% accuracy is given a weight of zero
 A classifier with less than 50% accuracy is given negative weight.

28. AI-Based Clinical Decision Support Framework
Experimental Comparison – Learning-based System
AdaBoost
Formal Definition:
 The equation for the final classifier:
No. of weak classifiers
Output of weak classifier
‘t’ {-1 , +1}
Weight applied to
classifier ‘t’We make our final decision simply
by looking at the sign of this sum

29. AI-Based Clinical Decision Support Framework
Experimental Comparison – Learning-based System
AdaBoost
Formal Definition:
 Weight of classifier:
 The first classifier (t = 1) is trained with equal probability given to all training
examples.
 After it’s trained, we compute the output weight (alpha) for that classifier.
 error rate (e_t ) is just the number of misclassifications over the training set divided
by the training set size.

30. AI-Based Clinical Decision Support Framework
Experimental Comparison – Learning-based System
AdaBoost
Formal Definition:
 Updating examples’ weights:
 If the predicted and actual output agree, y * h(x) will always be +1 (1*1 or -1*-1)
 If they disagree, y * h(x) will be negative.
 Misclassifications by a classifier with a positive alpha will cause this training example to be given a
larger weight. And vice versa.
 If a weak classifier misclassifies an input, we don’t take that as seriously as a strong classifier’s
mistake.
Vector of weights
Training example
number
Sum of all the weights
(normalization)
Correct output
predicted output

31. AI-Based Clinical Decision Support Framework
Experimental Comparison – Learning-based System

32. AI-Based Clinical Decision Support Framework
Experimental Comparison – Knowledge-based System
Use EulerSharp semantic reasoner for the knowledge-based reasoning

33. AI-Based Clinical Decision Support Framework
Conclusion
This approach of Integrating machine learning with ontological
reasoning makes use of the inherent advantages of both approaches in
order to offer the required accuracy for the medical domain.
This approach supports interoperability between different health
information systems. A decision making process should use all relevant
data from many distributed systems instead of a single data source to
maximize its effectiveness.
This approach provides a framework that is generic enough to be used
in other medical applications.