BME 6407: Machine Learning in

Healthcare

Dr. Taufiq Hasan

Director, mHealth Research Group
Associate Professor, Department of Biomedical Engineering (BME)
Bangladesh University of Engineering and Technology (BUET)
Former Research Scientist, Bosch R&D, Sunnyvale, CA (USA)
The Learning Problem

Female
Male

Voice pitch (Hz)

Height (cm)
 patient have
CVD or not?
Does the

l
ve
r le for r
ga d D a
su Bloo CV ug
of d s
sk oo
Ri h bl
g
Hi
Predicting Cardiovascular Disease (CVD)

for le
le D
sty ary CV eop
life ent of ry p
sk
Se
d Ri enta
d
Se for
eti
c D ts
CV ien
ab
Di of pat
sk c
Ri eti
ab
Di for
r D
ke CV s
of ker
o
Sm sk o
Ri Sm
for
D ol
el ol CV ter
lev ster of les
sk ho
ole Ri h C
g
Ch Hi

for CVD risk

Importance
Patient’s
data
Formulation
❖ Input: (patients clinical information)

❖ Output: (patient has CVD or is healthy)

❖ Target function: (ideal disease prediction formula)

❖ Data: (historical patient records)

Learning

❖ Hypothesis: (learned formula by ML algorithm)

❖ The hypothesis set: , (set of candidate formulas)

Flow Diagram of a Machine Learning Problem

Ideal function to predict cardiac disease

Historical data from patients

Final disease prediction formula

Set of candidate formulas

The Perceptron - A simple learning model
Perceptron Learning
Feasibility of Learning
Outside the training data
Probabilistic Interpretation
BME 6407: Machine Learning in
Healthcare

Dr. Taufiq Hasan

Director, mHealth Research Group
Associate Professor, Department of Biomedical Engineering (BME)
Bangladesh University of Engineering and Technology (BUET)
Former Research Scientist, Bosch R&D, Sunnyvale, CA (USA)
Error measures (Binary Classification)
Example: Consider the problem of detecting brain hemorrhage (internal bleeding) from CT
scan images. Assume that our target function predicts if a brain hemorrhage is detected or
not.

Hemorrhage detected

Healthy
False reject: The patient had hemorrhage but ML could not detect its presence.

False accept: The patient did not have hemorrhage but ML detected its presence.
Report Ranking Application Emergency Room Application
Review images by radiologist earlier for Perform surgery immediately if hemorrhage is
reporting in case hemorrhage is detected. detected. Do not wait for radiologist availability.
Confusion matrix
Error/performance measures
f = +1
F1 score is define as the harmonic mean of
precision and recall:

h = +1
Sensitivity vs Specificity

COVID RT PCR test: sensitivity and specificity were 72.1% and 98.7%,
Equal Error Rate (EER)

❖ In many cases, h(x) provides a “score” (or probability) instead of +1/-1 decision.
❖ In this scenario, a threshold is selected to make +1/-1 decision.
ROC (Receiver operating characteristic) curve