AI for Medicine
Lecture 8:
AI Applications in Medicine – Part III
Feb 07, 2022
Mohammad Hammoud
Carnegie Mellon University in Qatar
Today…
• Last Monday’s Session:
• AI Applications in Medicine – Part II
• Today’s Session:
• AI Applications in Medicine – Part III
• Announcements:
• Assignment 2 will be out by tonight. It is due on Feb 15 by midnight
• Quiz I will be on Feb 21
Outline
Introduction
AI Applications
What is AI? Administrivia
in Medicine
Some Applications of AI in Medicine
1. Diagnosing diabetic eye disease using deep learning [1]
2. Detecting anemia from retinal fundus images using deep learning [2]
3. Predicting cardiovascular risk factors from retinal fundus photographs
using deep learning [3]
4. Extracting symptoms and their status from clinical conversations using
recurrent neural networks [4]
5. Performing differential diagnosis using probabilistic graphical models [5]
6. Predicting osteoarthritis using machine learning [6]
Digital Health Has Become Ubiquitous
• Everyday millions of people turn to the Internet for health information
and treatment advice
• In Australia, around 80% of people search the Internet for health
information, and nearly 40% seek guidance online for self-treatment
• In the US, almost two-thirds of adults search the Web for health
information and roughly one-third utilize it for self-diagnosis
Digital Health and Search Engines
• A recent study showed that half of the patients investigated their
symptoms on search engines before visiting emergency departments
• Search engines (e.g., Google and Bing) are exceptional tools for
educating people, but they may facilitate misdiagnosis!
• Some governments have even launched “Don’t Google It” advertising
campaigns to urge their residents to avoid assessing their health using
search engines
Symptom Checkers
• In contrary, AI-based symptom checkers are tools that can assist
patients in self-diagnosing themselves
• They are constantly and instantly available!
• Studies show that more than 15 million people use symptom checkers
per month that are likely to keep growing
• However, the utility and promise of symptom checkers cannot be
materialized if they do not prove to be accurate in self-diagnosis
Avey: An AI-based Algorithm for Self-Diagnosis
• Avey utilizes probabilistic graphical models in AI to provide patients
with immediate and accurate self-diagnoses
...
Avey’s Algorithm
Avey: An AI-based Brain in Action
for Self-Diagnosis
Avey: Experimentation Methodology
Stage 1: Vingette Creation Stage 2: Vignette Standardization Stage 3: Vignette Testing on Checkers
𝑅𝑣 𝑅𝑣 ……. 𝑅𝑣
D1 1 2 𝑛
𝑅𝑣 1
𝑅𝑣 ……. 𝑅𝑣
2 𝑛
D2 Test 𝑅𝑣 𝑅𝑣 ……. 𝑅𝑣
V1 V1 Counter = 3 Vignettes on
1 2 𝑛
Checkers 𝑅𝑣 1
𝑅𝑣 ……. 𝑅𝑣
2 𝑛
V2 D3 V2 Counter = 5 𝑅𝑣 1
𝑅𝑣 ……. 𝑅𝑣
2 𝑛
. 𝑅𝑣 𝑅𝑣 ……. 𝑅𝑣
.
.
1 2 𝑛
. D4 Pool of
Medical Compile .
Sources Vignettes . Gold-Standard
. Vignettes
. D5
.
. Stage 4: Vignette Testing on Doctors
.
D6 𝑴𝑫𝟏 𝑅𝑣 𝑅𝑣 ……. 𝑅𝑣
Vn Vn Test 1 2 𝑛
Counter = 5
Vignettes on 𝑴𝑫𝟐 𝑅𝑣 1
𝑅𝑣 ……. 𝑅𝑣
2 𝑛
Doctors
D7 𝑴𝑫𝟑 𝑅𝑣 𝑅𝑣 ……. 𝑅𝑣
1 2 𝑛
Avey versus Other Symptom Checkers
Avey Ada WebMD K Health Buoy Babylon
100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
M1 M3 M5 Recall Precision F1-Measure NDCG
Avey versus Human Doctors
Avey Average MD
100%
90%
80%
70%
60%
50%
40%
30%
20%
10%
0%
M1 M3 M5 Recall Precision F1-Measure NDCG
Some Applications of AI in Medicine
1. Diagnosing diabetic eye disease using deep learning [1]
2. Detecting anemia from retinal fundus images using deep learning [2]
3. Predicting cardiovascular risk factors from retinal fundus photographs
using deep learning [3]
4. Extracting symptoms and their status from clinical conversations using
recurrent neural networks [4]
5. Performing differential diagnosis using probabilistic graphical models [5]
6. Predicting osteoarthritis using machine learning [6]
And the Power of AI is Yet to Usher
• The first medical image ever taken was in 1895 by Wilhelm Röntgen
• This allowed looking inside a human body without slicing someone open!
• Since then, medical imaging has represented the fastest
growing source of medical data and literally changed the
face of medicine
• Imaging can sometimes reveal a disease before we feel it
• For several diseases (e.g., cancer), the earlier we diagnose them the more
likely the patients will survive, let alone reducing pain, suffering, cost, etc.
Discovery and Visualization of Structural
Biomarkers from MRI using Machine
Learning
• That is why we screen people who exhibit no signs or symptoms yet
• But, the earlier we capture images, the smaller & smaller the visible
evidences of diseases become, until they vanish before our naked eyes
• How early can we detect diseases?
• There is a growing belief that there is an invisible side to imaging!
• These are small changes (or hidden patterns) that are imperceptible to humans
• Yet, they can be detected by AI!
Discovery and Visualization of Structural
Biomarkers from MRI using Machine Learning
• One in 10 develop knee osteoarthritis, which cannot be detected until
the damage has occurred
Discovery and Visualization of Structural
Biomarkers from MRI using Machine
Learning
• Here are scans that belong to different subjects, with different colors
representing different ingredients that make up the cartilage
Group A Group B
Which group has osteoartherits? Best experts cannot tell!
Discovery and Visualization of Structural
Biomarkers from MRI using Machine
Learning
• Here are scans that belong to different subjects, with different colors
representing different ingredients that make up the cartilage
Group A Group B
NO osteoarthritis in 3 years Osteoarthritis in 3 years
Discovery and Visualization of Structural
Biomarkers from MRI using Machine
Learning
• This can be predicted using a machine learning algorithm that applies
a technique known as “transport based morphometry”
• The algorithm is able to predict whether a person will develop
osteoarthritis 3 year down the line with an accuracy of 86.2%
• The algorithm can learn more with experience, hence, it will only get
better and better as we feed it with more scans
Discovery and Visualization of Structural
Biomarkers from MRI using Machine
Learning
• But what does this algorithm see that doctors cannot?
• The diffusion of water
The pooling of water at the center of the
Water is evenly distributed throughout the cartilage cartilage suggests that there is a weakening
of cartilage at that location
Discovery and Visualization of Structural
Biomarkers from MRI using Machine
Learning
• This can augment our capability and help us identify new targets for
treatments
• Can we apply this technique to other diseases (e.g., Alzheimer, autism,
schizophrenia)?
• If so, can we halt them before they even begin?
• Just like in 1985 when the first image was taken, the face of medicine is
about to change again
• At that time we were able to see the visible
• Now, we can see the invisible through AI!
Next Monday’s Lecture…
• Machine Learning: Core Concept
References
1. Gulshan, Varun, et al. "Development and validation of a deep learning algorithm for
detection of diabetic retinopathy in retinal fundus photographs." Jama 316.22 (2016):
2402-2410.
2. Mitani, Akinori, et al. "Detection of anaemia from retinal fundus images via deep
learning." Nature Biomedical Engineering 4.1 (2020): 18-27.
3. Poplin, R., et al. "Predicting cardiovascular risk factors from retinal fundus photographs
using deep learning. arXiv 2017." arXiv preprint arXiv:1708.09843.
4. Du, Nan, et al. "Extracting symptoms and their status from clinical conversations." arXiv
preprint arXiv:1906.02239 (2019).
5. Hammoud, M., et al. “Avey: An Accurate AI Algorithm for Self-Diagnosis”.
6. Kundu, Shinjini, et al. "Discovery and visualization of structural biomarkers from MRI
using transport-based morphometry." NeuroImage 167 (2018): 256-275.
