AI in Diagnostics

By Sylvia He
Introduction

While the drug market may be more lucrative than the diagnostic market, the diagnostic
market has an equal or more significant impact. For example, many more people have
received a COVID test than getting medication or in-hospital treatment.

In many countries, diagnostic errors affect patient care quality and cause complications
and even deaths. In addition, diagnostic errors often lead to financial compensation to
the affected patients, which further adds to the rising cost of healthcare.

In addition to rising costs and staff shortages, the volume of medical images is
outpacing the availability of specialists who can process them, especially in low- and
middle-income countries. The healthcare industry is looking to reduce costs via
automating repetitive processes and enabling doctors to make faster, more accurate
decisions. As a result, when private companies and public institutions began exploring
AI in drug discovery and development, they also tried to apply AI to clinical diagnostics.

Watson the Pioneer

IBM's AI platform, Watson, was the first platform to make a splash in the diagnostic
world. During the 2010s, IBMS had formed more than 50 partnerships with prominent
institutions such as the US Department of Veteran Affairs, Memorial Sloan Kettering,
Johnson & Johnson, and Medtronic to develop AI diagnostic tools to analyze medical
images, patient records, or genetic data.

Image analysis was viewed as a low-hanging fruit since the researchers can train the AI
algorithm with the existing massive image datasets. However, hospitals in India, Korea,
and the US have found that while Watson could analyze medical images, its
performance and accuracy were uneven with different cancers. Moreover, AI in its
current form does not excel in solving complex, nuanced problems with incomplete or
disorganized data, such as finding useful information in the research literature or patient
records.

There was some good news, however. AI was found to excel in more straightforward
tasks, such as analyzing genetic data to reach a binary conclusion or performing
repetitive tasks, such as obtaining basic patient information.

A recent paper by researchers at University Hospitals Birmingham NHS has reviewed

more than 20,000 research papers on AI's application in medicine. The study found that
high-quality deep learning algorithms could detect various diseases as accurately as
health professionals. On the other hand, a pathologist can make equally good or even
better diagnostic decisions than AI with unlimited time.

Since giving pathologists unlimited time to analyze samples is unrealistic, many

researchers have proposed a hybrid model, where the AI system assists doctors in
making effective diagnoses.

By 2019, the US FDA has approved more than 30 AI algorithms for healthcare; the
applications range from identifying bone fractures in images and detecting diabetes to
picking up signs of stroke in CT scans. Potentially, the hybrid model may broaden the
application of AI to more diseases and diagnostic functions.

Disease Detection

In the tradition of Watson, AI can be used to detect various diseases. For example, the
National Institute of Health in the US has launched an initiative that promotes the use of
AI to analyze CT lung scans of COVID-19 patients and improve treatment.

Medical imaging is also often used to screen for cancer. For example, determining if a
case of breast cancer is malignant or benign is challenging. Misdiagnosis can lead to
unnecessary procedures for benign breast cancer patients and missed treatment for
malignant patients. AI can help improve the accuracy of image analysis and reduce the
number of false positives and false negatives. For example, QuantX is the first FDA-
approved AI platform to evaluate breast abnormalities. QuantX quantifies the
abnormalities into a single score that helps radiologists reach a diagnosis. AI is also
being used to classify different types of skin cancer. Lastly, the Augmented Reality
Microscope integrates AI into routine workflows and helps increase the accuracy of
detecting metastatic breast cancer and identifying prostate cancer.

In addition, AI can be used to analyze chest X-rays and identify heart abnormalities.
Also, AI can help automate tasks such as measuring the aortic valve, carina angle, and
pulmonary artery diameter. Several related devices have obtained regulatory approval
from the US FDA to go to the market. Apple Watch was the first consumer-available
product that allowed users to take an ECG from their wrist and send it to physicians.
Zebra's Medical Vision platform quantifies the calcium deposit level, which clogs up
blood vessels and may cause a stroke or blood clotting in a patient's coronary artery.
Lastly, Bay Labs' EchoMD AutoEF platform analyzes a patient's echocardiogram and
assesses the level of blood flow and the risk of heart failure.

Bone fractures can be hard to detect on standard images. AI may pick up subtle
variations in the image that indicate a fracture that requires surgery. This way, AI can
help lower the number of false-negative patients and reduce the pain they have to
endure.

Neurological diseases, such as amyotrophic lateral sclerosis (ALS) and primary lateral
sclerosis (PLS), often have similar symptoms that are difficult to distinguish. False-
positive results are common. AI can help distinguish between ALS and PLS via image
analysis.

Lastly, Aidoc, an FDA-approved app, allows radiologists to identify acute brain

hemorrhages, a life-threatening condition, in head CT scans.

Disease Prediction

AI can also help doctors predict the likelihood of a patient developing certain conditions
or screen for the patients who are likely to have a disease.

A Swedish University has developed an algorithm that can identify patients at risk of
developing septic shock, a life-threatening and often unpredictable condition. Birth
asphyxia is hard to predict or prevent. American and Brazilian researchers have
developed a fuzzy logic system that analyzes a mother and the baby's data to predict
the likelihood of the baby needing resuscitation so that the doctors can be more
prepared. Researchers are also developing an algorithm that predicts patients' risk
of heart attack, breast cancer, or even suicide.

Improving the Diagnostic Process

Because AI excels in performing repetitive tasks accurately, it can play a significant role
in reducing personnel fatigue and documentation errors. Furthermore, in the current
climate of COVID-19, where face-to-face interaction between healthcare workers and
patients may increase the chance of infection, AI may serve as a virtual protective gear
for the medical staff.

For example, chatbots, which are AI robots with speech recognition capability, can
gather information from patients, identify patterns in their symptoms, and make
diagnoses and treatment recommendations to doctors. When the symptoms point to
common flu, the chatbot can recommend over-the-counter medication; if the symptoms
are more serious, the chatbot can direct the patient to a consultation with the doctor.
Babylon Health, a startup in the UK, has developed such a chatbot.

There are also remote patient monitoring programs. Ada Health, a Berlin-based
company, has developed an AI platform to track patient health and offer
recommendations based on patient symptoms and other health information. On the
other hand, a voice-based virtual nurse program checks on patients between office
visits and provides automatic alerts to physicians. This way, virtual nurses can reduce
patient anxiety, increase compliance with medication regimens, and maintain patient
satisfaction.

In addition, AI can synergize with personalized medicine to achieve optimal patient

treatment.

Due to their different genetic makeup, patients respond to drugs and treatment doses
differently. The same dose of a drug will be less effective but safer in patients who
metabolize (or break down) the drugs quickly, while it will be more effective but
potentially more toxic in patients who metabolize slowly. Therefore, customizing the
dosage of a drug for a patient will improve his response to the treatment.

Biomarkers are indicator molecules found in fluids, usually blood, in the body. They can
help doctors ascertain if a patient has a disease, is at risk of developing a disease, or
will respond to a drug. Biomarkers can make disease diagnosis more secure and
cheaper, but they are too numerous for humans to consider together. As a result,
genetic testing usually focuses on one or a few genes. However, with AI, doctors can
consider more genetic variation of a patient and make more accurate decisions on
which drug to use and how much.

Conclusion

The diagnostic field is more risk-averse than drug discovery because diagnostics have
direct application to patients. Therefore, the adoption of AI in diagnostics has faced and
will continue to face heavy skepticism and concern, as illustrated by the relatively low
success rate of AI diagnostic platforms gaining regulatory approval since 2018.

One way to alleviate such concern will be to define AI as an assistant role instead of the
leading role in diagnostic tests. In addition, training the AI model for diagnostics is even
more important than drug discovery since diagnostics are directly relevant to patients.
Therefore, the sharing of training data will be critical.