AI based Medical Diagnostics - what I did before advertising....

More than 20 years ago, in 1998, I was a Medical Entrepreneur who was fascinated by Medical AI technology. This was long before the advent of AI/ML, TensorFlow, Generative Adversarial Networks, or any the similar terms bandied about today. This was the Jurassic Era, where Intelligence did not reside in the cloud but individual computers, with intelligence locked into their hard disks... I was a pioneer of one of those early technologies, and here is my story....

Coronary Artery Disease (CAD) - a big problem

I had spent the last 18 months in Cardiac Catherization labs at hospitals across India. The Cath-Labs were part of the cardiac care facilities, and were teeming with heart patients who had come from far far corners of India seeking diagnostic treatment. With worried family in tow, many of these sick, frightened patients had travelled from the smaller cities, in search of a cure. For many of the patients a heart attack was the first symptom. For the unlucky few who could not even make it to the hospital. death by heart attack was the first and last symptom.

for some unlucky few a fatal heart attack was the first and last symptom..

CAD : A Silent Killer

CAD is a disease of the arteries supplying the heart. The arteries, which start out smooth and elastic, get plaque on their inner walls, which can make them more rigid and narrowed. This restricts blood flow to the heart muscle, which can then become starved of oxygen. The plaque could rupture, leading to a blockage of blood flow to the heart, resulting in a heart attack. In some cases the heart attack is so powerful that the patient dies quickly, within one hour of the onset of the attack, in a syndrome called Sudden cardiac Death, or SCD.

In general CAD does not manifest itself early. It can remain lurking and undetected, and strike when least expected. The case of Dr. James Fixx, the guru of jogging is an example of SCD; an avid runner, he died suddenly at the peak of his success at 52, while jogging. A post-mortem suggested that CAD had blocked his arteries and triggered a massive heart attack, or what is called a "fulminant" (explosive) heart attack. He led a healthy and fit lifestyle.

He had no symptoms...

CAD: 30% of all deaths worldwide, growing in India...

CAD is a major cause of mortality and morbidity all over the world. According to a report of World Health Organization (WHO) in 2005, cardiovascular disease (CVD) caused 17.5 million (30%) of the 58 million deaths that occurred worldwide. India is one of the reasons for alarm, with an alarming increase over the past several decades in the prevalence of CHD and cardiovascular mortality, with one study claiming a 4-fold rise of CAD prevalence in over the past 40 years.

The Problem: Lack of a reliable, easy to use, non-stress, non-invasive way of "ruling-out" CAD...

CAD is well spread among the large Indian population, cutting across geography and Socio economic classes. Many patients came to the big cities from rural or semi-urban areas, looking for hospitals with advanced diagnostic tests, the gold standard being Coronary Angiography. An coronary angiography is a highly invasive photograph of the coronary vasculature; a Radio-opaque die is injected into the coronary vasculature and photographs of the vasculature are taken by an X-Ray machine. In some cases it can result in adverse conditions such as allergic reaction to the radio-opaque die. The X-Rays fry your body with X-Rays; an entire year's exposure can be exceeded in a few seconds of photographs.

While Angiography was was the gold standard for detection of CAD, it clearly could not be a front-line, or screening protocol. What if we could have a test which was (1) simple to administer, (2) non-invasive (3) did not require the patient to exercise while the test was being conducted and (4) did not require any diagnostic expertise or sophistication in assessing the test result ?

This test could act as a early warning screening test, which would be deployed at all areas across the country, from urban to rural. Administered by a trained paramedic at thousands of primary care centers across the country, the tests could be applied at scale to large swathes of the population. It would be a highly accurate in its diagnostic of the presence of CAD or other heart disease. It would have false positives, but very low false negatives.

The crushing load on Cardiologists, confronted with the costs incurred by the patients could be solved at scale; only the few who needed extra tests would be channeled into the diagnostic system, with a large majority of healthy patients saving unnecessary costs of doctor visits and diagnostics.

Enter CUPID

At that time, we heard about this technology that was branded CUPID. The company was based in New York, and had just what was needed; a 12 -Lead resting ECG that promised very high diagnostic accuracy. Based out of technology created by 2 Chinese Scientists at the university of Beijing in the 60's, the technology featured a "brain" that was trained by mapping the ECGs and the Coronary Angiograms of tens of thousands of cardiac patients.

Spectral Analysis of ECG waveforms

An ECG signal is a time-domain signal , which maps amplitude of the ECG waveform over time. This needs to be visually inspected by the doctor recording the ECG and then diagnosis made on the basis of the shape of the ECG. Looking at a Time-domain signal is a very low-resolution way of looking at a waveform. Spectral Analysis of time domain signals is an established practice of adding a huge amount of fidelity to hitherto noisy time domain signals. Basically, mathematical transformations are applied, including Laplace and Fourier Transformations, that allow for very nuanced components of the signal to be detected; Consider an oceanographer who was studying the oceans in Australia circa 1915. Applying spectral analysis to the waves , he could make the following statement:

"Using spectral analysis we are able to see the presence of a very subtle oceanographic signal with an amplitude of 2 mm and a wavelength of 2 Km....that is most likely emanating from over 5,000 KM away, most likely from the Bay of Bengal..."

From a diagnostic perspective, the huge profusion of mathematical by-products that arise from these transformations are a blessing as well as a curse. CUPID claimed to have solved this problem, by building a brain where the mathematical by-products of a resting 12-lead ECG were mathematically correlated to the coronary angiograms of thousands of patients, resulting in an AI based brain that was capable of high diagnostic accuracy. The device was being used in Brazil with significant claimed success, and had an interesting peer-reviewed paper or two on the science of spectral analysis as applied to cardiac diagnostics.

This was the future of Cardiac Screening and Care. I would change the world.

Mustering all the money I had made from the coronary stent business, which was the princely sum of 5 Lac rupees, I decided to launch the technology in India, while simultaneously executing a study which would take patients, study their CUPID scores pre-angiography and actually map the results that were shown by the Angiogram.

Dr.Suresh Vijan was kind enough to give me a stall at one of the largest Cardiology Conventions in Mumbai at a cheap price. We invited the CEO of CUPID, who would be present and we would actually test out the technology and give diagnostic assessments of the patient on the spot. Several of the people we tested were doctors themselves, from smaller towns,. and impressed with the potential of such a technology. Everything seemed warm and positive.

The next task was to conduct an actual "Shootout" test, data from which could be used to validate the technology. Dr Brian Pinto, a leading interventional cardiologist agreed to do this as the Nanavati Hospital Cath-Lab. The cardiologists at the hospital were intrigued and we actually created a marketing event at the hospital.

CUPID failed.

Test after test, the diagnostic output of Cupid was measured. Initially it seemed to work, with us getting relatively correct results. However as the day went by we were checking more and more patients, CUPID was struggling to match the Angiographic diagnosis. It also seemed to require a large amount of interpretation, which was being done by the CEO of CUPID.

Dangerously, it was beginning to give False Negatives, which is the kiss of death for any diagnostic modality. The hypothesis of the test is that the patient has a cardiac problem. A false negative is a failure to detect a problem, something which is far more costly than a false positive - which just requires more tests and time.

It was then that I threw in the towel. The technology was promising and seemed rigorous, but was not living up to its claim. It would require a huge amount of incremental research, development and investment in marketing adoption for it to work....

My learnings from the journey

1. Anything which sounds too good to be true is probably just that. Be judicious in examining these more critically rather than believing what you want to hear.
2. Do more due-diligence on the datasets that go into building such a brain. Yes the data fed into the brain was large, but was based on Chinese patients, whose vasculature, diet and other habits are significantly different from Indians.
3. Even if such a technology is found, it is possible that there are vested interests that crimp the adoption of such a technology. Solving for chaos and waste can eliminate many of the revenues generated from an in-efficient ecosystem.

2019 - AI based diagnostics - a booming space

Cut to today. AI has come of age, with a huge ability to build sophisticated brains capable of very complex inferential and predictive power. I continue to believe in the power of using Spectral Analysis to build a spectacularly powerful screening test which would be cheap, non-invasive, non-stress and administerable by anyone with a Smartphone ! The Math is robust, the computation power is available on tap on the cloud - what is missing is the data. If we have clean data from Cath-Labs coming in, along with nominal patient data (family history, smoker, sex etc ), it is possible to build and deploy this brain.

The technology, would generate in its wake, a large market for second opinions. if a patient is recommended for further tests with some alerts by the brain, the patient can choose to send the test data to any other doctor who would then charge a fee for a second opinion, based on the same data. Cardiologists could make money sitting off a beach instead of expensive clinics. Medicare can be Uberized !