Tathagato Rai Dastidar, who goes by the moniker “Tatha”, is a tech industry veteran of 21+ years. He is a BTech and PhD in Computer Science from Indian Institute of Technology, Kharagpur. He started his career back in the year 2000 at National Semiconductor, building software for semiconductor design automation. He had a brief entrepreneurial stint in the semiconductor design automation domain in 2007-08. Over the years, he has held technical leadership positions in premier technology companies like Yahoo!, American Express and GraceNote.
It’s a wild wild west …
The ongoing pandemic (I purposely use the word “ongoing” rather than “recent” – this thing is far from over) has exposed a whole lot of holes in the healthcare system. Not just near home, but worldwide. The healthcare infrastructure has been pushed to its very limits. Not surprisingly, it has buckled under the pressure. Lives have been lost to lack of timely diagnosis, to lack of treatment facilities. I am going to explore some of the gaping holes in our healthcare system today and explore ideas around how and why AI can make a difference.
The timely diagnosis of a disease, or quantifying its progress, is just as important as the treatment itself. In fact, it guides the treatment. In this article, I am going to restrict myself to the diagnosis and quantifying part only.
Getting to the roots of the problem …
Early diagnosis of diseases, while an essential element in a well-functioning healthcare system, is difficult to achieve today for large swathes of the population. This is due to multiple factors.
First and foremost is the lack of qualified medical professionals. In India, the doctor-to-patient ratio is around 1:1700, which is lower than the recommended value. But the number hides the fact that the majority of doctors (close to 90% per some estimates) are concentrated in urban centres, while more than 65% of India’s population lives in villages. Thus, the actual doctor-to-patient ratio in large parts of the country is far worse than what the number above suggests. The hope of access to quality healthcare near home is nipped in the bud for many of us.
Second, most medical professionals are overworked and stretched to their limits. This has been showcased time and again during the pandemic. Overwork leads to variability, inaccuracy and inefficiency in decision making. Doctors are human too, just like the rest of us.
Third, quality medical diagnostic equipment is still expensive. While the availability of (relatively) lower cost point-of-care diagnostic devices has sharply increased during the pandemic, the majority of equipment is still optimised for central laboratories handling high sample volumes. Thus, they remain expensive. Setting up quality diagnostic centres in areas with low population density, remains economically difficult to achieve. Yet, if the pandemic has taught us one lesson, it is the need for healthcare to be decentralised.
How can AI help?
One thing that is certain in this puzzle is that we cannot create enough doctors to bridge the existing shortage overnight. The only way to bridge the gap is to help the doctors be more efficient and expand their reach to a larger section of the population. This is where AI can play a pivotal role.
Let’s explore a few areas where technology, and in particular AI, can make a difference. Outreach of doctors can be expanded manyfold through telemedicine. But simply expanding the reach, without bringing in efficiency, will only lead to more overwork for the doctors. This is where AI can help – in automating the routine tasks that they perform.
One common area in diagnostics is the examination of visual medical data to detect abnormalities. Visual medical data consists of things like ECG graphs, X-rays, MRIs, glass slides with tissue on it for examination under the microscope, and so on. Here the examining doctor looks for abnormal patterns in the visual evidence. As we all know, it is easy to miss or overlook things, especially subtle ones, when one is overworked or stressed. Can AI make a difference? The answer, it turns out, is a resounding ‘Yes’. AI can automate the analysis of visual data to detect patterns and point out abnormal ones to the doctor. It cannot, and in my opinion, should not, aim to replace the doctor, but simply to aid and assist her in the diagnosis process.
How far has AI progressed in this? Quite a lot, actually. We will go through a few examples.
ECG is an extremely useful tool in detecting heart ailments, proven and perfected over decades. More than 300 million ECGs are recorded annually. Yet, till a few years ago, almost all automated algorithms for detecting abnormalities in ECG fell woefully short. There were too many false alarms, rendering the algorithm practically useless. But today, with deep learning and AI, this has been made possible. A group in Stanford was able to diagnose arrhythmia (a heart problem) with cardiologist level accuracy, using signals from a single lead wearable ECG monitor, with the help of a deep neural network. Many wearable devices today incorporate some form of single lead ECG monitoring and analysis. Even for the conventional 12-lead ECG, an AI based detection algorithm for atrial fibrillation received the USFDA clearance.
Another common family of diagnostic procedures is medical imaging or radiology. X-rays, CT scans, MRI scans – all are part of this family. Traditionally, radiologists have examined these images. With the advent of digital imaging (no one takes X-rays on a film these days!), teleradiology has become extremely popular. Radiologists located in one part of the world can examine scans from all over the world. Teleradiology centres are a common thing today. AI has made inroads here too. Several AI algorithms have come up which address different types of medical images. Many start-ups building AI tools for analysis of radiology images have grown in India too.
In the world of microscopy, too, AI has made great inroads. AI based tools for examination of biopsies have received regulatory approval in multiple countries including the US. Digital haematology (analysis of blood) is also a hugely popular field. World over, many start-ups are operating in this area, though the number is definitely less than those in the radiology world.
In all these areas, there are documented studies where the AI is shown to be superior to doctors of average skill when it comes to detecting abnormalities. This is no mean feat!
Given the progress in the last few years, the day we see a digital AI assistant on every clinician’s desk may not be too far away!
OK, so what’s the catch?
Building AI for medical problems comes with a whole lot of challenges, unique to this field, which AI for other applications may not face.
The first and foremost is the availability of quality data. ECG and radiology have been digitised over several decades. Loads of quality data are publicly available. But in the world of microscopy, the picture isn’t as rosy. Microscopy, till today, is largely a manual job. Digital microscopes are few and far between. Thus, digital microscopic images of cancerous tissues or blood cells are not available in large quantities. Any start-up in this domain needs to solve the data availability problem first, even before commencing on the AI journey.
Next comes the annotation. When we get an MRI scan, we can sometimes get the radiologists’ report as well. But how do we relate the observations made by the radiologists to specific areas of interest in the scan? How do we demarcate the boundaries of a tumour from the report? In case of microscopic images, no annotated report is usually available, apart from the high-level diagnosis, which could relate to some parts of the tissue only. Thus, in most cases, additional annotation is required. Annotation of medical data is expensive and time consuming. Medical professionals are overworked as it is. Can they afford to spend more time in annotation of data in the hope of someday being able to avail the benefits of AI? How do they handle their current caseloads, then? Unlike many other fields of application of AI, laypeople simply cannot annotate medical data. The variability among doctors on the same data is also substantial in many cases, leading to confusion as to what the ground truth really is!
Third problem is the concern of privacy. Are hospitals allowed to share medical data, even anonymised? Was explicit patient consent obtained for the data being shared? If data can’t be shared without consent, do we need to only collect data prospectively, like in a clinical trial? How can the huge volumes of retrospective data be used for building AI? These are all tough ethical questions with no easy answers. Any start-up operating in this field thus has to carefully tread through these questions.
The next challenge in my list is that of regulatory approvals. How should AI tools be positioned? Are they mere assistants to a human, or are some AI aiming to replace the human (or at least make the human necessary for only the final sign off)? If the AI is giving out a diagnosis, who is responsible if it goes wrong? Even if it is not diagnosing on its own, and merely suggesting a diagnosis to a doctor, then too the question of liability remains. We all know the power of suggestion. What if the doctor makes a mistake based on the erroneous suggestion of the AI? All these questions are made more complicated due to the fact that human lives (and perhaps livelihoods too) are at stake here. Like the questions on ethics, these too are tough questions with no easy answers. But answers to these are what we should be looking for right now, because, as we saw earlier, the day where AI exceeds average human efficacy for routine tasks is already here.