The responses have been given by Udayan Dasgupta, Co-founder and Chief Data Scientist, Tricog Health
All around the world, people are facing unprecedented challenges and uncertainties as a result of the COVID-19 pandemic. As InnovatioCuris (IC), we are always on a lookout for healthcare innovations that are affordable and provide quality care. In the wake of this need of the hour, InnoHEALTH magazine scouted and interviewed some innovative startups to build an army of health transformers to mobilize and address this global health crisis.
Kritika Arora and Varsha Prasad interviewed Udayan Dasgupta, Co-founder, and Chief Data Scientist, Tricog Health on behalf of InnoHEALTH magazine.
By interviewing Tricog Health, we aim to understand and review the role of AI as a decisive technology to analyze, prepare us for prevention, and fight with COVID-19 (Coronavirus) and other such pandemics.
Reasons behind the company name, any story behind it?
The founders wanted to develop medical technology that could be used to diagnose complicated conditions, like heart attacks, early and even predict their onset. They envisioned the product to be portable and easy to use and one that could be used by a non-specialist to get accurate medical diagnosis anywhere, anytime. Being huge fans of StarTrek, they imagined that this modern-day medical TRIcoder (https://en.wikipedia.org/wiki/Tricorder), of StarTrek fame, could be designed by leveraging Artificial Intelligence (AI) or COGnitive science. By combining these two words (TRIcoder + COGnitive), Tricog was born – with the mission of magically delivering an expert medical diagnosis to the real world by leveraging the power of AI.
What made you start the company? Tell us briefly about your journey?
The founder, Dr. Charit Bhograj, a renowned interventional cardiologist, routinely saw critical heart attack patients approach him hours after onset of symptoms. And like other cardiologists, he always knew that the clinical outcomes could have been significantly improved if the patients presented themselves earlier. And this could have been possible if the patient had done a simple diagnostic test, called an Electrocardiogram (ECG). This growing sense of frustration made him follow the patient’s journeys and realize that the root cause for the delay was often because the patient did not have access to a medical center that could conduct the ECG exam accurately, quickly, and affordably.
ECGs are the primary means of diagnosing serious cardiac conditions like heart attacks. While ECG equipment is fairly commonplace, reading ECGs requires the skills of a cardiologist or an experienced physician. Given that such expertise is limited in India (there are less than 3 cardiologists per million people in India) and in large portions of the developing world, misdiagnosis and delayed diagnosis are rampant. Furthermore, a delay of an hour in diagnosing heart attacks (golden hour) has been shown to significantly increase the risk of mortality and permanent damage to the heart. To make matters worse, cardiovascular disease is a growing problem since it is linked to factors like age, sedentary lifestyles, poor eating habits, and high levels of stress.
As machine embedded algorithms are only ~70% accurate in diagnosing heart attacks, most physicians prefer getting ECGs read by cardiologists, or, even worse, not keeping ECG machines as they cannot read it themselves, further exacerbating the scarcity of getting a quick and accurate diagnosis. This causes significant delays in diagnosing cardiovascular conditions and increases risks to the patient’s life. The worsening state of the problem and the importance of the golden hour make the search for a scalable solution, to provide fast, accessible, and affordable cardiac care to the masses, especially important.
To solve this precise issue, he teamed up with a group of engineers, Dr. Zainul Charbiwala (a hardware geek), Dr. Udayan Dasgupta (an algorithms expert) and Abhinav Gujjar (a software guru) to start Tricog, a company focussed on accelerating medical care and improving outcomes by enabling access to accurate cardiac diagnoses in an affordable and timely manner.
How is your company using AI? Tell us about the specialisations where it is being used and problems that you are solving?
The company believes in democratizing medicine by accurately diagnosing medical conditions in an affordable and timely manner. In this journey, they started off by introducing a medical service focussed on accelerating cardiac care by providing instant ECG diagnosis for diagnostic ECGs captured from remotely placed ECG machines – ECG being the primary diagnostic test for detecting heart attacks and other critical cardiac conditions.
This solution consists of cloud-connected ECG machines placed at various remote centers, which continuously push ECGs to the company’s Cloud where a proprietary algorithm called THA first analyzes the ECGs and provides the preliminary interpretation for verification by the in-house team of cardiac specialists who are present 24/7/365 company’s centralized ECG Analysis Hub. The physician-verified reports are sent back to the remote center via SMS/APP, who can then coordinate with neighborhood hospitals to provide further care.
They are currently in the process of using a similar approach to develop solutions for providing instant diagnosis for treadmill tests (TMT) and echocardiography tests.
Explain us a typical day in office, how does an AI expert spend their day? Tell us some under the hood details?
Developing algorithms to solve challenging real-world problems is an extremely enjoyable journey. It typically starts off by clearly defining the problem, in terms of possible inputs and their expected outputs, followed by extracting features from the data, annotating the data, and visualizing it in multiple ways to gain insights. This stage also involves working with domain experts (in this case, cardiologists) to understand how they diagnose ECGs. Depending on the amount of data available and armed with the insights gained from the previous data exploration stage, the AI engineer then starts evaluating various approaches to solve the problem – be it signal processing methods, machine learning approaches, or deep learning models. The data is typically split into training and test data sets so that the algorithms can be trained on the training data and their performance evaluated on the unseen test data sets. This stage can also involve reading and implementing published papers that have solved similar problems. The work is typically iterative in nature as approaches are tried, evaluated, and modified to improve performance on the test data sets. Once the engineer is satisfied with the performance, he may do further optimizations and architecture changes for efficient real-time deployment and a final round of validation with domain experts on large unseen data sets. When the domain experts and AI engineers are satisfied and the solutions are deployed, the AI engineer picks up the next unsolved problem and restarts the journey. Overall depending on the stage, every day in this journey is different as the engineer struggles with a series of puzzles and challenges along the path – but the sense of joy and achievement when each of them is solved is well worth the effort, says Mr. Dasgupta.
Tell us the challenges you have faced and are facing in development/implementation of AI?
Implementing effective AI in the area of healthcare, which can provide real value to physicians, has several challenges. A few of them are as follows:
- The promise of AI stems from the fact that it should improve with more data, but collecting such vast amounts of real-world data can be very difficult. People typically use a service (and share data) with an expectation of a result and especially in the initial phases when the AI is rudimentary, providing them with accurate results which would allow the service to scale (and lead them to have the confidence to share more data), becomes very difficult.
- Several medical diagnoses, including reading ECGs, is a mixture of art and science and can be subjective. Compounding this issue is the fact that ground truth can be hard to get and is often substituted with physician’s opinions that can vary across physicians. This can lead to large interobserver variability and getting universal reference data for training algorithms can be difficult.
- Cardiac abnormalities occurring in the broad population have a lot of variations (there are over 200 abnormalities that can be picked up by an ECG), and the data distribution is extremely skewed. This means that several conditions are very rare and the data collection process for such conditions is typically slower, making data-hungry AI algorithms struggle. And of course, the overriding concern and challenge are that given the nature of the task, the AI has a very small margin of error given the large implication of incorrect decisions.
How can you overcome some of these challenges? Do you want to share any instances where in the past your team was able to overcome any challenges?
The first challenge was devising a model that would continuously improve in performance over time. By using the hybrid model (Human + Machine Intelligence) they could roll out a service through which large amounts of data could be collected, and the algorithms improved. The human cognitive load reduced as the machine algorithms got better. According to Mr. Dasgupta, all throughout the process the end-customers always got accurate human-verified diagnoses, allowing people to believe in the diagnosis and the service to scale.
The second challenge, namely interobserver variability, was a very tricky challenge. Their approach to solve this issue was threefold:
- Standardized reading procedures and rules followed by doctors,
- Developing a proprietary algorithm approximating these rules which also provided evidence to the doctors to explain the algorithms’ conclusions. The viewing system that incorporated the algorithm results were carefully designed to promote usability while combating automation bias to keep the Human-Machine model accurate.
- Implementing rigorous quality control protocols coupled with continuous testing and training to help physicians adhere to the standardized reading procedures.
Implementing AI algorithms that could work well with smaller amounts of data and provided improved performance when larger amounts of data became available, involved designing various algorithms to solve the same problem, with varying dependency on data size. The first version of algorithms typically used little data and performance was improved by hand-tuning. As more data was available, other data-hungry algorithms, whose performance naturally improved with more data, were brought into the fold and combined with the earlier set of algorithms to improve the overall performance.
In summary, by building out this technology backed medical service, the start-up developed and deployed a business model which could deliver accurate results from the very beginning, but one whose efficiencies would continuously improve with more data. It was always apparent that a human-only approach would not scale and would be prohibitively expensive. By using the hybrid model (Human + Machine Intelligence) they rolled out a service through which large amounts of data could be collected, and the algorithms could be improved. Through this process the human cognitive load reduced and system efficiencies improved over time while maintaining high overall accuracy of the hybrid system at all times.
How reliable are these AI tools from a clinical perspective, tell us about the regulatory approvals, which stage are you in currently?
The start-up’s solution is currently deployed in over 3000 centers in both rural and urban locations and has diagnosed over 3.5 Million patients with 50% being abnormal and over 120000 of them being critical. They have recently implemented their first statewide STEMI program in Goa, and are working with multiple other state governments on similar programs. Outside of India, they also have rapidly scaled services in South East Asian countries (Philippines, Malaysia, Indonesia) and Africa (Kenya, Nigeria, Cameroon), shared Mr. Dasgupta.
The efficacy of the human-AI based platform has been key in creating this outsized impact. They are currently pursuing regulatory approvals for their solutions so that they can market these tools in other markets.
It was interesting to know that the algorithms have surpassed the performance of FDA approved algorithms in head-to-head testing on real-world resting ECGs containing over 200 abnormalities. Their reliability from a clinical perspective is evaluated continuously, out of ~150,000 ECGs they diagnose every month are verified by doctors. In such real-world tests it has been seen that doctors only modify about 1 out of every 5 diagnoses that the algorithm produces. The AI is expected to get even better as more data is collected, especially for rarer conditions. These efficiency gains are a testament to the fact that over the last 24 months, the ECG load has increased by ~10x while limiting the medical team growth to less than 50%.
Share the customer benefit with your solution, and the role AI will play?
Doctors, the customers, value the company’s services because of its accuracy and quick turnaround times, says Mr. Dasgupta. AI has been pivotal in meeting this expectation while keeping the service affordable. With the use of medical sensors becoming widespread, large amounts of data will be collected and it will be practically impossible for human-only approaches to analyze the data, derive insights, and develop treatment plans while keeping the process timely and affordable. AI will be increasingly adopted to help with these tasks. It will start with the simpler tasks of data analysis and will help physicians derive insights and then plan treatment. Over time, based on human interactions, AI will start taking on more complicated tasks, thereby increasing efficiencies of the healthcare systems. However in medicine, given the variety of complexities of the human body and the subjectivity of the diagnosis, human judgment will remain important – making physicians and AI work in tandem, especially for complicated cases.
Are there any general issues associated with AI products and services?
The main issue in the adoption of AI in healthcare is the concerns (often valid) about trusting technology for taking critical decisions, especially given that the impact of an incorrect decision is large. This lack of trust is not helped by the fact that some of the recent AI techniques, like deep learning, are typically opaque, i.e. although they may deliver the correct diagnosis, they are unable to explain the results for the physician to understand its reasoning. And although some methods work very well when the unseen data is similar to past data, they may struggle to generalize learnings over data from disparate sources.
What efforts are you making to overcome associated disadvantages?
The start-up has never used AI to independently diagnose critical heart conditions and has rather used it as an assist tool for their in-house physicians. Also, throughout this journey, they have focussed on two aspects:
- Always providing visual evidence to the physicians which justifies the algorithms’ decisions
- Always learning from the physicians’ reviews, especially when they differ from the algorithm’s conclusions so that over time the algorithms improve and better mimic human decision making. With these guiding principles and constantly working with the physicians, the end-users, the algorithm has been able to gain their trust for the majority of diagnosis and also their patience as they work with the algorithm designers on the remainder of the conditions where the algorithm continues to improve.
What differentiates you from your competitors?
The virtuous Human <=> AI cycle embedded in Tricog’s business model has been key in building a robust algorithm platform tested on real-world data which has provided increasing value to the business over time. Their algorithms started off by leveraging a few public domain databases and using signal processing based approaches for feature detection and measurements, coupled with an expert system built in consultation with the in house medical team. However as their ECG diagnosis service became popular, it helped in collecting large volumes of high fidelity real-world data which were annotated in real-time by the in-house medical team.
Another key advantage is that the solution was always architected to be cloud-first, which is in sharp contrast to even high-end traditional ECG machines with in-built diagnosis algorithms leveraging sophisticated, but fixed, signal processing techniques. By moving the data and computations to the cloud, a few interesting benefits occurred,
- The computational resource constraint was removed
- The algorithms could remain dynamic and change over time
- The aggregation of data allowed machine learning and deep learning approaches to be incorporated within the product.
- The elaborate ECG viewing system allowed a lot more supplemental information to be shown to the cardiac specialists, leading to a quicker and more accurate diagnosis.
Being the first physician verified instant ECG diagnosis service, they seem to have effectively identified and effectively solved a deep problem in the healthcare landscape.
Where do you see your organisation in 10 years? Any brief message for our readers?
The start-up intends to provide technology accelerated solutions to real-world medical problems. It has started off by making cardiac care accessible, affordable, and timely. Developing solutions for the Indian market has successfully led them to take its solutions to countries in Southeast Asia and Africa. Going forward they intend to grow these services in other geographies while developing differentiated services for delivering quality healthcare at home.
Interviewed by Kritika Arora and Varsha Prasad
