Rutwik Shah

About Me

Hi there! Thanks for stopping by at this corner of the internet. As a physician scientist turned healthcare executive, I use my clinical and research expertise to build health A.I. products and manage industry partnerships at one of the top academic medical centers in the world. I have been fortunate to have a career that has spanned multiple roles which allowed me to understand the healthcare ecosystem in two very different settings, earlier in India and now in the US. Working at the intersection of medicine and machine learning, I love tackling complex problems by developing technology driven solutions that can have impact at scale.

While my interests span the spectrum of medical devices and digital health, I have developed considerable expertise in building A.I. solutions and deploying them in the clinic. My role allows me to work with cross functional teams involving clinicians, developers, data scientists and business managers and I absolutely love it! I feel excited to be at the leading edge of creating a new healthcare paradigm, where we unlock the potential of the Cambrian data explosion and create human-machine symbiosis to assist physicians, improve patient outcomes and address unmet clinical needs.

1. Clinical MLOps

The hard problem in A.I. is no longer model training and validation. Instead, the challenge now is in being able to deploy models such that they demonstrate explainability, generalizability, privacy, and integration within existing workflows. This is especially true for healthcare where the ethical and safety stakes are much higher. My team and I are designing solutions that take a systems approach through the entire A.I. life cycle of discovery, development, and deployment. We are building the next generation technology stack that aligns clinical data, tools, and platforms. Our framework enables model monitoring for drift and continual learning based on new data and physician feedback. As a clinical lead for data science teams, I utilize domain knowledge of clinical informatics, regulatory specifications, and user requirements to build solutions which are interpretable and usable in real world scenarios.

2.Multimodal Machine Learning

No physician makes a clinical decision by only looking at a single lab report, or an EKG or an x-ray or just based on your medication history. Yet, all our current clinical A.I. solutions are being trained narrowly on unimodal patient data which do not factor in the multiple considerations that go into a doctor’s decision. A key reason for this approach has been the lack of data interoperability and data siloes in healthcare. In collaboration with NVIDIA, my team is building the underlying platform which allow models to seamlessly retrieve patient data from multiple modalities such as imaging (PACS), clinical notes (EMR), and labs during inference and in real time. Assembling data like Lego blocks, our platform trains ML models to mimic real world physician decisions and generate holistic predictions based on a 360-degree view of the patient health.

3.Swarm Intelligence

Current A.I. algorithms are trained on "ground truth" labelled by clinical experts. But a significant challenge is overcoming differences of opinion observed among experts themselves, and agreeing on what the ground truth is in the first place. Such inter-reader variability cannot be solved simply by increasing the size of the training data or compute. To find a solution to this problem, I turned to the biological swarms of birds and insects for inspiration. Turns out nature has already devised a way to leverage the wisdom of the crowds to carry out complex tasks in real time. Working in collaboration with Unanimous AI, I conducted the first of its kind experiments in medical A.I., using a digital swarm platform modelled on behavior of bee swarms. Using the swarm platform, we had multiple radiologists collaborate in a blinded manner and in real time to give consensus opinions for lesions in the knee viewed MRI scans. Not only did the swarm opinions outperform the individual and majority vote decisions, but it also outperformed a state-of-the-art AI system built to detect knee degeneration on MRI imaging. Link to preprint.