In December of the previous year, as OpenAI’s ChatGPT grappled with finding practical applications, Google embarked on an ambitious journey to harness the power of large language models (LLMs) in the realm of healthcare. This endeavor culminated in the birth of Med-PaLM, an open-sourced, specialized LLM meticulously crafted for medical purposes.
Since its inception, this pioneering team unveiled a series of scaled-up healthcare LLMs, including the noteworthy Med-PaLM-2 and Med-PaLM-M. These models, designed with precision and purpose, have already begun to leave a profound impact on human lives. Notably, Med-PaLM-2 is currently undergoing rigorous testing at prestigious healthcare institutions like the Mayo Clinic.
One of the remarkable contributors driving these groundbreaking projects is Vivek Natarajan, an AI researcher at Google Health. Hailing from the vibrant cultural tapestry of Tamilian heritage interwoven with deep Bengali roots, Vivek’s professional journey has been an odyssey that spans from engineering internships at Qualcomm to roles with Meta AI, ultimately finding a fulfilling and transformative path within the dynamic field of medical AI.
Yet, beneath these remarkable achievements lies a captivating story of why Vivek Natarajan made the pivotal decision to transition into the world of medical AI.
Vivek Natarajan’s journey into the intersection of healthcare and AI was profoundly shaped by the healthcare landscape in India. As of 2023, India continues to grapple with substantial healthcare challenges, marked by inadequate medical infrastructure and a severe scarcity of medical professionals, especially in rural areas. The stark reality is that India boasts a doctor-to-patient ratio of only 0.7 doctors per 1,000 people, significantly below global standards. Moreover, with just 0.9 hospital beds per 1,000 population, and a mere 30% of those beds available in rural regions, the accessibility of healthcare is a pervasive issue.
Vivek’s upbringing in various regions of India exposed him to the formidable challenges that people in small towns and villages faced when seeking medical care. The arduous journeys of traveling long distances, often on foot and through extreme conditions, resulted in delayed diagnoses, poorly managed chronic conditions, and tragically, untimely deaths. These challenges cut across socioeconomic lines, underscoring the profound healthcare disparities that persisted in these areas.
The sheer injustice of people enduring such hardships for basic healthcare deeply troubled Vivek, and he felt an unrelenting drive to effect change. In 2013, he embarked on his mission by creating ‘Ask the Doctor, Anytime Anywhere,’ an app designed to democratize access to healthcare. The app’s very name encapsulates its purpose—to connect individuals with healthcare professionals whenever and wherever they need assistance.
Admittedly, ‘Ask the Doctor’ was constructed using older machine learning techniques, relying on a blend of expert systems and rule-based approaches. It soon became apparent that the app had limitations and did not deliver the desired outcomes, ultimately leading to its discontinuation. However, even during those formative years, Vivek possessed an unwavering conviction that AI held the potential to be the linchpin in resolving the healthcare challenges that people faced.
This intuitive belief ultimately guided him towards contributing to groundbreaking projects like Med-PaLM, where AI’s transformative capacity to revolutionize and democratize healthcare access became a tangible reality.
After completing a bachelor’s degree at NIT Trichy in Electronics Engineering and earning a master’s degree in Computer Science from UT Austin in 2015, Vivek began his career with Meta AI. This was his first job, and despite it being in the pre-transformer era, he gained a profound appreciation for the potential of deep learning. At Meta AI, he worked on various fronts, including speech recognition, conversational AI, and multimodal AI. His contributions extended to critical platforms like Newsfeed and Messenger.
However, life took a different turn during this period. Vivek’s father started showing signs of an aggressive form of Parkinson’s disease, which couldn’t have been identified sooner due to limited care options and resources. This personal experience reignited his commitment to a problem he had always cared deeply about—using AI to democratize access to healthcare and bring world-class medical expertise to billions.
Coincidentally, this was also when researchers from Google Brain and DeepMind, renowned for their groundbreaking medical AI papers, were forming Google Health AI, aligning perfectly with Vivek’s mission. So when Greg Corrado, co-founder of Google Brain and head of Google Health AI, offered him the chance to join, Vivek took it up without hesitation. Since then, he’s had the privilege of collaborating with esteemed AI researchers like Greg and Dr. Alan Karthikesalingam, working toward their vision of making an AI doctor accessible to billions.
The development of Med-PaLM and related projects was driven by a core concept—leveraging general-purpose language models like PaLM and GPT-4, which excel at predicting text but lack specialized medical knowledge. The challenge was clear: they needed to transform these models into medical experts. In essence, they aimed to send AI to “medical school” so that they could learn from high-quality medical domain information spanning human biology, the practice of medicine, and clinical expert demonstrations and feedback. This learning process is akin to a residency program for medical professionals after completing medical school.
However, they encountered primary obstacles along the way. The scarcity of large-scale medical datasets posed a significant challenge, primarily due to privacy concerns and the fact that healthcare in many global regions, especially in the global south, wasn’t fully digitalized. Additionally, a pressing concern was the presence of bias in language models used for healthcare. These biases—cultural, social, racial, and gender-based—could lead to unequal access to care, misdiagnoses, and treatment disparities. The root of this problem lay in these models relying on extensive datasets that mirrored historical healthcare inequities, potentially resulting in inaccurate diagnoses and treatment recommendations for marginalized communities.
To address these challenges, they fine-tuned the Med-PaLM models, derived from the general-purpose language models like PaLM, specifically for medical applications. This process involved utilizing high-quality medical datasets and clinical expert demonstrations. These datasets covered various critical areas, such as professional medical exams, PubMed research, and user-generated medical questions. Notably, they made use of openly available resources like the HealthSearchQA dataset from Google, which played a pivotal role in the development of Med-PaLM and similar projects.
In their efforts, they introduced an evaluation rubric for assessing large language models in medical applications, with a key focus on mitigating bias. Moreover, in Med-PaLM 2, they introduced adversarial questions evaluation, which specifically targeted sensitive topics like vaccine misinformation, COVID-19, obesity, mental health, and suicide. These topics had a high potential to exacerbate bias and healthcare disparities through the spread of medical misinformation. Their approach to mitigating bias involved rigorous evaluation and expert clinician demonstrations to train the model. While it’s undoubtedly a complex challenge, they are steadily making progress in this area.
The fine-tuning approach they employed varied based on the available data. For the first Med-PaLM, they utilized prompt tuning, where the majority of the large language model parameters remained fixed, and only a small set of additional parameters were learned. However, for subsequent versions like Med-PaLM 2 and Med-PaLM M, their team had access to more data, enabling them to fine-tune the models end-to-end. This approach significantly enhanced their performance and brought them closer in alignment with medical expertise.
As Vivek envisions the future, he sees immense promise in the role of large language models (LLMs) in the field of healthcare and beyond. These models, while not equivalent to human intelligence, offer exciting research opportunities. He sees vast potential in exploring LLMs’ applications in biology and neurology, including tasks like analyzing the human genome and decoding complex brain signals, opening up doors to a deeper understanding of these intricate domains.
While he has no immediate plans to revisit building a similar app like “Ask the Doctor,” Vivek firmly believes that his work on Med-PaLM and medical AI as a whole at Google will eventually lead to something akin to democratized healthcare access. They’ve made incredible progress with LLMs in the past year, and it seems that the dream of making an AI doctor accessible to billions is no longer confined to the realm of science fiction. Fingers crossed!