On May 7, the Digital Data Design Institute at Harvard hosted Leading with AI: Exploring Business and Technology Frontiers. The conference featured a panel moderated by Robert Huckman, Albert J. Weatherhead III Professor of Business Administration at Harvard Business School, and four panelists: Francesca Dominici, Clarence James Gamble Professor of Biostatistics, Population, and Data Science at Harvard T.H. Chan School of Public Health; Nicolò Fusi, General Manager of Microsoft Research; Sham Kakade, Co-Director of the Harvard University Kempner Institute; and Armen Mkrtchyan, Origination Partner at Flagship Pioneering. The panel discussed possible short- and long-term implications of AI in healthcare and human health in general.
Each panelist began by answering Professor Huckman’s question, “In the next 1 to 2 years, what are the most interesting applications of AI that could potentially impact not just healthcare but human health?” From there, a fascinating discussion touched on many important issues related to AI and healthcare.
Key Insights
Insight
Francesca Dominici on studying the intersection of generative AI, climate and health.
“The reality is that we know that probably tomorrow we’re going to start having these massive heat waves or the next wildfires, and what can we do at the local level to prevent the largest number of deaths?”
Through her research, Professor Dominici is seeking to use AI to analyze and learn from large amounts of data that can drive the development of local climate adaptation strategies to prevent hospitalizations and deaths due to extreme weather events.
Insight
Nicolò Fusi on integrating information in drug discovery related to particular cancers and better matching therapies with patients:
“[I hope to see] a lot more advancement on rare diseases and rare cancers also, which are not rare collectively. If you take them all together, they actually make a good chunk of all cancer patients, but the possibility and potential of actually integrating information in context and reasoning about particular cancer in context is extremely powerful…”
Fusi pointed out that once large generative models are pre-trained, they are efficient in context. Combining enormous amounts of data and the increased efficiency of the models, in the short term, he hopes to see advances in drug discovery for rare diseases and rare cancers and a focus on delivering the right therapies to the right patients.
Insight
Sham Kakade on leveraging AI capabilities to summarize web searches and patient information to put them in relevant context:
“For an average case, either say something valuable to me as a patient or valuable to me as a doctor, I’d love it if it could just drop my health history into the context drop…they could do a web search, pull things into the context, and give me a summary…This could be a huge win for helping doctors to remember just all the conversations they’ve had with a person…”
In healthcare settings, Professor Kakade points to the power of AI tools that can search and pull information into context (for example, from medical publications) and provide a summary of medical history, conversations, and research for doctors and patients.
Insight
Armen Mkrtchyan on the entrepreneurial opportunity to enable preemptive health and medicine:
“…think of it as AI coming up with a hypothesis but then figuring out how to validate or invalidate those hypotheses…we have spent a ton of effort [on] preemptive health and medicine, how much we can go upstream of the disease, try to figure out the biomarkers, try to figure out what are the interventions and create them.”
In the short term, Mkrtchyan feels that AI can improve efficiency and efficacy to help “close the loop scientifically.” AI can integrate data, knowledge, and experiments to build and test hypotheses. As an outcome, for both physical and mental health, AI can enable preemptive health and medicine to identify biomarkers and associations for disease and potential interventions, both drug-based and other activities.
Other Issues
As they addressed Professor Huckman’s initial question, panelists also touched on several other key issues, including:
- Public health and privacy: Preemptive medicine and prevention can cover a large population or a targeted one (“precision public health”). How do we protect individuals’ data and privacy? Can a large model using long context help to avoid privacy issues?
- Reliability: Despite AI’s promise, we still need human intervention when making clinical recommendations, advising on disease prevention, and ensuring AI-enabled designs are safe to produce. How do we trust what the models tell us?
- Data quality and representativeness: When integrating multimodal data (e.g., images, biological), we need a threshold for data quality. When using data from different populations, how can we be sure the model is trained on different populations and that the data represents different populations appropriately?
- Business models and regulation: Given data privacy and regulatory issues, we need business models and regulations in place to enable data sharing to make these systems work. How will insurers be incentivized to invest in these systems? How will providers be compensated for working with patients?
And finally, the panel reflected on how they might answer the initial question 5 years from now, and agreed that we will still be exploring the potential benefits of AI, including increased access, reduced costs, personalized medicine and prevention, and integrated models and agents. At the same time, we will continue to address concerns around sharing these capabilities broadly and equitably worldwide, guided by relevant business models and regulations.
Meet the Speakers
Francesca Dominici is a data scientist focused on analyzing data to explore the health impacts of environmental threats and inform policy. Dr. Dominici received her B.S. in Statistics from University La Sapienza in Rome, Italy and her Ph.D. in Statistics from the University of Padua in Italy. She did her postdoctoral training at the Bloomberg School of Public Health at Johns Hopkins University.
Nicolò Fusi is a General Manager at Microsoft Research (MSR) in Cambridge, MA, where he leads multidisciplinary teams on artificial intelligence, biomedical machine learning, and statistics to accelerate scientific discovery through generative AI. Fusi received a Ph.D. from the University of Sheffield and Bachelor’s and Master’s degrees in Computer Science from the University of Milan.
Sham Kakade is the Gordon McKay Professor of Computer Science and Statistics and the Co-director of the Kempner Institute at Harvard University. His research focuses on the engineering, scientific, and mathematical aspects of deep learning. Kakade received a Bachelor of Science in Physics from Caltech and a Ph.D. in Computer Science from the University College London’s Gatsby Computational Neuroscience Unit.
Armen Mkrtchyan is an Origination Partner at Flagship Pioneering, leading its Pioneering Intelligence™ initiative to institutionalize and expand the use of artificial intelligence and drive Flagship’s AI research to identify breakthrough innovation platforms. Mkrtchyan was awarded a Ph.D. in Aeronautics & Astronautics from MIT and a B.Sc. in Electrical Engineering from the University of North Dakota.
Additional Resources
Francesca Dominici
- Comparing traditional and causal inference methodologies for evaluating impacts of long-term air pollution exposure on hospitalization with Alzheimer’s disease and related dementias (Article) – Impacts of three air pollutants (fine particulate matter, nitrogen dioxide, and summer ozone) on elderly patients’ rate of first hospitalization with ADRD diagnosis.
- Using Data Science to Study Air Pollution Effect on COVID-19 Outcomes (Podcast) – Informing public health policy through research on environmental health science, data science, climate change, and health policy.
Nicolò Fusi
- Virchow2: Scaling Self-Supervised Mixed Magnification Models in Pathology (Article) – The result of scaling data and model size in models for computational pathology applications.
- Tag-LLM: Repurposing General-Purpose LLMs for Specialized Domains (Article) – How to repurpose general LLMs into effective task solvers for specialized domains, such as physical and biomedical sciences.
Sham Kakade
- Learning an Inventory Control Policy with General Inventory Arrival Dynamics (Article) – Learning and backtesting inventory control policies in a quantity-over-time arrivals model (QOT).
- Pareto Frontiers in Neural Feature Learning: Data, Compute, Width, and Luck (Article) – How sparse initialization and increasing network width can yield improvements in sample efficiency.
Armen Mkrtchyan
- Biology and AI: Twin Engines Built for Breakthrough Innovation (Essay) – Leveraging AI for biology and vice versa — creating a symbiotic relationship to power innovation.