Research
Computational Precision Health develops new methods in computer science, machine learning and AI, statistics, data science, and health informatics to enable and evaluate new health interventions for real-world impact.
Computational Precision Health is a new paradigm for developing and deploying adaptive precision interventions for real-world impact. The science of CPH is the development and use of novel computational algorithms, tools, and architectures to increase precision in problem formulation, develop precision solutions that adapt to person, place, time, and context, and deploy and evaluate these solutions more efficiently and effectively.
Precision Problem Formulation: New computational methods in causal inference enable us to ask deeper and more precise questions than ever before. We can do more than predict. We can begin to understand why things happen, how to intervene, and when to change our actions. We can formulate problems that capture the complexity of the real world, including the reality that patients have more than one disease at a time, and that health care systems are highly complex organizations that often function within tight resource constraints, structural biases, and perverse incentives.
CPH faculty include practicing clinicians and leaders in health care operations and policy, ensuring that problem formulation is appropriately scoped and grounded in the real world.
Computational Precision Solutions: CPH solutions emerge from iterative cycles of problem formulation, development and testing of computational approaches, and real-world evaluation of prototype solutions. Solutions aim to augment human intelligence – whether of clinicians, patients, family caregivers, or public health practitioners – to deliver personalized interventions that incorporate uncertainty, time, and context.
CPH is building open software architectures and partnering with designated Computational Health in Practice (CHiP) sites at UCSF Health and beyond to facilitate rapid iteration and testing of computational precision solutions.
Precision Deployment: Unlike a drug that never changes once developed and marketed, the very nature of computational solutions is that their underlying models and real-world performance will drift as data, decision, and valuation contexts change over time. Novel standards-based and privacy-preserving architectures and evaluation methods are needed to ensure that computational solutions continue to deliver fair and equitable outcomes and are demonstrably safe and effective over time in each and every setting.
CPH faculty are experts at novel approaches to evaluation design, including use of new causal inference methods to evaluate real-world evidence, and are leaders in regulatory science and policy for AI/ML and digital health.
Core Research Areas
The majority of CPH faculty research falls into one or more of the following core research areas.
See select recent publications by CPH faculty
Research Highlights
Examining the New Ethical Issues Raised by Generative AI
The relational nature of large language models and other forms of generative AI raises additional questions for medicine, posing new ethical challenges beyond existing frameworks like the Belmont principles of beneficence, respect for persons, and justice. Drs. Ida Sim and Christine Cassel call for the establishment of proactive health AI ethics centers and regulatory bodies to ensure ethical standards are met as AI technologies rapidly evolve.
Leveraging AI and knowledge graphs to predict Alzheimer’s disease onset and risk
Dr. Marina Sirota, Dr. Sergio Baranzini and colleagues have used machine learning to predict Alzheimer’s disease up to 7 years before symptoms appear. Finding that high cholesterol and, for women, osteoporosis, were both predictive of risk, the work demonstrates the promise of combining AI, knowledge graphs and clinical data to ID genetic markers and other determinants of risk.
Conformal Meta-learners for Predictive Inference of Individual Treatment Effects
Machine learning (ML) to predict treatment response often gives only a single estimate of average treatment effect, without quantifying uncertainty based on errors in the model or individual variation. Dr. Ahmed Alaa and colleagues here introduce a new method that can provide valid inference of treatment effects for individual patients, and a reliable uncertainty estimate, using any ML model.