CAREER: Ethical Machine Learning in Health: Robustness in Data, Learning and Deployment

Health is an area of immense potential for machine learning (ML), due to the increasing complexity of care management and large volume of data becoming available. Recent work has shown that models in healthcare lack robustness, and do not perform equally well across all patients and settings. Recent work in general model robustness have failed to translate to health settings in part because they do not consider the full range of patients, conditions, and contexts that models will be used in. This project will create new ways to improve model robustness, and empower researchers to target more ethical deployments. This research will identify improvements for data use and model training that prioritize actionable models in health, by focusing on the nuance and complexity of health data. Ultimately these advances will also contribute to machine learning in other high-stakes areas such as lending, education and legal systems, that rely on routinely collected data to generate insights. Beyond the direct and long-term societal impact of these advances, this work will help lay the foundation for a new undergraduate-focused summer course focusing on bringing a larger pipeline of students into machine learning in health. The importance of patient safety combined with poor model robustness limits the practical utility of ML in healthcare, and ethical deployment requires developing methods and metrics to ensure state-of-the-art models are robust. This project targets three ways to develop robust health models: ensuring representations and downstream models withstand incorrect data associations, achieving fair and robust model learning, and enhancing post-hoc robustness to outlier data during testing. First, targeting representational robustness to data error and change, it will build resilient models across patient subpopulations and variations in care through contrastive self-supervision in deep metric models. Second, in model learning, it will improve algorithms for stable training, balancing fairness/robustness trade-offs by combining private and public data for clinical prediction tasks. Third, it will target test-time methods for outlier detection and extending pre-trained models to cover minority subgroups. The project will result in methods that address robustness in data, learning, and testing, as crucial steps toward ethically deploying health models. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.