Hawes Hall, Classroom 201, Harvard Business School
Deep Causal Behavioral Policy Learning for Healthcare
The vast majority of clinical knowledge cannot be found in clinical textbooks or guidelines, but is acquired tacitly through years of practice. Much of this knowledge is encoded in the actions of clinicians responding to the high-dimensional, multimodal, and dynamically evolving contexts of their patients, captured in electronic health records. We present an approach for leveraging this knowledge to improve healthcare delivery in diverse non-randomized settings. Our proposed methodology uses deep learning algorithms to learn the distribution of high-dimensional clinical action paths given observed history, and identifies the causal link between action paths and patient outcomes. Specifically, our approach: (1) identifies the causal effect of provider assignment on clinical outcomes; (2) identifies the optimal provider given patient characteristics; (3) learns the distribution of clinical action paths that the optimal provider would take given evolving patient information; and, (4) identifies differences in care decisions that causally translate into differences in patient outcomes. Underlying this strategy, we construct a large clinical behavioral model (LCBM) based on transformer architecture and demonstrate its ability to estimate clinical behavioral policies (dynamic behavioral regimes). We propose a novel interpretation of a behavioral policy learned using the LCBM: that it is an efficient encoding of the complex, often implicit, knowledge used to treat a patient. This allows us to learn a space of policies for a range of healthcare applications, including causally rigorous quality measurement, provider coaching, upskilling, and clinical decision support.
Discussants:
- Issa Dahabreh (Associate Professor of Epidemiology, Harvard T.H, Chan School of Public Health)
- Yige Li (Research Fellow, Harvard Medical School)
