Although reinforcement learning (RL) has tremendous success in many fields, applying RL to real-world settings such as healthcare is challenging when the reward is hard to specify and no exploration is allowed. In this work, we focus on recovering clinicians' rewards in treating patients. We incorporate the what-if reasoning to explain clinician's actions based on future outcomes. We use generalized additive models (GAMs) - a class of accurate, interpretable models - to recover the reward. In both simulation and a real-world hospital dataset, we show our model outperforms baselines. Finally, our model's explanations match several clinical guidelines when treating patients while we found the previously-used linear model often contradicts them.

TLDR: We extract clinicians' treatment goals by interpretable GAM modeling and what-if reasoning

Although Generlized Additive Models (GAMs) are accurate and interpretable, current GAMs do not have differentiability and scalability. In this work we propose a deep learning version of GAM and GA2M that often outperform regular GAMs while remaining interpretable.

TLDR: We develop a deep-learning version of Generalized Additive Model (GAM) and GA2M that is both accurate and interpretable.

Generalized additive models (GAMs) are useful for data bias discovery and model auditing. But do they always tell the true story of your data, or just its own hallucinated patterns? Also, which GAM algorithm is more accurate and less lying? In this paper we benchmark total 7 different GAMs variants and conclude that tree-based models are more trustworthy. We also design several metrics to decide which GAM is better.

TLDR: We compared total 7 different GAMs and showed which GAM is more trustworthy.

What makes an image be labeled as a cat? What makes a doctor think there is a tumor in a CT scan? These questions are inherently causal, but typical machine learning (ML) models rely on associations rather than causation. In this paper, we incorporated human causal knowledge into the ML models to make them robust, and show our models still have high accuracy when the environment changes. This is crucial for models to transfer across different environments e.g. different hospital sites in medical applications.

TLDR: We make our models robust to environment shifts by making them depend on the causal features more and spurious features less.

TLDR: We characterize a feedback loop problem that clinicians changing their decisions based on an imperfect ML system that changes the future data distribution.

TLDR: We propose using generative models to ask counterfactual questions to interpret a black-box model (e.g. DNN).

TLDR: We propose a reinforcement learning approach to help better allocate healthcare resouces for measurement scheduling.