Learning Explainable Treatment Policies with Clinician-Informed Representations: A Practical Approach

Digital health interventions (DHIs) and remote patient monitoring (RPM) have shown great potential in improving chronic disease management through personalized care. However, barriers like limited efficacy and workload concerns hinder adoption of existing DHIs; while limited sample sizes and lack of interpretability limit the effectiveness and adoption of purely black-box algorithmic DHIs. In this paper, we address these challenges by developing a pipeline for learning explainable treatment policies for RPM-enabled DHIs. We apply our approach in the real-world setting of RPM using a DHI to improve glycemic control of youth with type 1 diabetes. Our main contribution is to reveal the importance of clinical domain knowledge in developing state and action representations for effective, efficient, and interpretable targeting policies. We observe that policies learned from clinician-informed representations are significantly more efficacious and efficient than policies learned from black-box representations. This work emphasizes the importance of collaboration between ML researchers and clinicians for developing effective DHIs in the real world.

💡 Research Summary

This paper addresses the pressing challenges of limited efficacy, clinician workload, and lack of interpretability that have hampered the adoption of digital health interventions (DHIs) and remote patient monitoring (RPM) systems, especially in the context of managing type‑1 diabetes (T1D) among youth. The authors propose a complete, end‑to‑end pipeline that learns explainable treatment policies by first constructing low‑dimensional representations of patient states and clinician actions, then estimating conditional average treatment effects (CATEs) for each state‑action pair, and finally ranking patients under capacity constraints to decide who should receive an intervention.

The dataset consists of three IRB‑approved clinical trials comprising 281 participants who wore continuous glucose monitors (CGMs) that transmitted glucose readings every five minutes. For each weekly review, clinicians either sent a text‑based treatment recommendation or did not. The raw patient state includes static demographics, time‑varying covariates (age, pump use), and a 2‑week CGM window (4032 dimensions). The raw action is the free‑form message text. The reward is defined as the change in Time‑in‑Range (TIR) during the week following the message.

Two families of state‑action representations are explored. The “black‑box” baseline directly compresses the high‑dimensional data: CGM traces are embedded using TS2Vec (a contrastive time‑series encoder) or reduced with UMAP, while messages are encoded with a 728‑dimensional PaLM language model and clustered via K‑means to obtain discrete action types. The “clinician‑informed” approach leverages domain knowledge that clinicians already use for decision‑making. For states, clinically meaningful summaries such as percentage of time in target range, low‑glucose episodes, pump usage, and age are extracted, then reduced to a handful of dimensions. For actions, a few‑shot labeling pipeline with Gemini Pro automatically tags each message with interpretable clinical categories (e.g., “increase pre‑dinner insulin dose”, “adjust diet”). This yields low‑dimensional, conditionally consistent representations that satisfy the authors’ Assumption 1, guaranteeing that any two original messages mapping to the same action label have identical expected outcomes given the same state representation.

With these representations, the authors estimate CATEs ρ(s,a)=E