Latent-state models for precision medicine

• DOI: 10.48550/arxiv.2005.13001
• Publication Date: 2020-01-01
• Authors: Zekun Xu; Eric B. Laber; Ana‐Maria Staicu; Emanuel Severus
• Keywords: Methodology (stat.ME); FOS: Computer and information sciences; 0101 mathematics; 01 natural sciences; Statistics - Methodology; 3. Good health

Observational longitudinal studies are a common means to study treatment efficacy and safety in chronic mental illness. In many such studies, changes may be initiated by either the patient or their clinician can thus vary widely across patients timing, number, type. Indeed, observational pathway of STEP-BD bipolar depression, one motivations for this work, no two have same history even after coarsening clinic visits weekly time-scale. Estimation an optimal regime using data is challenging as cannot naively pool together with history, required methods based on inverse probability weighting, nor it possible apply backwards induction over decision points, done Q-learning its variants. Thus, additional structure needed effectively information within time. Current scientific theory illnesses maintains that patient's disease status conceptualized transitioning among small number discrete states. We use inform construction partially observable Markov process model health trajectories wherein observed outcomes dictated latent state. Using model, we derive evaluate estimators under paradigms quantifying long-term health. The finite sample performance proposed estimator demonstrated through series simulation experiments application study. find method provides high-quality estimates strategy settings where existing approaches applied without ad hoc modifications.