TIDAL: A Temporal Causal Diffusion Framework for Visualizing Knee Osteoarthritis Treatment Outcomes

Le résumé fourni par la source

Generating medically-relevant, patient-specific counterfactual images from longitudinal medical data requires addressing confounding bias that static causal methods cannot handle. We introduce TIDAL (Temporal IPW Diffusion Adversarial Learning), a causal diffusion framework with three core algorithmic contributions: (1) a temporal propensity architecture that processes variable-length patient histories via LSTM to estimate interval-based multi-treatment probabilities, structurally distinct from static IPW methods that assume fixed covariates; (2) integration of temporal IPW weights into diffusion training that accounts for treatment sequences rather than single assignments; (3) adversarial training on trajectory-encoding latent representations that enforces treatment-invariance conditioned on longitudinal history. We provide theoretical justification via a risk decomposition theorem showing these mechanisms address complementary error sources: IPW bounds weighting error from imperfect propensity estimates, while adversarial training minimizes representation leakage. Evaluated on knee osteoarthritis X-rays from the Osteoarthritis Initiative, the only public dataset with longitudinal imaging, treatments, and clinical covariates, TIDAL achieves 5.24% improvement in SSIM and 15.9% reduction in treatment effect error versus baselines. Ablations confirm both components are necessary. Our contributions advance machine learning for causal inference in sequential decision-making contexts beyond medical imaging.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.