MiranDa: Mimicking the Learning Processes of Human Doctors to Achieve Causal Inference for Medication Recommendation

TL;DR

MiranDa introduces a causal-inference framework that mimics physician learning by pairing supervised evidence-based training with gradient-space reinforcement learning guided by counterfactual outcomes expressed as $ELOS$ (estimated length of stay). By using two retrieval-based action-space expansions and a reward based on $ELOS$ differences, MiranDa refines medication recommendations within a hyperbolic geometric analysis of structured drug combinations. On MIMIC-III and MIMIC-IV datasets, MiranDa achieves near-identical real LOS for counterfactual evaluation and superior metrics (e.g., ROC AUC and PR AUC) while revealing procedure-specific attributes and sparser, more targeted medication regimens. This work advances a general, causal-inference–driven paradigm that can be applied to diverse medical tasks and beyond, though it notes computational overhead and the need for cautious interpretation in observational settings.

Abstract

To enhance therapeutic outcomes from a pharmacological perspective, we propose MiranDa, designed for medication recommendation, which is the first actionable model capable of providing the estimated length of stay in hospitals (ELOS) as counterfactual outcomes that guide clinical practice and model training. In detail, MiranDa emulates the educational trajectory of doctors through two gradient-scaling phases shifted by ELOS: an Evidence-based Training Phase that utilizes supervised learning and a Therapeutic Optimization Phase grounds in reinforcement learning within the gradient space, explores optimal medications by perturbations from ELOS. Evaluation of the Medical Information Mart for Intensive Care III dataset and IV dataset, showcased the superior results of our model across five metrics, particularly in reducing the ELOS. Surprisingly, our model provides structural attributes of medication combinations proved in hyperbolic space and advocated "procedure-specific" medication combinations. These findings posit that MiranDa enhanced medication efficacy. Notably, our paradigm can be applied to nearly all medical tasks and those with information to evaluate predicted outcomes. The source code of the MiranDa model is available at https://github.com/azusakou/MiranDa.

Figures (8)

• Figure 1: The training process of MiranDa
• Figure 2: The predictor (agent) of MiranDa. This predictor is also the baseline of this study.
• Figure 3: Retrieved process and the calculation of ELOS
• Figure 4: Morphology Comparison among Gradient Descent Strategies: Vanilla, Scaling, and Perturbation. This figure provides a demonstrative portrayal of gradient descent dynamics employing three illustrative strategies: a straightforward Vanilla approach; Vanilla augmented with Scaling; and an integrated strategy combining Vanilla, Scaling, and Perturbation. For clarity, the right subfigures zoom into selected regions to highlight differences. The red lines and arrows at the bottom of the figure represent the gradient contour plot and the gradient direction of the respective point, respectively.
• Figure 5: Comparative Analysis of Spatial Positioning in the Poincaré Model for Validation Set during Training. This figure shows the training trajectory of both MiranDa and the baseline method, and the disc at the center illustrates the actual spatial distribution from human doctors. MiranDa achieved its training in six epochs, compared to the baseline's eight epochs. The centrally positioned disc symbolizes genuine spatial distribution. Each point indicates a medication combination from a patient's hospitalization. The structured data information inherently featured in human doctor decisions is mimicked by MiranDa to replicate this hierarchical structure effectively. In contrast, the baseline method demonstrated a conspicuous absence of a discernable hierarchical structure, reflecting a potential discrepancy in the emulation of human-like decision-making processes. In the utilization of the Poincaré model for organizing or representing data, the position of an object approaches the center of the model, the attributes or characteristics it represents exhibit a higher level of foundational significance and abstraction within the entire structure. This implies that points situated near the center symbolize more fundamental or universal concepts, whereas those further from the center correspond to more concrete or specific instances.