Graph-Constrained Policy Boosts Clinical Code Prediction Accuracy

Predicting clinical codes (like ICD-10-CM) from unstructured discharge summaries is challenging due to the vast, sparse, and hierarchical nature of the label space. Most existing systems treat this as a flat multi-label classification, which often struggles with rare codes due to limited training signals. Researchers propose a novel graph-constrained traversal policy that reframes the problem as a finite-horizon decision process over a pruned code hierarchy. A single language model navigates this graph level by level, selecting valid child nodes until billable leaf codes are reached. This approach transforms extreme multi-label prediction into sparse, hierarchy-aware subset decisions, guaranteeing structurally valid outputs. Evaluated on MIMIC-IV discharge summaries, the supervised policy (SFT-1+) achieved significantly higher micro-F1 and macro-F1 scores than flat baselines, particularly improving performance for rare codes. The study also found that increasing supervised trajectory data consistently improved performance, while reinforcement learning offered no additional benefit over supervised continuation. This demonstrates that simple graph-constrained policy learning can be more effective than complex alternatives.