Making Clinical Language Models Auditable: CAST Introduces Concept-Guided Artifact Suppression Tuning

Summary

Clinical language models often achieve strong in-hospital accuracy but fail under deployment shifts because they rely on note-specific artifacts such as templates, separators, and boilerplate rather than genuine patient state. A paper by Jin Mu and Guanhua Chen (arXiv:2608.27397) proposes CAST (Concept-guided Artifact Suppression Tuning), an SAE-based framework for auditable clinical text classification. CAST uses Sparse Autoencoders to expose sparse, human-auditable features from intermediate Transformer activations, labels SAE latents via an LLM-assisted interpretation pipeline with ICD-10 retrieval constraints, suppresses verified artifact latents through residual subtraction during fine-tuning, and offers post-hoc per-concept attributions for auditing model decisions. On MIMIC-IV discharge-note mortality prediction, CAST outperforms comparable fine-tuned encoder baselines and remains competitive with strong LLM baselines, while producing feature-level audit trails that link each prediction to clinical concepts and to artifact concepts suppressed during training.