Making Clinical Language Models Auditable: Concept-Guided Fine-Tuning for Robust Prediction
CAST targets the shortcut features that can make clinical text models look accurate before they fail in deployment.
The paper proposes Concept-guided Artifact Suppression Tuning, using sparse autoencoders to surface auditable features inside Transformer activations. It labels those latent features with an LLM-assisted pipeline constrained by ICD-10 retrieval, then suppresses verified artifact latents during fine-tuning. On MIMIC-IV discharge-note mortality prediction, the authors say CAST improves over matching fine-tuned encoder baselines and stays competitive with strong LLM baselines. It also produces per-concept attributions showing which clinical concepts supported a prediction and which artifact concepts were suppressed. ArXiv · AI/CL/LG's note
The paper proposes Concept-guided Artifact Suppression Tuning, using sparse autoencoders to surface auditable features inside Transformer activations. It labels those latent features with an LLM-assisted pipeline constrained by ICD-10 retrieval, then suppresses verified artifact latents during fine-tuning. On MIMIC-IV discharge-note mortality prediction, the authors say CAST improves over matching fine-tuned encoder baselines and stays competitive with strong LLM baselines. It also produces per-concept attributions showing which clinical concepts supported a prediction and which artifact concepts were suppressed. ArXiv · AI/CL/LG's note
score 4