ISR Gating Mechanism: Building a Safety Barrier for Cardiovascular AI Diagnosis

Clinical large language models are widely used in the medical field, but they have a hidden risk—when the order of medical record sections changes, they may produce drastically different diagnostic results. The ISR gating mechanism detects this order-sensitive instability and automatically rejects unreliable predictions before critical decisions are made, providing a practical post-processing solution for medical AI safety.

Large language models show great potential in the medical field, but the long-neglected issue of section order sensitivity threatens system reliability. After rearranging the sections of the same medical record, the model's diagnostic results may be completely different. Especially in the diagnosis of cardiovascular diseases (such as acute myocardial infarction, heart failure, etc.), this instability directly relates to patients' lives, reducing the clinical value of AI or even bringing medical risks.

The Information Sufficiency Ratio (ISR) gating mechanism is an innovative post-processing method that intervenes after model prediction. Core process: 1. Generate variants by rearranging the sections of the original medical record multiple times; 2. The model predicts each variant independently; 3. Calculate the ISR value (prediction stability probability); 4. Reject the answer when it is below the threshold and transfer it to manual review.

The research team conducted experiments on the MIMIC-IV extended heart disease dataset (4476 records) using the Clinical-Longformer model. The results show: 32.4% of the medical records in the test set had their diagnostic labels flipped after section rearrangement, exposing the potential risks of current medical AI system deployment.

ISR gating brings a trade-off between coverage and accuracy: the baseline mode has 100% coverage but 84.3% accuracy; when the error rate upper limit is 5%, 44.2% coverage with 91.0% accuracy; mixed strategy (10% error rate) has 76% coverage with 93.0% accuracy. The technical implementation is lightweight, no need to retrain the model, can be seamlessly integrated into existing NLP pipelines, plug-and-play.

The ISR gating mechanism reveals the sensitivity of large models to input structure, which also exists in fields such as law and finance. Traditional accuracy metrics may mask fragility, requiring systematic testing of performance under input perturbations. The ISR gating mechanism provides a pragmatic solution, acknowledging technical limitations, and ensuring AI decisions are within controllable risks through an intelligent rejection mechanism—it is the safety philosophy for medical AI to move towards clinical practice.