Why 'Zero-Shot' Clinical Predictions Are Risky

Date: April 19, 2026

Source: How to interpret 'zero-shot' results from generative EHR models — Nature Medicine

Key Takeaways

• Large Language Models (LLMs) are effective at simulating scenarios, making them useful for generating baseline experiences across different contexts.
• However, true prediction—especially in domains requiring high reliability—demands rigorous calibration and testing to reach "oracle-level" accuracy, which exceeds current LLM capabilities.
• LLMs do not always rely strictly on the input data provided when generating predictions, making it essential to verify which data the model actually used.
• As a result, a new evaluation paradigm is required to responsibly apply LLMs to predictive use cases.

Proposed Evaluation Framework

Performance by Frequency

Evaluate how well the model performs on rare versus common medical events.

Calibration

Ensure that predicted probabilities (e.g., 30% risk) align with real-world outcome frequencies.

Timeline Completion

Measure how often the model fails to generate a complete patient timeline.

Shortcut Audits

Assess whether the model relies on non-clinical shortcuts (e.g., administrative codes) instead of true medical signals.

Out-of-Distribution Validation

Test model performance on fundamentally different patient populations without retraining.