AI Model Detects Physician Fatigue Through Clinical Notes

How AI Identifies Fatigue Through Medical Documentation

Clinical notes serve as crucial records of physician-patient encounters, but they may also contain subtle indicators of physician fatigue. A recent study analyzed 129,228 emergency department visits to train a machine learning model that identifies notes written by physicians likely experiencing fatigue. The model specifically targeted physicians who worked emergency department shifts on at least five of the previous seven days, representing high-workload conditions that often lead to exhaustion.

The trained model demonstrated remarkable accuracy in identifying notes written by high-workload physicians in hold-out test sets. Beyond just detecting fatigue from multiple consecutive shifts, the model also successfully flagged notes written during other high-fatigue scenarios including overnight shifts and periods of exceptionally high patient volumes. This suggests the model captures genuine patterns of fatigue rather than simply counting work hours.

• Model trained on 129,228 emergency department visits
• Identifies notes from physicians working 5+ shifts in 7 days
• Detects fatigue patterns across multiple high-stress scenarios
• Shows 19% lower testing yield for heart attacks with fatigue
• LLM-generated notes show 74% higher fatigue prediction than human-written notes

Impact on Patient Care and Decision-Making

The most concerning finding emerged when examining patient outcomes associated with notes flagged for fatigue. For each standard deviation increase in model-predicted fatigue, the yield of testing for heart attacks decreased by 19%. This indicates that when physicians show signs of fatigue in their documentation, their clinical decision-making for those same patients appears significantly compromised. The correlation between note quality and patient care quality highlights the real-world implications of physician exhaustion.

A key feature of notes written by fatigued doctors is the predictability of the next word, given the preceding context

Perhaps the most surprising finding involves large language models and their generated clinical notes. Since word prediction forms the core functionality of how large language models operate, researchers discovered that LLM-generated notes showed 74% higher predicted fatigue scores than physician-written notes. This raises important questions about whether AI-generated medical documentation might introduce unintended distortions or patterns that mimic human fatigue, potentially affecting how these notes are interpreted and used in clinical settings. The study underscores the need for careful evaluation of AI-generated medical text to ensure it does not inadvertently replicate problematic human patterns.