Towards Reducing Diagnostic Errors with Interpretable Risk Prediction.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems

Background:

• Diagnostic errors are often caused by difficulties in accessing relevant patient information within Electronic Health Records (EHRs).
• Delays in diagnosis and incomplete differential diagnoses contribute significantly to medical errors.

Purpose of the Study:

• To develop a method utilizing Large Language Models (LLMs) to identify evidential data in EHRs.
• To mitigate diagnostic errors by providing clinicians with evidence-based risk assessments for specific diagnoses.
• To reduce diagnostic delays by offering individualized risk estimates during clinical uncertainty.

Main Methods:

• Proposed a Neural Additive Model for evidence-based predictions and individualized risk estimations.
• Employed LLMs to infer temporally fine-grained retrospective labels of diagnoses from pre-diagnosis EHR text.
• Refined evidence sets based on model-learned correlations for improved accuracy.

Main Results:

• Demonstrated a method for LLMs to extract diagnostic evidence from EHRs.
• Developed a model providing individualized risk estimates to support clinical decision-making.
• Evaluated the approach's utility in a simulated clinical setting for differential diagnosis.

Conclusions:

• The proposed LLM-based method enhances access to critical evidence within EHRs.
• This approach has the potential to significantly reduce diagnostic errors and delays.
• The system aids clinicians in making more informed decisions when facing diagnostic uncertainty.