Unlocking the Power of EHRs: Harnessing Unstructured Data for Machine Learning-based Outcome Predictions.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Data Science

Background:

• Electronic Health Records (EHRs) offer vast clinical data but often omit crucial social and behavioral health factors.
• Unstructured clinical notes contain vital patient context (e.g., social isolation, stress) often missed by current ML models.
• Over-reliance on structured clinical data in EHR research may lead to health outcome disparities.

Purpose of the Study:

• To assess how including patient-specific context from unstructured EHR data impacts ML algorithm accuracy and stability for mortality prediction.
• To evaluate the added value of non-clinical, daily life factors alongside clinical data in outcome prediction models.

Main Methods:

• Utilized the MIMIC III database for analysis.
• Developed and evaluated ML algorithms incorporating unstructured EHR data for mortality prediction.
• Assessed the impact of social and behavioral factors on model performance.

Main Results:

• Incorporating patient-specific context from unstructured notes significantly improved ML model discriminatory power and robustness.
• Non-clinical factors were found to be crucial for accurate patient outcome predictions.
• Generative models like GPT-4 show promise for extracting this contextual information.

Conclusions:

• Patient-specific context from unstructured EHR data is vital for accurate and stable ML-based mortality prediction.
• Future research should leverage advanced models like GPT-4 to integrate social and behavioral data for enhanced clinical decision support.
• Considering holistic patient data, including social determinants of health, is essential for equitable and precise health outcome predictions.