Predicting high-cost care in a mental health setting

Area of Science:

• Health Informatics
• Clinical Prediction Modeling
• Natural Language Processing

Background:

• Digital health records contain vast information for predicting cost-relevant outcomes.
• Electronic health records (EHR) offer potential for automated prediction of healthcare costs.
• Leveraging dense information in EHR is a key area for healthcare innovation.

Purpose of the Study:

• To assess the predictive power of routinely recorded EHR data for key service outcomes.
• To determine if natural language processing (NLP) tools enhance these predictions.
• To evaluate prediction accuracy for in-patient duration, readmission, and high service cost.

Main Methods:

• Utilized clinical data from a UK mental healthcare provider's EHR.
• Combined structured EHR data with text-derived data (diagnoses, medications, symptoms).
• Employed logistic regression and validated models using receiver operating characteristic curves.

Main Results:

• Full models achieved areas under the curve (AUC) from 0.59 to 0.85.
• Prediction for high service use yielded the highest AUC (0.85).
• Incorporating NLP-derived data increased explained variance by 12-46% across scenarios.

Conclusions:

• EHR data can significantly improve routine clinical predictions by accessing previously unavailable information.
• NLP tools are crucial for unlocking the full potential of unstructured EHR data.
• High service use prediction demonstrated the most robust performance, indicating strong clinical utility.