SCOPE: predicting future diagnoses in office visits using electronic health records

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Decision Support

Background:

• Electronic Health Records (EHR) contain vast amounts of patient data.
• Physicians require efficient tools to navigate complex patient histories and predict diagnoses.
• Existing diagnostic prediction models may lack interpretability or scalability.

Purpose of the Study:

• To develop and evaluate an interpretable and scalable model for predicting likely diagnoses at patient encounters.
• To assist physicians in utilizing EHR data more effectively.
• To compare the proposed model's performance against deep learning methods.

Main Methods:

• Retrospective analysis of de-identified EHR data from 2,701,522 patients.
• Development of a multi-label classification model using a binary relevance strategy.
• Testing logistic regression and random forests as base classifiers with various time windows for data aggregation.
• Comparison with a recurrent neural network (RNN) based deep learning model.

Main Results:

• The best model, utilizing a random forest classifier with integrated demographic, diagnosis, and lab data, achieved a median AUROC of 0.904.
• Performance was comparable or superior to existing methods, including outperforming the deep learning model in AUROC.
• Model interpretability revealed meaningful feature associations, highlighting clinical relevance.

Conclusions:

• The proposed interpretable multi-label model demonstrates comparable performance to deep learning methods while offering enhanced simplicity and interpretability.
• The model is a promising candidate for clinical deployment in aiding diagnostic predictions.
• Further validation across multiple institutions is warranted.