Delirium Prediction using Machine Learning Models on Preoperative Electronic Health Records Data

Area of Science:

• Medical Informatics
• Clinical Research
• Machine Learning in Healthcare

Background:

• Electronic Health Records (EHR) primarily serve administrative functions but offer valuable data for medical research.
• Predicting patient outcomes, such as postoperative delirium, is a key application of EHR data.
• Postoperative delirium is a significant concern affecting patient recovery and healthcare costs.

Purpose of the Study:

• To evaluate the efficacy of preoperative Electronic Health Records (EHR) in predicting postoperative delirium.
• To compare the performance of seven distinct machine learning models for delirium prediction.
• To identify key patient factors influencing the risk of developing postoperative delirium.

Main Methods:

• Utilized preoperative Electronic Health Records (EHR) data for analysis.
• Implemented and compared seven machine learning models: linear models, generalized additive models, random forests, support vector machines, neural networks, and extreme gradient boosting.
• Assessed model performance using comprehensive prediction metrics, focusing on sensitivity.

Main Results:

• Random forests and generalized additive models demonstrated superior performance in predicting postoperative delirium compared to other models.
• These top-performing models showed particular strength in sensitivity for delirium prediction.
• Identified significant risk factors for delirium, including age, substance abuse history, socioeconomic status, medical problem severity, and attending surgeon.

Conclusions:

• Machine learning models, particularly random forests and generalized additive models, can effectively predict postoperative delirium using EHR data.
• Preoperative EHR data contains crucial information for identifying at-risk patients.
• Factors such as patient demographics, medical history, and clinical context are vital predictors of delirium.