Evaluating the Potential of Machine Learning for Discharge Management on Routine Health Insurance Data.

Area of Science:

• Health Informatics
• Machine Learning Applications
• Data Science in Healthcare

Background:

• Routine health insurance claims (HIC) data offer valuable longitudinal patient information.
• Challenges like data volume, variability, and sparsity hinder the application of machine learning (ML) in HIC data.
• The KI-THRUST project aimed to overcome these barriers using a large-scale HIC dataset.

Purpose of the Study:

• To evaluate the performance of various ML models in predicting clinical outcomes from HIC data.
• To assess the impact of different feature sets on predictive accuracy.
• To interpret ML model predictions using explainable AI (XAI) methods.

Main Methods:

• Utilized a dataset of 2 million anonymized HIC records from four German providers.
• Compared linear and nonlinear ML models: logistic regression, AdaBoost, Random Forest, and Neural Network.
• Employed XAI techniques like Integrated Gradients, Shapley values, and LIME for result interpretation.
• Investigated feature sets including demographics, ICD-10 codes for conditions, and procedures.

Main Results:

• AdaBoost demonstrated the highest overall performance, with logistic regression showing comparable results.
• Incorporating diagnostic and procedural information (M2) significantly enhanced predictive performance.
• Time-related information did not yield a significant performance gain.
• XAI methods provided insights into model predictions.

Conclusions:

• Large-scale HIC data, when combined with ML and XAI, holds significant potential for healthcare applications.
• ML models can effectively predict clinical outcomes like mortality and unplanned readmissions.
• Future applications include enhanced discharge management systems and other healthcare decision support tools.