Sequential Data-Based Patient Similarity Framework for Patient Outcome Prediction: Algorithm Development.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Clinical Data Analysis

Background:

• Electronic medical records (EMRs) contain valuable sequential patient data for outcome prediction.
• Traditional methods struggle with the heterogeneity and irregularity of EMR data for similarity measurement.

Purpose of the Study:

• Develop a patient similarity framework using both sequential and cross-sectional EMR data.
• Enhance patient outcome prediction accuracy.

Main Methods:

• Calculated sequence similarity using edit distance for timestamped events.
• Measured trend similarity via dynamic time warping and Haar decomposition.
• Incorporated demographic, laboratory, and radiological data for cross-sectional similarity.
• Validated using k-nearest neighbors classifiers for mortality and readmission prediction in acute myocardial infarction patients.

Main Results:

• The similarity framework significantly outperformed baseline models on both public and private datasets.
• Predictive performance for mortality improved over time with more available data.
• Random forest and logistic regression models showed the best performance for mortality and readmission, respectively, within the first week.

Conclusions:

• The developed patient similarity framework effectively addresses challenges with uneven EMR data.
• The framework improves predictive performance for patient outcomes, aiding in early warning systems.