Prediction using patient comparison vs. modeling: a case study for mortality prediction

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Prediction Models

Background:

• Electronic Medical Records (EMRs) contain valuable data for predicting health states and enabling proactive interventions.
• The complex structure of EMRs presents challenges for standard machine learning applications.
• Comparing distinct predictive approaches for EMR data is crucial for advancing clinical decision support.

Purpose of the Study:

• To compare two distinct methods for generating health predictions from EMR data: temporal feature extraction with predictive modeling versus patient similarity metrics.
• To evaluate the performance and scalability of these approaches in predicting patient mortality.
• To determine the optimal strategy for leveraging EMR information for clinical prognostication.

Main Methods:

• Extraction of high-level temporal features from EMR data to build predictive models.
• Development and application of a patient similarity metric for outcome prediction based on similar patient cohorts.
• Comparative analysis of both approaches using the MIMIC-II Intensive Care Unit (ICU) dataset for mortality prediction.

Main Results:

• The predictive modeling approach, utilizing extracted temporal features, achieved a higher Area Under the Curve (AUC) of 0.84 for mortality prediction.
• The patient similarity approach demonstrated poorer scalability and resulted in a less accurate model with an AUC of 0.68.
• Patient mortality was accurately predicted with a median lead time of 72 hours using the superior modeling approach.

Conclusions:

• Feature extraction and predictive modeling represent a more effective and accurate strategy for EMR-based mortality prediction compared to patient similarity methods.
• The findings highlight the potential of advanced machine learning techniques applied to EMRs for improving proactive healthcare interventions.
• Accurate and timely prediction of patient mortality is achievable, offering significant implications for critical care management.