A Database-driven Decision Support System: Customized Mortality Prediction.

Area of Science:

• Critical Care Medicine
• Medical Informatics
• Biostatistics

Background:

• Current standard scoring systems for mortality prediction may lack accuracy for specific patient subpopulations.
• Existing models often aim for broad applicability, potentially sacrificing precision for heterogeneous patient groups.

Purpose of the Study:

• To investigate the hypothesis that locally customized mortality prediction models offer superior accuracy compared to established scoring systems.
• To develop and validate customized models for distinct patient cohorts, including ICU patients with specific conditions and elderly cardiac surgery patients.

Main Methods:

• Development of mortality prediction models using Logistic Regression (LR), Bayesian Networks (BN), and Artificial Neural Networks (ANN).
• Models were trained and tested on specific patient subsets from the Multi-parameter Intelligent Monitoring for Intensive Care (MIMIC) database and a cardiac surgical registry.
• Performance comparison against the Simplified Acute Physiology Score (SAPS) and EuroSCORE for respective patient groups.

Main Results:

• Customized models demonstrated significantly better performance across all tested subsets.
• For acute kidney injury patients, ANN achieved an AUC of 0.875 vs. SAPS AUC of 0.642.
• For subarachnoid hemorrhage patients, BN achieved an AUC of 0.958 vs. SAPS AUC of 0.84.
• For elderly open-heart surgery patients, ANN achieved an AUC of 0.94 vs. EuroSCORE AUC of 0.648.

Conclusions:

• Locally customized mortality prediction models provide more accurate risk stratification than generalized scoring systems.
• Developing models tailored to specific patient subsets using local data is a viable alternative to broad external validation.
• This localized approach enhances the precision of mortality prediction in critical care and specialized surgical populations.