Enhancing mortality prediction in intensive care units: improving APACHE II, SOFA, and SAPS II scoring systems using...

Area of Science:

• Critical Care Medicine
• Biomedical Informatics
• Machine Learning in Healthcare

Background:

• Traditional intensive care unit (ICU) mortality prediction relies on logistic regression (LR) models using worst-value data, neglecting temporal dynamics.
• Existing scoring systems like APACHE II, SOFA, and SAPS II have limitations in capturing the full patient data trajectory within the crucial first 24 hours.

Purpose of the Study:

• To enhance ICU patient mortality prediction accuracy by employing Long Short-Term Memory (LSTM) algorithms.
• To leverage the temporal information from time-series data collected within the initial 24 hours of ICU admission.

Main Methods:

• Retrospective analysis of 1336 ICU patients from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database.
• Development of LSTM-based prediction models using three data representations: original hourly, categorized hourly, and hourly total scores for APACHE II, SOFA, and SAPS II variables.
• Performance evaluation using Area Under the Receiver Operating Characteristic Curve (AUC), sensitivity, specificity, F1 score, and Brier score, compared against traditional LR models.

Main Results:

• LSTM models significantly outperformed LR models across all scoring systems (e.g., APACHE II: LSTM AUC 0.898 vs. LR AUC 0.777).
• LSTM models using original hourly data demonstrated superior accuracy, particularly for the APACHE II system (P=0.021).
• All LSTM models showed substantial improvements in predictive performance compared to conventional LR methods.

Conclusions:

• Incorporating time-series data with LSTM models substantially improves ICU mortality prediction accuracy.
• This temporal data-driven approach offers a significant advancement over traditional static scoring systems.
• LSTM-based models hold potential for redefining clinical decision support tools in critical care.