Mortality Prediction in ICUs Using A Novel Time-Slicing Cox Regression Method.

Area of Science:

• Computational biology
• Medical informatics
• Machine learning

Background:

• Machine learning and data mining are increasingly used for clinical prediction in intensive care units (ICUs).
• A significant gap exists in fully utilizing the rich time-series data generated within ICUs.
• Existing methods often fail to optimally leverage the temporal dynamics inherent in ICU data.

Purpose of the Study:

• To address the underutilization of ICU time-series data for clinical prediction.
• To propose a novel machine learning approach, Time Slicing Cox regression (TS-Cox), for enhanced analysis.
• To improve the accuracy of mortality prediction in ICUs.

Main Methods:

• Developed Time Slicing Cox regression (TS-Cox), extending classical Cox regression for multi-dimensional time-series classification.
• Incorporated discriminative features from time-series data.
• Exploited temporal orders of features using a Cox-like function, differentiating from traditional classifiers like logistic regression and support vector machines.

Main Results:

• Empirical evaluation on the MIMIC-II database demonstrated the efficacy of the TS-Cox model.
• TS-Cox significantly outperformed baseline models in terms of AUC_PR, sensitivity, and PPV.
• The model effectively integrates feature discrimination and temporal information.

Conclusions:

• TS-Cox is a promising novel tool for mortality prediction in ICUs.
• The approach effectively leverages multi-dimensional time-series data by considering temporal dependencies.
• This method offers a significant advancement over traditional classification techniques for ICU clinical prediction.