Modeling length of stay as an optimized two-class prediction problem

Area of Science:

• Medical Informatics
• Clinical Prediction Modeling
• Health Services Research

Background:

• Predicting Intensive Care Unit Length of Stay (ICU LOS) is crucial for patient management and resource allocation.
• Dichotomizing ICU LOS into high-risk and low-risk groups is a common strategy, but optimal threshold selection remains challenging.

Purpose of the Study:

• To develop a novel method for simultaneously selecting the optimal dichotomization threshold for ICU LOS and building a predictive model.
• To identify high-risk patients requiring intensive care interventions.

Main Methods:

• A new performance statistic, MALOR, was introduced to compare model precision across different outcome dichotomizations, insensitive to prevalence.
• Class probability tree method was employed for predictive model development.
• Threshold selection and model development were performed concurrently.

Main Results:

• The method was applied to cardiac surgery patient data to dichotomize ICU LOS.
• The optimal precision for predicting ICU LOS in this cohort was achieved with a threshold of seven days.
• The developed predictive model demonstrated maximal precision at this threshold.

Conclusions:

• The presented method enhances predictive modeling by optimizing the outcome definition for prediction.
• This approach is applicable to various prediction problems requiring outcome dichotomization.
• The method is robust to changes in the prevalence of positive cases across different thresholds.