OBJECTIVES: The purpose of this study was to identify septic patients at greater risk of longer stay in the ICU, using a random forest machine learning algorithm.

METHODS: This retrospective study used ICU data of septic patient survivors from the Medical Information Mart for Intensive Care III between 2001 and 2012. High risk patients were defined as having >90% probability of having longer LOS. Longer LOS was defined as LOS equal to or greater than the median. Demographics, lab values, and clinical characteristics captured 12 hours before to 24 hours after ICU admission were used as LOS predictors. A random forest predictive model was built in a training dataset and validated separately. The representative tree of the forest was extracted to visualize predictive patient subgroups of interest. The predictive accuracy (sensitivity and specificity) of the algorithm was evaluated using area under the receiver operating curve (AUC).

RESULTS: We identified 712 patients, of whom 56.5% were male. The mean age was 62.3 years, mean LOS was 8 days, and the median LOS was 5 days. The top predictors of LOS in order of most important included: mechanical ventilation use, Oxford Acute Severity of Illness Score, Logistic Organ Dysfunction Score (LODS), Glasgow Coma Scale, Simplified Acute Physiology Score II, Acute Physiology Score III, vasopressor use, quick Sepsis-related Organ Failure Assessment, urine output, and shock index. High risk patients were identified as those on mechanical ventilation with a “heart rate (HR) >118 bpm and Systemic inflammatory response syndrome criteria >3” or “HR ≤118 bpm, LODS >10, and urine output ≤1361 ml”. The AUC was 0.86, sensitivity was 0.88, and specificity 0.71.

CONCLUSIONS: Identifying septic patients at risk of potentially longer LOS may help decision makers optimize therapeutic intervention and resource allocation. Using random forest methodology, we identified subgroups instead of individual risk factors.