Predicting ICU Length of Stay Using Engineered Temporal and Physiological Features

Timely prediction of intensive care unit (ICU) length of stay (LOS) can improve bed management, staffing, and early care escalation. This study develops and evaluates early-horizon LOS prediction models using the MIMIC-IV database of adult ICU admissions. From the first hours of each stay, we derived physiological and temporal features including vital-sign and laboratory trends (rates of change, short-term variability), cumulative burdens (e.g., tachycardia minutes), and composite indicators of organ dysfunction while enforcing strict horizon censoring to avoid label leakage. We implemented the prediction as (a) continuous LOS regression and (b) as classification of prolonged stay (72 hours) at the 6-, 12-, and 24-hour decision points. Regression (regularized linear baselines, gradient-boosted trees, and recurrent neural networks) models were fitted using nested cross-validation and performance measured by mean absolute error, classification performance measured by area under the receiver operating characteristic curve, calibration, and decision-curve analysis performance. The analysis of features revealed temporal instability (e.g. increased lactate, increased creatinine and urine-output deviations) and ventilatory parameters as predictors. Across horizons, learned models demonstrated better calibration and higher net clinical benefit than age-only or early-warning-score baselines. Findings suggest that incorporating fine-grained temporal dynamics within the first day of ICU care can support earlier discharge planning and resource prioritization.