Design and Development of a Machine Learning Model for Predicting ICU Patients' Length of Stay

Background and aim Predicting the Length of Stay (LOS) for patients in the Intensive Care Unit (ICU) can aid in improving care management and resource allocation. Compared to traditional scoring systems, machine learning methods usually provide more accurate LOS predictions, highlighting the need for improved precision. This study aims to introduce and assess the stacked ensemble machine learning model capabilities to predict ICU LOS for both short- and long-term patient groups, using Acute Physiology and Chronic Health Evaluation (APACHE) IV-derived features. Methods We used approximately 148,000 patient records from the eICU Collaborative Research Database. To predict patient LOS, we first divided the patients into two groups (short- and long-term), based on their actual ICU LOS. Subsequently, we developed two stacked ensemble learning models for each patient group. Results For short-term patients, the Mean Absolute Error (MAE) was 1.037, whereas for long-term patients it was 1.997. Particularly, for patients with long-term actual ICU LOS (median 10.6 days), the respective predictions were clinically acceptable, suggesting the model's dynamics in real ICU environment applications. The clinical utility of these predictions can help clinicians in managing patient care in the ICU. Validation of these findings on external datasets may increase the applicability in daily clinical practice. Limitations include the use of data only from United States ICUs, selection of input features, and lack of external validation, which may affect the further generalizability and applicability of our model to different clinical settings. Conclusion The clinical utility of these predictions can help clinicians in managing patient care in the ICU. Future validation of these findings on external datasets may increase the applicability in daily clinical practice.