Predicting Diabetes Outcomes in ICU Patients Using AdaBoost Machine Learning

Sanchez-Gomez J.S.; Bravo J.T.; Sobrino I.C.A. · 2025 IEEE 4th Colombian BioCAS Workshop, ColBioCAS 2025 - Conference Proceedings (2025) · DOI: 10.1109/colbiocas66416.2025.11276155

Resumen

The timely diagnosis of Diabetes Mellitus (DM) in Intensive Care Unit (ICU) patients is critical for improving patient outcomes, particularly in preventing complications such as infection, multi-organ failure, and elevated mortality rates. However, early detection remains challenging due to subtle metabolic symptoms and incomplete patient histories. This study presents an AI-driven approach using a machine learning (ML) model to predict adverse outcomes in ICU patients with diabetes, utilizing the AdaBoost algorithm, a popular ensemble learning method. The predictive model was trained on a dataset from the Global Open Source Severity of Illness Score (GOSSIS) initiative, which includes over 90 clinical features such as demographic data, vital signs, and laboratory measurements. The model achieved an Area Under the ROC Curve (AUC) of 0.83, indicating good discriminative power, and an accuracy of 83.28 %, suggesting high overall classification performance. Despite a relatively high specificity $(93.95 \%)$, recall was lower $(41.95 \%)$, signaling a missed opportunity for early identification of adverse events in some patients. Key predictor variables included glucose levels, body mass index, age, renal function markers (e.g., creatinine), and bicarbonate levels. While the model shows promise for clinical application, further optimization, particularly improving recall, is necessary to enhance its sensitivity and early detection capabilities. This approach demonstrates the potential for machine learning to augment clinical decision support systems, improving the prognosis for ICU patients with diabetes.

Instituciones: Universidad Peruana de Ciencias Aplicadas