Machine Learning-Driven Predictive Model for Early Detection of Patient Vital Sign Deterioration

Background: Delayed recognition of physiological deterioration remains a major challenge in acute care, often resulting in preventable adverse outcomes. Conventional early warning score systems are largely based on fixed threshold values, which can limit their ability to reflect the evolving and individualized patterns of patient instability, especially in resource-limited healthcare settings. Objective: This study aimed to develop and evaluate a machine-learning-based early warning framework for predicting patient vital-sign deterioration using routinely collected clinical data from a tertiary hospital in Nigeria. Methods: Vital-sign records obtained from ABUAD Multi-System Hospital, Ado-Ekiti, between January 2019 and February 2025 were retrospectively analyzed. The incidence of an unscheduled ICU admission, cardiac arrest, or the start of an emergency critical care intervention within a 24-hour prediction window was considered clinical deterioration. Data preprocessing included cleaning, normalization, and feature engineering based on a Modified Linear Early Warning Score (MLEWS). Four supervised machine-learning algorithms: Random Forest, Gradient Boosting, Support Vector Machine, and K-Nearest Neighbors, were trained and evaluated. Model performance was assessed using regression error metrics, classification indicators, cross-validation, receiver operating characteristic analysis, and SHAP (SHapley Additive exPlanations) -based interpretability. Results: Among the evaluated models, the Random Forest model demonstrated the strongest overall performance, achieving a cross-validated accuracy of 0.97 (±0.0116), sensitivity of 0.94, and specificity of 0.93. The area under the receiver operating characteristic curve was 0.98 for the training set and 0.96 for the testing set, indicating excellent discrimination and generalizability. Respiratory rate, oxygen saturation, and systolic blood pressure were identified as the most influential predictors. Conclusion: The proposed system accurately predicts patient deterioration and provides clinically interpretable insights using routine vital-sign data. Its performance supports integration into clinical decision-support systems to facilitate timely intervention and improve patient outcomes in resource-constrained settings.