Predicting Shock in Pediatric Patients Through Thermal Gradients and Machine Learning: A Multi-Model Approach

Abstract

Early detection of hemodynamic compromise in pediatric patients is critical for timely and effective intervention in intensive care. This study evaluates the use of thermal gradients, specifically the temperature difference between the abdomen and foot, as non-invasive physiological markers to improve prediction of shock. The dataset included thermal gradients, pulse rate, age, and four time-stamped measurements, enabling models to anticipate circulatory deterioration across different prediction horizons. These forecasting windows were examined to assess how far in advance the onset of shock could be reliably predicted. Several machine learning models were compared, and the best approach achieved an AUC of 0.84, with sensitivity of 0.90 and specificity of 0.74. Although methodological differences make direct comparison with previous studies challenging, this performance surpasses that reported in the existing literature. These findings highlight the potential of combining thermal gradients with conventional vital signs to enhance early and reliable risk stratification and support clinical decision-making in pediatric intensive care.