Interpretable machine learning and deep neural networks for ICU admission prediction in paediatric respiratory patients

Respiratory disorders represent a significant health challenge globally, especially in paediatric populations. Traditional diagnostic methods including clinical evaluation, imaging, pulmonary function tests, and invasive procedures, etc., often involve high costs, discomfort, and risks of misdiagnosis. In this context, machine learning (ML) techniques, particularly Deep Neural Networks (DNNs), offer a promising alternative by effectively modelling complex clinical and demographic data to improve prediction accuracy. This study proposes different ML based approaches, and DNN based approaches coupled with cross-entropy and triplet losses, to accurately predict Intensive Care Unit (ICU) admission for paediatric patients diagnosed with respiratory diseases. The DNN architecture consists of multiple fully connected layers that learn hierarchical feature representations, enabling robust prediction of ICU admission risk. The model achieved an accuracy of 97% and a precision of 98%, outperforming conventional ML methods such as decision trees, Random Forests, Logistic Regression, AdaBoost, XgBoost, K-Nearest Neighbours (KNN), and CatBoost. Furthermore, explainable AI techniques such as Shapley Additive Explanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and Explain Like I’m 5 (ELI5) were employed to interpret the decisions made by algorithms, identifying key predictive features such as cyanosis, heart rate, pneumococcal and paleness. These findings demonstrate that DNNs, combined with explainability frameworks, can enhance predictive performance and transparency in paediatric respiratory disease diagnosis, thereby potentially improving clinical decision-making and patient outcomes.