Machine Learning-Based Predictive Risk Assessment for Preterm Infants: A Clinical Decision Support Approach

In neonatal care, premature babies with underdeveloped organs often require ventilatory support and subsequent transfers to specialist regional hospitals. These transfers impose substantial financial burdens on healthcare systems and families due to significant logistics and relocation costs. There is a critical need for a risk stratification mechanism to categorize premature neonates based on the severity of their conditions, facilitating appropriate clinical decisions. Specifically, this would help in deciding whether to transfer only the critically ill babies to a regional Neonatal Intensive Care Unit (NICU) or to treat low-risk babies at local hospitals. Current literature on this topic is sparse and lacks substantial contributions to the assessment of preterm babies' conditions. This study introduces an AI-based risk stratification tool aimed at optimizing neonatal resource use, thus minimizing overall costs. The tool was developed using synthetic data under the guidance of a consultant paediatrician from Cambridge University Hospitals NHS Foundation Trust and uses clinical parameters like Tidal Volume, Peak Inspiratory Pressure (PIP), Positive End Expiratory Pressure (PEEP), Respiratory Time, Inspiratory Time, and Oxygenation levels. This paper uses machine learning models for risk categorization and compares the performance of several models: Random Forest Classifier, Gradient Boost, AdaBoost, and Artificial Neural Networks (ANN). Results indicate that in terms of accuracy score the Random Forest model outperforms other models with a score of 92.42%, while AdaBoost had the highest AUC of 94.50%. This tool could enhance healthcare quality by providing a software-based resource allocation solution for neonatal healthcare providers.