Artificial intelligence algorithm to predict the requirement of neonatal endotracheal intubation within 3 h: application for clinical practice

Introduction: Timely intervention, such as endotracheal intubation (EI), is crucial for managing acute respiratory distress in the neonatal intensive care unit (NICU). Delays in EI can lead to significant adverse effects in neonates. This study aimed to develop a highly accurate predictive model to forecast the requirement for EI, allowing for proactive clinical planning and intervention up to 3 h in advance. Method: We developed a multimodal deep learning model designed to simultaneously analyze distinct data types. The model utilizes numeric initial clinical data and time-series vital sign data collected over the preceding 1-3 h. To rigorously evaluate the model's reliability and clinical applicability, we performed comprehensive external validation using independent patient datasets, specifically assessing generalization and bias. Result: The constrained model successfully predicted the requirement for EI with high predictive power across various forecasting intervals (up to 72 h in 1-h increments). Internal validation yielded an accuracy of 0.9579 and AUC of 0.9323, while external validation maintained high generalization (accuracy 0.9411, AUC 0.9336). Discussion: The proposed multimodal deep learning model provides an effective tool for the advance prediction of EI requirements in neonates. Given its high accuracy, confirmed generalization capabilities through external validation, and potential to prevent severe respiratory distress problems by facilitating proactive care, this model holds wide and significant applicability in clinical NICU environments.