Development And External Validation Of A Pretrained Deep Learning Model For The Prediction Of Non-accidental Trauma

We developed and externally validated PABLO, a pretrained, longitudinal deep learning model for the prediction of NAT. This is the largest study of NAT prediction to date, and the first to report external validation of the predictive model. PABLO significantly outperformed both traditional machine learning algorithms and a state-of-the-art deep-learning model. The data used for prediction are ubiquitous in health systems, allowing for implementation in any setting collecting patient demographic, diagnostic, and procedural data. PABLO may have benefited from a pretraining methodology that combines both contextual masked representation learning and multiple diagnosis forecasting tasks.