DeePedIMV: Deep learning model for early prediction of invasive mechanical ventilation in critically ill children

Acute deterioration requiring invasive mechanical ventilation (IMV) is common in critically ill children and may arise from respiratory, cardiac, neurological, or systemic causes. Timely recognition of IMV need is challenging because of the heterogeneity of pediatric intensive care unit (PICU) patients and variability in clinical decision-making.

To develop and evaluate DeePedIMV, a deep learning model designed to predict IMV requirements in advance and support early clinical recognition of deterioration.

We analyzed retrospective electronic health records of patients aged <18 years admitted to a tertiary PICU over a 10-year period. Using time-series clinical data, DeePedIMV was trained to predict IMV up to 8 h before intubation. Performance was compared with pediatric early warning score (PEWS) and conventional machine-learning models. Threshold-based metrics were assessed at matched specificity levels.

Among 1318 admissions (688 IMV cases, 52.2%), DeePedIMV achieved the highest AUROC (0.88) compared with Random Forest (0.82), XGBoost (0.80), and modified PEWS (0.62). It also demonstrated superior precision-recall performance (AUPRC 0.47). At matched specificity levels, DeePedIMV achieved higher positive predictive value and likelihood ratios than comparator models, with lower simulated alert frequency. Performance remained consistent across age groups, with particularly strong accuracy in infants.

DeePedIMV effectively predicted IMV requirements up to 8 h before clinical deterioration. Although this single-center retrospective study requires external validation and prospective evaluation, the model shows potential as a data-driven decision-support tool in pediatric critical care.