History-conditioned temporal machine learning for personalized ventilator weaning in ARDS

Ventilator weaning in Acute Respiratory Distress Syndrome (ARDS) involves substantial inter-patient heterogeneity and rapidly evolving physiological trajectories. Although weaning trials are routinely performed, the extent to which their associations with extubation outcomes vary across patient histories remains poorly characterized in observational ICU settings. To develop and evaluate a history-conditioned double machine learning framework for identifying heterogeneous associations between weaning trials and extubation outcomes in longitudinal critical care data. We conducted a retrospective cohort study of 4,287 adult ARDS patients who received ≥ 24 h of invasive mechanical ventilation in the MIMIC-IV database (2008–2019). The framework integrates cross-fitted ensemble nuisance models (XGBoost, Random Forest, LSTM, and Temporal Convolutional Networks) with orthogonalized score estimation to adjust for high-dimensional baseline and time-varying covariates. Data were chronologically split into training, validation, and a later admissions test set. The primary outcome was successful extubation, defined as remaining free from invasive mechanical ventilation for at least 48 h without reintubation. The proposed approach achieved strong discriminative performance (AUROC 0.869, 95% CI 0.858–0.880; AUPRC 0.907), outperforming logistic regression, gradient boosting, non-temporal DML, and recurrent neural network baselines (all p < 0.01). Patient-specific association estimates demonstrated substantial heterogeneity, ranging from negative to strongly positive values, indicating clinically meaningful variation in the relationship between weaning trial initiation and extubation success across evolving patient states. History-conditioned double machine learning provides a data-driven framework for uncovering heterogeneous associations in longitudinal ICU time series under observed clinical practice. This approach offers a useful basis for understanding variability in ventilator liberation outcomes and may inform future prospective studies in ARDS care.