Abstract
PURPOSE: Benefit or harm from early mobilisation (EM) in mechanically ventilated patients may vary by individual patient characteristics. We used machine learning to predict individualised treatment effects (ITEs) in the "Early Active Mobilization during Mechanical Ventilation in the ICU" (TEAM) trial. METHODS: This was a secondary analysis of the TEAM trial using a causal inference approach to estimate ITEs, which compared enhanced EM to usual care EM. Baseline variables in the original publication were used as predictor variables. The primary outcome was death by day 180. The dataset was randomly split into two halves (train and test) by site. In the training data, fivefold cross-validation was used to compare six candidate machine learning algorithms. The best-performing model was evaluated in the test dataset. Patients were stratified into tertiles based on predicted ITEs, reflecting estimated benefit, no effect or harm. RESULTS: We included 687 patients from 40 sites, and 141 (20.5%) patients died by day 180. Predicted ITEs in the test cohort ranged from an absolute 34.0% reduction to a 39.3% increase in mortality with enhanced EM. The interaction term between the model predictions and treatment assignment demonstrated significant heterogeneity of treatment effect (p = 0.006). Patients predicted to respond poorly to enhanced EM therapy were more likely to receive vasopressors, have diabetes and have lower RASS scores at baseline, compared to patients predicted to have benefit. CONCLUSION: Using baseline characteristics, a machine learning model identified patients with estimated benefit or harm with enhanced EM. Future testing of a personalised approach to mobilisation in the ICU is warranted.