Prediction of Respiratory Decompensation in Patients Receiving Home Mechanical Ventilation: Machine Learning Model Development and Validation Study

Background: Chronic respiratory diseases often require long-term ventilatory support, leading to a growing number of patients treated with home mechanical ventilation (HMV). Despite advancements in telemonitoring with real-time tracking of noninvasive mechanical ventilation enabled by integrated software in HMV devices, early signs of respiratory decompensation may go unnoticed, leading to emergency visits and hospitalizations, which burden both patients and health care systems. Objective: This study aims to develop and evaluate a machine learning-based model capable of predicting respiratory decompensation events, defined as emergency visits or hospitalizations due to acute deterioration in the patient's underlying respiratory condition. These events reflect episodes of worsening respiratory status that require urgent medical attention. The model uses data from HMV telemonitoring platforms, with the aim of improving patient outcomes. Methods: This retrospective study analyzed data from 482 patients on HMV monitored via ResMed, Philips, and Breas platforms, collected between March 2021 and November 2024 at the Germans Trias i Pujol Hospital in Catalonia, Spain. Data included device use, compliance, mask leakage, and ventilator settings. Decompensation was defined as emergency department visits or hospitalizations. A windowing strategy captured the 5 weeks prior to events. Multiple machine learning models were trained using grid search to identify the optimal hyperparameters, prioritizing recall in order to minimize false negatives. Models were evaluated using 10-fold cross-validation. Finally, Shapley Additive Explanations (SHAP) analysis was used for model interpretability. Results: The final dataset included 157 data windows, balanced for positive and negative cases. Among the models tested, logistic regression achieved the highest recall (mean 0.94, SD 0.06, 95% CI 0.90-0.98) though with moderate accuracy (mean 0.60, SD 0.05, 95% CI 0.56-0.64). The random forest classifier achieved the best balance across metrics (accuracy: mean 0.66, SD 0.10, 95% CI 0.59-0.73; recall: mean 0.78, SD 0.15, 95% CI 0.67-0.89; F1-score: mean 0.70, SD 0.10, 95% CI 0.63-0.77). SHAP analysis revealed that higher use, leakage, and compliance in the week before a decompensation event were key predictors, suggesting compensatory behavior or early clinical deterioration. Overall performance remained moderate, reflecting limitations in sample size, incomplete longitudinal records for daily data, and the absence of key physiological measurements. Conclusions: This study demonstrates the feasibility of predicting respiratory decompensation using data from HMV telemonitoring systems. Tree-based ensemble models, particularly random forests, provided the most balanced performance, while SHAP analysis offered clinically relevant insights. Although performance was moderate, the findings support further development of predictive tools to enable timely telemedical interventions. Limitations include sample size, missing physiological parameters, and a single-center design. Future research should expand to multicenter datasets and incorporate additional clinical variables to enhance model robustness and generalizability.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.