This study was conducted at Germans Trias i Pujol Hospital (HUGTiP), a reference hospital in Catalonia (Spain) that serves a population of 1,448,812 people across 70 municipalities. The study included data collected between March 2021 and November 2024. Patients were recruited voluntarily before their inclusion in the study. The study population comprised hospitalized patients who required HMV initiation after an acute exacerbation in the IRCU or in general hospitalization wards at HUGTiP. It also included patients already on HMV who required therapy adjustments before discharge. Additionally, patients attending outpatient clinics who required new HMV initiation or adjustments in follow-up consultations were included in the analysis.

The study population, comprising both women and men, had a mean age of 69.3 years, with most individuals aged between 60 and 80 years. Chronic obstructive pulmonary disease accounted for more than half of the patients treated with NIMV, representing the main indication for ventilatory support. Other underlying conditions included obesity hypoventilation syndrome (often associated with obstructive sleep apnea) and restrictive thoracic or neuromuscular disorders, such as postsurgical thoracic sequelae, amyotrophic lateral sclerosis, and muscular dystrophies.

To ensure appropriate follow-up and reliable data acquisition, only patients residing within the hospital’s coverage area were considered eligible for HMV. Furthermore, patients needed to have a designated caregiver to assist in therapy adherence and monitoring.

Patient monitoring was conducted through validated commercial telemonitoring platforms commonly used in HMV: ResMed, Philips, and Breas. These platforms allow for real-time tracking of NIMV therapy. The home respiratory therapy provider, Esteve Teijin, which is responsible for the geographic area of HUGTiP, granted access to its telemonitoring system through a licensed workstation located at the sleep unit in the Pulmonology Department at HUGTiP.

The telemonitoring system facilitated automated data extraction from ventilators, providing detailed information on therapy adherence and device performance through weekly and daily reports. The weekly reports included information on ventilator use patterns, ventilation settings, and device and mask characteristics. More specifically, the extracted data comprised weekly continuous compliance, last recorded values of expiratory and inspiratory positive airway pressure, ventilation mode, ramp activation and duration, device type, mask type, and size. Daily reports focused on key physiological indicators, such as use (in min), apnea-hypopnea index, and leakage levels.

To provide a comprehensive view of patient status, ventilator-derived parameters were supplemented with hospital admission and emergency department visit dates. Consequently, the study included 482 patients for whom both telemonitoring reports and corresponding date records were available.

This study complies with the approval of the Ethics Committee of HUGTiP (approval code PI-21‐236) and adheres to the ethical principles outlined in the Declaration of Helsinki [10], ensuring the highest standards of research ethics. Informed consent was obtained from all participants (or their legal guardians, where applicable) prior to enrollment. No financial compensation was provided for participation, and involvement was entirely voluntary. To ensure confidentiality, all personal identifiers, including names and medical record numbers, were replaced by the hospital’s information system professionals with unique study codes. These codes were used for data analysis, ensuring that no real patient information is disclosed. Patient autonomy and rights were respected throughout the study, with participant well-being prioritized above scientific or societal interests.

Each variable was first rescaled using min-max normalization, computing each feature’s global minimum and maximum across all reports. Next, to emphasize how recent each measurement was relative to a decompensation event, each report was multiplied by a decreasing time-decay weight. The closest report to the event received a weight of 1, while older reports were down-weighted by a constant decreasing factor, up to 0.5 for the oldest report. Finally, the dataset was balanced to ensure representation of both classes.

Different machine learning models were trained with grid search for hyperparameter tuning, ensuring that the best possible parameter configurations were identified. The models included logistic regression (LR), random forest classifier (RFC), decision tree classifier (DTC), naïve Bayes (NB), k-nearest neighbors (KNN), and extra trees classifier (ETC). For each model, a predefined set of hyperparameters was exhaustively explored. For LR, the hyperparameters included the regularization strength C, ranging from 0.01 to 100, the penalty types l1 and l2, the solvers lbfgs, liblinear, and saga, and maximum iterations ranging from 1000 to 1,000,000. The RFC and ETC were both tuned by varying the number of estimators between 10 and 500, adjusting the maximum depth from shallow to unrestricted (up to 30), and modifying the minimum number of samples required to split a node or form a leaf, while also exploring different splitting criteria such as gini, entropy, or log_loss. DTCs were similarly optimized by adjusting depth, split, and leaf requirements and splitting criteria to balance model complexity and generalization. KNN was tuned by varying the number of neighbors from 1 to 11, the weighting method between uniform and distance, and the algorithm among auto, ball_tree, kd_tree, and brute. Finally, for NB, the variance smoothing parameter ranged from 1·e-12 to 1·e-7.

Given the clinical context, where minimizing false negatives is crucial, grid search was configured to prioritize maximizing recall, thus ensuring that the number of false negatives is minimized, as failing to identify a positive case is more critical than a false positive in this setting. All models were implemented using the open-source Python (version 3.11.12; Python Software Foundation) library scikit-learn [11].

Additionally, two separate neural network models were developed. The first was a neural network (NN) with a depth ranging from 2 to 5 layers, selected based on performance. This model consisted of alternating fully connected layers and rectified linear unit activations, ending with a sigmoid output layer for binary classification. The second model was a recurrent neural network (RNN), composed of stacked RNN layers with tanh activation, followed by fully connected layers and rectified linear unit activations, also terminating in a sigmoid output. The optimal architecture for each model was empirically determined. Both models were trained using binary cross-entropy loss, and the limited-memory Broyden-Fletcher-Goldfarb-Shanno and Adam optimizers were evaluated and compared. The entire pipeline was implemented using the open-source Python library PyTorch [12].

Models were evaluated through 10-fold cross-validation, using accuracy, precision, recall, and F₁-score as performance metrics. Results are reported as mean (SD) with 95% CIs.

In this study, the Shapley Additive Explanations (SHAP) method [13] was used to assess the interpretability of the predictive models developed. Once the optimal model was identified, a SHAP explainer was created. This explainer is designed to compute SHAP values, where each value indicates the contribution of a corresponding feature toward the final prediction, thus breaking down the prediction into interpretable parts. Summary plots that display the global importance of features across the entire dataset were generated. These plots rank features by their average impact on predictions and reveal the direction of their influence, making it easier to interpret how changes in feature values affect the model’s outcomes.

The dataset resulted in a balanced sample size of 157 windows; thus, each fold holds 16 windows for testing and the remaining 141 for training.

Results show that LR with C=0.1, a maximum of 1000 iterations and the liblinear solver, achieved the highest mean recall of 0.94 (SD 0.06; 95% CI 0.90‐0.98) but exhibited a mean overall accuracy of 0.60 (SD 0.05, 95% CI 0.56‐0.64), a mean F₁-score of 0.71 (SD 0.03, 95% CI 0.69‐0.73), and a mean precision of 0.58 (SD 0.04, 95% CI 0.55‐0.61; Table 1). RFC, with 200 trees, a maximum depth of 5, and default splitting and leaf parameters, reached the highest mean accuracy of 0.66 (SD 0.10, 95% CI 0.59‐0.73), along with a mean recall of 0.78 (SD 0.15, 95% CI 0.67‐0.89), a mean precision of 0.66 (SD 0.13, 95% CI 0.57‐0.75), and mean F₁-score of 0.70 (SD 0.10, 95% CI 0.63‐0.77). The ETC, a related ensemble approach, using 50 trees with similar depth constraints, achieved a mean accuracy of 0.62 (SD 0.08, 95% CI 0.56‐0.68), mean recall of 0.81 (SD 0.12, 95% CI 0.72‐0.89), and mean F₁-score of 0.69 (SD 0.07, 95% CI 0.64‐0.74).

The single DTC with entropy criterion and unrestricted depth reached a mean accuracy of 0.57 (SD 0.12, 95% CI 0.48‐0.66), mean recall of 0.67 (SD 0.17, 95% CI 0.55‐0.79), mean precision of 0.58 (SD 0.09, 95% CI 0.52‐0.65), and mean F₁-score of 0.61 (SD 0.11, 95% CI 0.53‐0.69). KNN, with three neighbors and distance weighting, produced a mean accuracy of 0.60 (SD 0.15, 95% CI 0.49‐0.71), mean recall of 0.70 (SD 0.20, 95% CI 0.56‐0.84) and mean F₁-score of 0.64 (SD 0.15, 95% CI 0.53‐0.75). NB, tuned with a var_smoothing parameter of 1·e-11, showed a mean accuracy of 0.57 (SD 0.09, 95% CI 0.51‐0.64) and mean F₁-score of 0.60 (SD 0.08, 95% CI 0.54‐0.66). The neural network with hidden layers of 100 and 50 neurons achieved a mean accuracy of 0.61 (SD 0.06, 95% CI 0.57‐0.65), mean recall of 0.64 (SD 0.12, 95% CI 0.55‐0.73), and mean F₁-score of 0.63 (SD 0.07, 95% CI 0.58‐0.68), while the RNN, comprising 2 simpleRNN layers of 100 and 50 units followed by dense layers, yielded comparable metrics (accuracy: mean 0.61, SD 0.11, 95% CI 0.53‐0.69; recall: mean 0.70, SD 0.18, 95% CI 0.57‐0.83; and F₁-score: mean 0.65, SD 0.11, 95% CI 0.57‐0.73).

As shown in Figure 1, the SHAP interpretation of the RFC. The labels (1) through (5) correspond to the number of weeks preceding a decompensation event: (1) indicates the week immediately before the event, (2) two weeks prior, and so on, up to (5), which represents 5 weeks before decompensation. The SHAP values highlight the importance of use patterns in the model’s predictions. Specifically, higher patient use and continuous compliance during the week immediately prior to the event (1) appear to be strong indicators of respiratory decompensation. Lower compliance in weeks (3) and (4) also contributes to the model’s prediction, while higher compliance in weeks (1), (2), and (5) is similarly associated with increased risk. Elevated leakage values in the period leading up to the event also emerge as relevant features influencing the model’s output.

Although LR achieved the highest recall (mean 0.94, SD 0.06, 95% CI 0.90‐0.98), it showed lower precision (mean 0.58, SD 0.04, 95% CI 0.55‐0.61) and accuracy (mean 0.60, SD 0.05, 95% CI 0.56‐0.64), indicating a tendency to overpredict positive cases. This may be advantageous in clinical scenarios where identifying all potential positives is prioritized, but it comes at the cost of increased false positives. In contrast, the RFC, configured with 200 trees and a maximum depth of 5, provided a more balanced performance across all metrics, including the highest accuracy (mean 0.66, SD 0.10, 95% CI 0.59‐0.73) and solid values for recall (mean 0.78, SD 0.15, 95% CI 0.67‐0.89), precision (mean 0.66, SD 0.13, 95% CI 0.57‐0.75), and F₁-score (mean 0.70, SD 0.10, 95% CI 0.63‐0.77). This suggests that RFC may be more suitable for general-purpose prediction in this context, balancing sensitivity and specificity.

Other models such as DTC and KNN performed moderately, with KNN exhibiting high variability across folds, indicating sensitivity to the specific train-test splits. DTC and NB recorded the lowest accuracy among all models, while NB also showed the lowest F₁-score. This underperformance of NB suggests that its strong independence assumptions limit its ability to capture data patterns. Ensemble approaches like the ETC also yielded competitive results, reinforcing the robustness of tree-based methods. Both neural networks achieved intermediate performance; however, their capacity to learn more intricate patterns may have been limited by the relatively small sample size. Recurrent models generally require larger datasets and longer temporal sequences to effectively learn temporal dependencies. In this study, the relatively short input sequences and small cohort may have restricted the RNN’s capacity to model temporal dynamics robustly. Larger datasets, finer-grained temporal data, and extended temporal windows could allow the exploration of more advanced sequential architectures, such as attention-based models in future applications.

SHAP analysis from the RFC indicated that increased compliance, use, and leakage in the week preceding decompensation could be meaningful predictors. Although at first glance, it may seem counterintuitive that higher compliance is associated with increased risk, this pattern could reflect increased ventilator dependence in patients whose health is deteriorating. Patients may use the device more intensively as their condition worsens, while elevated leakage could indicate improper mask fitting under higher use. Furthermore, variations in global compliance patterns over time, such as a drop in compliance during periods (3) and (4), may be indicative of behavioral changes that precede decompensation.

While SHAP does not directly assess temporal trends or causality, these insights highlight the potential clinical relevance of use trends, suggesting that both increases and fluctuations in ventilator parameters could serve as early warning signals. Nonetheless, interpretability remains a challenge: models with moderate performance may limit the robustness of conclusions, and potential confounding factors that are not captured in the dataset—particularly the absence of detailed physiological variables—could reduce predictive power or lead to misattributions.

One of the primary limitations of this study is the absence of enough physiological explanatory variables. While many such variables are inherently unavailable from HMV data, respiratory rate, although a common monitored parameter, was not available in the reports used for the analysis. The absence of broader physiological or clinical variables (eg, comorbidities and vital signs) could have limited the models’ ability to account for the full clinical picture, reducing the predictive power.

Another notable constraint is the sample size. Although the dataset comprised 482 patients, and the strategy of constructing windows aimed to maximize the sample size, extracting windows with complete necessary information remained a challenge: generally, complete daily data across all 5 consecutive week reports could not be obtained consistently. For this reason, to address data sparsity, available daily measurements were averaged, as mentioned previously in the methodology. However, this approach could compromise temporal resolution, and due to the characteristics of the missing data, data augmentation techniques were insufficient to overcome this limitation.

Finally, the study is limited by its monocentric design, as all patients were sourced from a single center, which may impact the generalizability of the findings.

Although no studies in the literature directly use HMV parameters to predict respiratory decompensation [6], constructed an RF model that achieved an average accuracy of 0.69 (SD 0.10) for predicting ventilatory complications in critically ill intensive care unit patients, using invasive mechanical ventilation parameters. This performance is comparable to our model, which achieved an average accuracy of 0.66 (SD 0.10, 95% CI 0.59‐0.73) using RFC models.

This study introduced a novel framework for predicting respiratory decompensation using only HMV data. Ensemble tree models, particularly a random forest, achieved the best balance of 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), while SHAP analysis revealed that higher use, leakage, and compliance 7 days before, that is, period (1), could be an indicator of respiratory decompensation.

Despite the promising nature of these findings, overall performance remained moderate, reflecting the limitations imposed by incomplete longitudinal daily records and the lack of key physiological measurements. Combined with the relatively small sample size, these constraints position the present results as exploratory rather than ready for clinical application. Nevertheless, the findings suggest the potential of using NIMV data for early decompensation identification. Future research should focus on incorporating more clinical and contextual features and validating larger, multicenter cohorts to improve predictive power. With such advancements, this approach could translate into a clinically useful, noninvasive decision support tool for the proactive management of patients on home ventilation.