Sociodemographic, anthropometric, clinical, polysomnographic, and CPAP usage data from the OSA therapy telemonitoring program managed by Oxigen Salud will be used. Access, storage, and subsequent analysis of these variables require addressing the management of a high volume of data acquired daily from different sources and in various formats, as well as verifying them in terms of integrity and accuracy. Additionally, since the number of OSA patients is constantly growing, it must be scalable. In accordance with these requirements, the first stage of this project will focus on a thorough study of the state of the art in big data systems applied to health problems. This will include the selection of appropriate tools and technologies for data storage, access, and management, as well as verification.

In this stage, complementary approaches will be applied: unsupervised vs. supervised learning; conventional learning (ML) vs. deep learning (DL):

• Unsupervised learning: OSA subgroups will be identified that share similarities within the data used, helping to discover whether patients with OSA of the same age range, comorbidities, sex, etc., follow similar patterns. A post-hoc prospective analysis of these groups will allow us to assess their adherence behaviors and evaluate their shared information to identify new factors affecting adherence.
• Supervised learning: classification and regression methods based on ML and DL will be used. The former will be used to classify subjects as "adherent" or "non-adherent", while the latter will automatically estimate the number of hours of CPAP use. Additionally, independent models will be created to predict adherence in the short term (3 months), medium term (6 months), and long term (12 months).

ML and DL models have traditionally been perceived as "black boxes". This perception is particularly important in the context of healthcare, as it reduces clinicians' trust in their predictions, negatively affecting the real-world implementation of these tools. To address this issue, different XAI methods will be implemented and applied. These techniques have been designed and evaluated in several fields of application, facilitating the interpretation of predictions and the identification of new laws or hidden patterns in the data. Their use at this stage of the project will allow us to explain and interpret the ML and DL models, as well as better understand why such predictions/results are provided.

For binary classification (adherence vs. non-adherence), the performance of the ML and DL models will be evaluated in terms of: sensitivity, specificity, accuracy, positive predictive value, negative predictive value, and area under the ROC curve.

For regression approach (estimating the number of hours of CPAP usage), the performance of the models will be evaluated in terms of concordance: intra-class correlation coefficient, Bland-Altman plots, and Cohen's kappa.

All statistical metrics will be derived from the prospective test dataset to ensure proper validation of the methodology.

A cost-effectiveness analysis will be conducted to evaluate the feasibility of the automatic early prediction model of adherence to treatment in patients with OSA. First, we will analyze the estimated costs of non-adherence to CPAP based on cutting-edge studies. Then, we will evaluate the costs associated with the implementation of our proposal, as well as the estimated number of patients who will benefit from it. Finally, we will estimate the cost savings and profit margin for the real-world deployment of the proposal, and consider potential actions to protect the project results.