PHENOXLEEP

Executive summary

The problem

Obstructive sleep apnea (OSA) is a chronic disease characterized by recurrent episodes of absence or reduction of airflow during sleep, causing transient hypoxemia and arousals. This condition is associated with increased sympathetic activity, repeated oxygen desaturations, and sleep fragmentation, leading to cardiovascular, metabolic, and neurocognitive problems. OSA has a high prevalence, being nocturnal polysomnography (PSG) the gold standard test for diagnosis. This test monitors multiple signals to estimate the apnea-hypopnea index (AHI), which determines the presence and severity of OSA.

However, OSA is a heterogeneous disorder with diverse predisposing factors, pathophysiological mechanisms, clinical presentations, and health outcomes. Patients with a similar AHI may have very different underlying causes, symptoms, and prognoses. This suggests the existence of different phenotypes of OSA, which are more homogeneous subtypes of the disease. Identifying these phenotypes is crucial for developing personalized diagnostic and treatment strategies.

OSA affects 13%-33% of adult men and 6%-19% of women worldwide. However, OSA is an underdiagnosed disease, with estimates suggesting that the actual number of patients could double

Hypothesis

Previous studies suggest the presence of different OSA phenotypes related to comorbidities, which in turn may explain the relatively high proportion of OSA cases that do not respond, or only partially respond, to OSA treatment. Despite this preliminary evidence using machine learning techniques, the delineation of these phenotypes remains challenging, probably due to the limitations associated with conventional feature-engineering methods.

Emerging approaches in multitask deep learning combined with new techniques in Explainable Artificial Intelligence (XAI), applied to the information collected during the PSG procedure, can delineate new OSA phenotypes linked to various comorbidities and variations in treatment effectiveness.

Can new OSA phenotypes related to different comorbidities and variations in treatment effectiveness be defined using multitask deep learning approaches and XAI techniques applied to information obtained from PSG?

Main objective

The main objective of the project is to design, evaluate, and interpret deep learning-based solutions to accurately define new clinically relevant OSA phenotypes from information collected during PSG. It is proposed to use multitask strategies to train algorithms capable of detecting cardiovascular, cognitive, and sleepiness-related morbidities in OSA patients. In addition, XAI methods will be implemented to identify physiological patterns related to these morbidities, thus enabling the definition of phenotypes.

Specific objectives

To design, develop, and optimize high-performance models to detect OSA and its comorbidities based on DL approaches and new MTL strategies

Initially, the following DL algorithms will be assessed: (i) CNNs, including residual connections and attention mechanisms; (ii) RNNs; (iii) Transformers; and (iv) a combination of CNNs with RNNs and transformers, based on pre-trained CNNs. Furthermore, the following MTL architectures and optimization strategies will be initially explored: (i) shared-trunk; (ii) cross-stitch; (iii) Prediction distillation; (iv) soft parameter sharing; and (v) loss weighting.

To detect new patterns/attributes inherent to PSG (signals, clinical, and sociodemographic data) linked to OSA co-morbidities

XAI algorithms enable the identification of paramount information, such as patterns within a raw signal, that elucidates predictions, particularly those derived from DL methods. As an initial proposal, we will use the following approaches: (i) SHAP; (ii) Grad-CAM.

To identify OSA phenotypes at hig risk for cardiovascular, cognitive, and/or hypersomnolence morbidity

Utilizing the inputs (e.g., signals) relevance derived from XAI for each OSA co-morbidity, we will delineate phenotypes through a supervised approach. As a starting point, we will employ SHAP clustering.

To assess treatment effectiveness of sleepiness and neurocognitive-related phenotypes

Taking advantage of the PSGs and the information from the APPLES database, we will evaluate the long-term effectiveness of CPAP therapy on sleepiness and neurocognitive function for each phenotype. For this, statistical correlation analyses will be used over the longitudinal data.

Responsible team

Project tasks

Organization of polysomnographic databases
Bibliographic review of automatic methods
Implementation of selected automatic methods
Automatic OSA phenotype definitions
Statistical comparison of methods, discussion and drawing of conclusions
Dissemination and transfer of results
Management and coordination

Checking the quality of 18,703 AF, SpO2, PRV, ECG/HRV, and EEG signals from the Apnea Positive Pressure Long-term Efficacy Study (APPLES), Cleveland Family Study (CFS), Multi-Ethnic Study of Atherosclerosis (MESA), MrOS Sleep Study (MrOS), Sleep Heart Health Study (SHHS), Wisconsin Sleep Cohort (WSC) databases. Analysis of sociodemographic and clinical data to ensure proper merging of these datasets, detect possible biases, relate clinical variables with information extracted from the recordings, and/or detect the presence of comorbidities in patients. Design of validation strategies based on the morbidities of the subjects.

Compilation and analysis of the latest scientific studies related to DL, MTL, XAI, and clustering. Method comparison and study for project application.

Implementation of the selected DL methods to OSA detection. Implementation of MTL methods that will subsequently be applied to jointly identify OSA and its morbidities. Implementation of XAI and clustering procedures to identify physiological patterns and OSA phenotypes, respectively. All methods are implemented using Python tools.

Training and validation of DL models to individually detect OSA from AF, SpO2, PRV, ECG/HRV, and EEG signals. Application of DL models using MTL as a training paradigm to jointly identify OSA and its morbidities. Study of the set of physiological patterns via XAI that potentially contribute to each OSA-related comorbidity. Definitions of OSA phenotypes by applying supervised clustering methods.

Study and application of the most appropriate DL and MTL approaches to address the joint detection of OSA and its morbidities. Discussion on the results to characterize the physiological patterns linked to OSA morbidities. Gaining insights on the heterogeneity of OSA by defining new phenotypes linked to cardiovascular, cognitive, and sleepiness-risk morbidities. Phenotype treatment effectiveness evaluation.

Dissemination of preliminary results in conferences. Publication of at least 18 papers on the final results in prestigious international journals (Q1 or Q2). Preparation of dissemination material: website, social networks, organization of seminars, awards, and press releases. Preparation of reports on the evolution of the project for delivery to the Observing Promoter entities (OPs): Oxigen Salud S.A., Five Flames Mobile S.L.L., Centro Regional de Servicios Avanzados S.A., Fundación CARTIF, Hospital Universitario de Jaén, and Hospital Universitario Río Hortega. Biannual meetings to discuss the results. Preparation of reports on the development of potential patents.

Coordination and control of each of the tasks and subtasks.

Publications

Articles in indexed journals

Enrique Gurdiel, Fernando Vaquerizo-Villar, Javier Gomez-Pilar, Gonzalo C. Gutiérrez-Tobal, Félix del Campo, and Roberto Hornero, Beyond the Ground Truth, XGBoost Model Applied to Sleep Spindle Event Detetion”, IEEE Journal of Biomedical and Health Informatics, vol. 29, NO. 7, July, 2025, DOI: 10.1109/JBHI.2025.3544966

Fernando Vaquerizo-Villar, and Verónica Barroso-García, Special Issue: “Artificial Intelligence for Biomedical Signal Processing”, Bioengineering, 2025, 12, 753, DOI: 10.3390/bioengineering12070753

Adrián Martín-Montero, Fernando Vaquerizo-Villar, Clara García-Vicente, Gonzalo C. Gutiérrez-Tobal, Thomas Penzel, Roberto Hornero, Heart rate variability analysis in comorbid insomnia and sleep apnea (COMISA), Scientific Reports, vol. 15 (17574), May, 2025, DOI: 10.1038/s41598-025-02541-7

Teresa Arora, Fernando Vaquerizo-Villar, Roberto Hornero, David Gozal, Sleep irregularity is associatedwith night-time technology,dysfunctional sleep beliefs andsubjective sleep parametersamongst female universitystudents, Scientific Reports, 15, 1, 12, February, 2025, DOI: 10.1038/s41598-025-90720-x

Communications in international conferences

Javier Gomez-Pilar, Adrián Martín-Montero, Fernando Vaquerizo-Villar, Máximo Domínguez-Gerrero, Daniela Ferreira-Santos, Pedro Pereira-Rodrigues, David Gozal, Roberto Hornero, Gonzalo C. Gutiérrez-Tobal, Phenotypic Characterization of Sleep Apnea Using Clusters Derived from Subject-Based SpO2 Weighted Correlation Networks, 47th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC 2025), ISBN: 979-8-3315-8618-8, Copenhague (Denmark), July 14 - July 17, 2025

Clara García-Vicente, Gonzalo C. Gutiérrez-Tobal, Javier Gomez-Pilar, Fernando Vaquerizo-Villar, Adrián Martín-Montero, Máximo Domínguez-Gerrero, David Gozal, Roberto Hornero, Explainable Hybrid Convolutional and Transformer Network for Pediatric Sleep Apnea Diagnosis Using Nocturnal Oximetry, 47th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC 2025), ISBN: 979-8-3315-8618-8, Copenhague (Denmark), July 14 - July 17, 2025

Communications in national conferences

Related news

Back

Contact

Name*

Last name*

Email*

Phone number*

Your message*

Biomedical Engineering Group

Roberto Hornero Sánchez
Gonzalo César Guitierrez Tobal
Paseo Belén 15 - CP. 47011 Valladolid, España
Tlfn +34 983 18 5570