No abstract

Objective: Portable oximetry has been shown to be a promising candidate for large-scale obstructive sleep apnea screening. In polysomnography (PSG), the gold standard OSA diagnosis test, the oxygen desaturation index (ODI) is usually computed from desaturation events occurring during sleep periods only, i.e. overnight desaturations occurring during or overlapping with a wake state are excluded. However, for unattended home oximetry, all desaturations are taken into account since no reference electroencephalogram is available for sleep staging. We aim to evaluate the hypothesis that the predictive power of oximetry for OSA screening is not impaired when reference sleep stages are not available. Approach: We used a PSG clinical database of 887 individuals from a representative São Paulo (Brazil) population sample. Using features derived from the oxygen saturation time series and demographic information, OxyDOSA, a published machine learning model, was trained to distinguish between non-OSA and OSA individuals using the ODI computed while including versus excluding overnight desaturations overlapping with a wake period, thus mimicking portable and PSG oximetry analyses, respectively. Main results: When excluding wake desaturations, the OxyDOSA model had an AUROC = 94.9 ± 1.6, Se = 85.9 ± 2.8, Sp = 90.1 ± 2.6 and F1 = 86.4 ± 2.7. When considering wake desaturations, the OxyDOSA model had an AUROC = 94.4 ± 1.6, Se = 88.0 ± 2.0, Sp = 87.7 ± 2.9 and F1 = 86.2 ± 2.4. Non-inferiority was demonstrated (p = 0.049) at a tolerance level of 3%. In addition, analysis of the desaturations excluded by PSG oximetry analysis suggests that up to 21% of the total number of desaturations might actually be related to apneas or hypopneas. Significance: This analysis of a large representative population sample provided strong evidence that the predictive power of oximetry for OSA screening using the OxyDOSA model is not impaired when reference sleep stages are not available. This finding motivates the usage of portable oximetry for OSA screening.

Objective: In this research, we introduce a new methodology for atrial fibrillation (AF) diagnosis during sleep in a large population sample at risk of sleep-disordered breathing. Approach: The approach leverages digital biomarkers and recent advances in machine learning (ML) for mass AF diagnosis from overnight-hours of single-channel electrocardiogram (ECG) recording. Four databases, totaling n = 3088 patients and p = 26 913 h of continuous single-channel electrocardiogram raw data were used. Three of the databases (n = 125, p = 2513) were used for training a ML model in recognizing AF events from beat-to-beat time series. Visit 1 of the sleep heart health study database (SHHS1, n = 2963, p = 24 400) was used as the test set to evaluate the feasibility of identifying prominent AF from polysomnographic recordings. By combining AF diagnosis history and a cardiologist’s visual inspection of individuals suspected of having AF (n = 118), a total of 70 patients were diagnosed with prominent AF in SHHS1. Main results: Model prediction on SHHS1 showed an overall S e = 0.97 , S p = 0.99 , NPV = 0.99 and PPV = 0.67 in classifying individuals with or without prominent AF. PPV was non-inferior (p = 0.03) for individuals with an apnea-hypopnea index (AHI) ≥15 versus AHI < 15 . Over 22% of correctly identified prominent AF rhythm cases were not previously documented as AF in SHHS1. Significance: Individuals with prominent AF can be automatically diagnosed from an overnight single-channel ECG recording, with an accuracy unaffected by the presence of moderate-to-severe obstructive sleep apnea. This approach enables identifying a large proportion of AF individuals that were otherwise missed by regular care.

Background: The 12-lead electrocardiogram (ECG) is a standard tool used in medical practice for identifying cardiac abnormalities. The 2020 PhysioNet/Computing in Cardiology Challenge addresses the topic of automated classification of 12-lead ECG. Methods: Two machine learning strategies were implemented: a feature engineering approach based on the engineering of physiological features (or "digital biomarkers") and a deep learning approach. Two sets of features were engineered: (1) capturing the interval variation between consecutive heartbeats, commonly called heart rate variability (HRV) measures and (2) using morphological biomarkers (e.g. QT interval, QRS width). A total of 16 HRV and 97 morphological biomarkers were implemented in python for each lead. A random forest (RF) model was trained using 5-fold cross validation to optimize the model hyperparameters. For the deep learning approach, a residual neural network (ResNet) architecture was used. The RF and ResNet were also combined in an ensemble learning (EL). The dataset was divided into 80%-20% stratified training-test sets. Results: on the local test set we achieved a Challenge score of 0.65 using the FE approach, 0.52 using the DL approach and 0.66 using the EL approach. For technical reasons we did not manage to score our models on the Challenge hidden test set.

No abstract