We used sleep monitoring data from a study that investigated the prevalence, characteristics, risk factors and type of sleep apnea (SA) in 579 patients with paroxysmal atrial fibrillation. Most patients were screened for two nights, resulting in 1,043 sleep recordings that each contained data from one night. SA was diagnosed using the Nox T3 portable sleep monitor. An experienced sleep specialist scored the recordings manually using Noxturnal software. A total of 157 women (27%) and 422 men (73%) were examined; 477 (82.7%) had an apnea–hypopnea index (AHI) ≥ 5/hr, whereas moderate to severe SA (AHI ≥ 15/hr) was diagnosed in 243 patients (42.1%). The AHI derived from automatic and manual scoring showed a good agreement (Pearson's r coefficient of 0.96). The median difference in AHI was very small (i.e., 0.72 [mean difference, 1.06]), but was statistically significant (p < .0001). Automatic scoring classified sleep recordings with more than 90% accuracy into SA categories of mild (AHI ≥ 5/hr), moderate (AHI ≥ 15/hr) and severe (AHI ≥ 30/hr). We found a minor (11%–21%) mis‐estimation of the number of recordings right above and below the boundary separating mild and moderate SA. The accuracy of automatic scoring differed from recording to recording, especially regarding the sensitivity of detecting disrupted breathing events. We found low to moderate agreement for the duration of disrupted breathing events (r = .53), for which the automatic scoring led to a statistically significant overestimation by 5.22 s (p < .0001).

Obstructive Sleep Apnea (OSA) is a common, but severely under-diagnosed sleep disorder that affects the natural breathing cycle during sleep with periods of reduced respiration or no airflow at all. It is our long-term goal to increase the percentage of diagnosed OSA cases and reduce the time to diagnosis with user friendly and cost-efficient tools for sleep analysis at home. As a first step towards this goal, we study in this work whether a small subset of those physiological signals that are used in classical OSA diagnosis in combination with automatic classification allow to detect apnea events. We study the performance of five data mining techniques to classify epochs of data from the Apnea-ECG and MIT-BIH databases from PhysioNet as either disrupted or normal breathing. The data is only slightly preprocessed (rate reduction and normalization). We focus in this study on respiratory signals from the nose, abdomen, chest, and oxygen saturation. We measure for any combination of these signals the accuracy, sensitivity, specificity and Kappa statistics of classification with Artificial Neural Network, Support Vector Machine, Decision Tree, K-Nearest Neighbor and Random Forest. For Apnea-ECG, we achieve an accuracy of 96.6% with a combination of respiration data from the chest and nose as input data; and an accuracy of more than 90% for all signal combinations. Interestingly, these good results are also achieved with the simple KNN technique. The results for MIT-BIH are lower, because of noise, smaller size, and some class imbalance. The accuracy does not significantly improve with the number of signals included in the signal combinations.We conclude that one signal might be sufficient to detect disrupted breathing, if the data set is of sufficient quality and size, and that respiration from the abdomen is the preferrable choice when taking into account both classification performance and patient comfort.

Improper or erroneous labelling can pose a hindrance to reliable generalization for supervised learning. This can have negative consequences, especially for critical fields such as healthcare. We propose an effective new approach for learning under extreme label noise for medical applications like sleep apnea, that is based on under-trained deep ensembles. Each ensemble member is trained with a subset of the training data, to acquire a general overview of the decision boundary separation, without focusing on potentially erroneous details. The accumulated knowledge of the ensemble is combined to form new labels, that determine a better class separation than the original labels. A new model is trained with these labels to generalize reliably despite the label noise. We evaluate our approach on the tasks of sleep apnea detection and sleep apnea severity classification, and observe performance improvement in kappa from 0.02 up-to 0.55.