0512 ABED: Automatic Sleep-Disordered Breathing Event Detection

Introduction Sleep disordered breathing (SDB) affects more than 10% of the global population and is associated with cardiovascular diseases. Manual annotation of SDB events is time consuming and suffers from inter-rater variances and biases. We propose Automatic Sleep-Disorder Breathing Event Detection (ABED), a method capable of localizing and classifying obstructive apneas (OA), central apneas (CA), and hypopneas (HYPO). Methods ABED was trained on 5,456 PSGs and evaluated on 1,099 PSGs from 4 different cohorts. The respiratory modalities included measures of nasal flow, oral thermistor, thoracoabdominal plethysmography, and blood oxygen saturation. These were passed directly as input for ABED. EEG and EOG were passed through another automated detector to ascertain arousals and wake probabilities, which was also passed as input. ABED uses an end-to-end deep learning architecture based on state-of-the-art object detection methods Unified-Real-Time-Object-Detection, ResNet for feature extraction and bi-directional long-short-term-memory layers for temporal processing. ABED was trained on 240 second windows and apnea classes were balanced by sampling less prevalent classes (OA, CA) multiple times during training. ABED was optimized based on event class (cross-entropy loss) and position (Huber loss) weighted towards the class. At the point of inference, non-maximum suppression was used to remove overlapping predictions. Results Results show that the apnea-hypopnea index (AHI) values based on ABED’s predictions correlate well with experts (R^2=0.86;β_0=0.95,β_1=3.29 ) and predict sleep apnea diagnosis (AHI≥15) and severity (no, mild, moderate, severe) with a mean accuracy of 89% and 76%, respectively. ABED was able to predict the central apnea index (CAI) with a high correlation (R^2=0.82;β_0=1.16,β_1=0.51 ). When inferring individual events, ABED achieved an overall F1 score of 0.78 in distinguishing between the presence of an event or regular breathing. For specific event types, the F1 scores were: 0.72 OA, 0.51 for CA, and 0.66 for HYPO. Conclusion ABED was trained and tested on a multi-cohort dataset and demonstrates generalizing properties on intra-cohort variability along with consistent scoring, making ABED suitable as support for medical experts in SDB annotation. Overall apnea event detection performance demonstrates that ABED can discern between apneic events and regular breathing. Additional work will involve integration of additional cohorts. Support (if any)