Search

Study Objectives: This study aimed to (1) improve sleep staging accuracy through transfer learning (TL), to achieve or exceed human inter-expert agreement and (2) introduce a scorability model to assess the quality and trustworthiness of automated sleep staging., Methods: A deep neural network (base model) was trained on a large multi-site polysomnography (PSG) dataset from the United States. TL was used to calibrate the model to a reduced montage and limited samples from the Korean Genome and Epidemiology Study (KoGES) dataset. Model performance was compared to inter-expert reliability among three human experts. A scorability assessment was developed to predict the agreement between the model and human experts., Results: Initial sleep staging by the base model showed lower agreement with experts (κ = 0.55) compared to the inter-expert agreement (κ = 0.62). Calibration with 324 randomly sampled training cases matched expert agreement levels. Further targeted sampling improved performance, with models exceeding inter-expert agreement (κ = 0.70). The scorability assessment, combining biosignal quality and model confidence features, predicted model-expert agreement moderately well (R² = 0.42). Recordings with higher scorability scores demonstrated greater model-expert agreement than inter-expert agreement. Even with lower scorability scores, model performance was comparable to inter-expert agreement., Conclusions: Fine-tuning a pretrained neural network through targeted TL significantly enhances sleep staging performance for an atypical montage, achieving and surpassing human expert agreement levels. The introduction of a scorability assessment provides a robust measure of reliability, ensuring quality control and enhancing the practical application of the system before deployment. This approach marks an important advancement in automated sleep analysis, demonstrating the potential for AI to exceed human performance in clinical settings., (© The Author(s) 2024. Published by Oxford University Press on behalf of Sleep Research Society. All rights reserved. For commercial re-use, please contact reprints@oup.com for reprints and translation rights for reprints. All other permissions can be obtained through our RightsLink service via the Permissions link on the article page on our site—for further information please contact journals.permissions@oup.com.)

Background: Atrial fibrillation (AF) is often asymptomatic and thus under-observed. Given the high risks of stroke and heart failure among patients with AF, early prediction and effective management are crucial. Given the prevalence of obstructive sleep apnea among AF patients, electrocardiogram (ECG) analysis from polysomnography (PSG) offers a unique opportunity for early AF prediction. Our aim is to identify individuals at high risk of AF development from single‑lead ECGs during standard PSG., Methods: We analyzed 18,782 single‑lead ECG recordings from 13,609 subjects undergoing PSG at the Massachusetts General Hospital sleep laboratory. AF presence was identified using ICD-9/10 codes. The dataset included 15,913 recordings without AF history and 2054 recordings from patients diagnosed with AF between one month to fifteen years post-PSG. Data were partitioned into training, validation, and test cohorts ensuring that individual patients remained exclusive to each cohort. The test set was held out during the training process. We employed two different methods for feature extraction to build a final model for AF prediction: Extraction of hand-crafted ECG features and a deep learning method. For extraction of ECG-hand-crafted features, recordings were split into 30-s windows, and those with a signal quality index (SQI) below 0.95 were discarded. From each remaining window, 150 features were extracted from the time, frequency, time-frequency domains, and phase-space reconstructions of the ECG. A compilation of 12 statistical features summarized these window-specific features per recording, resulting in 1800 features (12 × 150). A pre-trained deep neural network from the PhysioNet Challenge 2021 was updated using transfer learning to discriminate recordings with and without AF. The model processed PSG ECGs in 16-s windows to generate AF probabilities, from which 13 statistical features were extracted. Combining 1800 features from feature extraction with 13 from the deep learning model, we performed a feature selection and subsequently trained a shallow neural network to predict future AF and evaluated its performance on the test cohort., Results: On the test set, our model exhibited sensitivity, specificity, and precision of 0.67, 0.81, and 0.3, respectively, for AF prediction. Survival analysis revealed a hazard ratio of 8.36 (p-value: 1.93 × 10 -52 ) for AF outcomes using the log-rank test., Conclusions: Our proposed ECG analysis method, utilizing overnight PSG data, shows promise in AF prediction despite modest precision, suggesting false positives. This approach could enable low-cost screening and proactive treatment for high-risk patients. Refinements, including additional physiological parameters, may reduce false positives, enhancing clinical utility and accuracy., Competing Interests: Declaration of competing interest None., (Copyright © 2024. Published by Elsevier Inc.)