Objective: Epilepsy prediction algorithms offer patients with drug-resistant epilepsy a way to reduce unintended harm from sudden seizures. The purpose of this study is to investigate the applicability of transfer learning (TL) technique and model inputs for different deep learning (DL) model structures, which may provide a reference for researchers to design algorithms. Moreover, we also attempt to provide a novel and precise Transformer-based algorithm. Methods: Two classical feature engineering methods and the proposed method which consists of various EEG rhythms are explored, then a hybrid Transformer model is designed to evaluate the advantages over pure convolutional neural networks (CNN)-based models. Finally, the performances of two model structures are analyzed utilizing patient-independent approach and two TL strategies. Results: We tested our method on the CHB-MIT scalp EEG database, the results showed that our feature engineering method gains a significant improvement in model performance and is more suitable for Transformer-based model. In addition, the performance improvement of Transformer-based model utilizing fine-tuning strategies is more robust than that of pure CNN-based model, and our model achieved an optimal sensitivity of 91.7% with false positive rate (FPR) of 0.00/h. Conclusion: Our epilepsy prediction method achieves excellent performance and demonstrates its advantage over pure CNN-based structure in TL. Moreover, we find that the information contained in the gamma (γ) rhythm is helpful for epilepsy prediction. Significance: We propose a precise hybrid Transformer model for epilepsy prediction. The applicability of TL and model inputs is also explored for customizing personalized models in clinical application scenarios.
11 August 2023 in IEEE Journal of Biomedical and Health Informatics
Objective: Obstructive sleep apnea (OSA) is a common sleep-related breathing disorder that can lead to a wide range of health issues if left untreated. This study aims to address the lack of research on personalized models for single-lead electrocardiogram (ECG)-based OSA detection, by proposing an automatic semi-supervised algorithm for automated low-cost personalization fine-tuning. Methods: We utilize a convolutional neural network (CNN)-based auto-encoder (AE) with a modified training objective to detect anomalous region of OSA. An indicator based on model outputs is utilized as a benchmark measure to assign pseudo-labels with confidence to each sample. Finally, we perform validation of the semi-supervised algorithm on the same database and cross-database scenarios. Results: By introducing semi-supervised personalization, the accuracy, AUC, and mean absolute error (MAE) of the general model (GM) of 35 subjects from the same database are improved from 86.3%, 0.915, and 5.178 to 90.3%, 0.948, and 2.593. Simultaneously, in the validation of 25 subjects from a cross-database, the accuracy, AUC, and MAE of the GM are enhanced from 75.6%, 0.800, and 9.149 to 84.3%, 0.881, and 3.509. Conclusion: The improved version of AE demonstrates excellent adaptability in identifying abnormal features in OSA, employing a data-driven approach to assign pseudo-labels for unknown data automatically. Additionally, leveraging the pseudo-labels through a semi-supervised fine-tuning strategy provides a solution to overcome the limitation of clinical annotations, facilitating low-cost implementation of personalized models. Significance: The semi-supervised approach proposed in this paper provides a high-performance and annotation-free solution for personalized adjustment of automatic OSA detection.
Electrocardiogram (ECG) is an efficient and commonly used tool for detecting arrhythmias. With the development of dynamic ECG monitoring, an effective and simple algorithm is needed to deal with large quantities of ECG data. In this study, we proposed a method to detect multiple arrhythmias based on time-frequency analysis and convolutional neural networks. For a short-time (10 s) single-lead ECG signal, the time-frequency distribution matrix of the signal was first obtained using a time-frequency transform method, and then a convolutional neural network was used to discriminate the rhythm of the signal. ECG data in multiple databases were used and were divided into 12 classes. Finally, the performance of three kinds of time-frequency transform methods are evaluated, including short-time Fourier transform (STFT), continuous wavelet transform (CWT), and pseudo Wigner-Ville distribution (PWVD). The best result was obtained by STFT, with an accuracy of 96.65%, an average sensitivity of 96.47%, an average specificity of 99.68%, and an average F1 score of 96.27%, respectively. Especially, the area under curve (AUC) value is 0.9987. The proposed method in this work may be efficient and valuable to detect multiple arrhythmias for dynamic ECG monitoring.
Early detection of widespread undiagnosed sleep apnea is crucial for preventing its severe health complications. However, large-scale diagnosis faces inaccessible monitoring and trust barriers in automated analysis, particularly due to the absence of transparent artificial intelligence frameworks capable of monitoring adaptation. Here, we develop Apnea Interact Xplainer, a transparent system enabling sleep apnea diagnosis through flexible channel analysis across clinical and home settings. Analyzing 15,807 polysomnography recordings from seven independent multi-ethnic cohorts, our system achieves accuracies of 0.738-0.810 for four-level severity classification, with 99.8% accuracy within one severity grade and R-squared of 0.92-0.96 for apnea-hypopnea index prediction on external test cohorts. The system provides multi-level expert-logic interpretable visualization of respiratory patterns enabling transparent collaborative decision-making. Notably, it achieves a sensitivity of 0.970 for early sleep apnea detection using only oximetry signals, while providing nightly risk assessment and intelligent monitoring reports. This study establishes a paradigm shift in advancing early and cost-effective sleep apnea diagnosis through transparent artificial intelligence.
7 September 2023 in IEEE Transactions on Instrumentation and Measurement
Objective: The automatic sleep apnea (SA) detection based on deep learning (DL) and single-lead electrocardiogram (ECG) has been extensively studied. We aim to explore the impact of different DL model structures and label mapping length (LML) on personalized transfer learning (TL), providing personalized TL applicability conditions based on proposed hybrid Transformer model. Methods: Two DL models, a pure convolutional neural network (CNN)-based model (PCM) and proposed hybrid Transformer-based model (HTM), are included in the study. Eight different LMLs are considered. Furthermore, various personalized TL strategies are introduced to thoroughly explore the impact. Finally, two-sided t-tests are utilized to evaluate the significance. Results: In the same database, the average accuracy and AUC for the personalized PCM are 0.8412 and 0.9002 respectively with p<0.001, while the hybrid Transformer-based personalized model achieves an average accuracy of 0.8537 and an average AUC of 0.9147 with p<0.001. Across databases, the accuracy and AUC of personalized HTM reach 0.8271 and 0.8724 respectively, and p<0.001. Conclusion: The increase of LML has a beneficial impact in general model (GM) and personalized model for different model structures. The HTM exhibits better performance in both GM and personalized TL compared to the PCM. Additionally, personalized TL achieves significant improvement when utilizing only positive samples within the same database. However, it is more advantageous to utilize only negative samples when performing cross-database personalized TL. Significance: This paper provides guidance for personalized TL in SA detection and provide a HTM-based personalized TL method.