0484 Heart Rate Variability and Deep Learning Analysis of Obstructive Sleep Apnea Using ECG from Polysomnography
Kiyong Kim,Tae-Won Yang,Ji Yoon Kim,Woo Ri Choi,Kyung Won Kwon,So Young Lee,Seung Hwan Kim,Byeong Gu Jeon,Nak Gyeong Ko,Young-Soo Kim,Oh-Young Kwon,Do-Hyung Kim
DOI: https://doi.org/10.1093/sleep/zsae067.0484
IF: 6.313
2024-04-20
SLEEP
Abstract:Abstract Introduction Previous studies suggested that obstructive sleep apnea (OSA) can affect the autonomic nervous system. Patients with OSA appear to have a higher sympathetic component, a lower parasympathetic component, and greater autonomic nervous system (ANS) imbalance. We compared heart rate variability (HRV) with existing studies and confirmed classification accuracy through deep learning analysis (DLA), using electrocardiogram (ECG) data extracted from polysomnography (PSG). Methods We retrospectively surveyed people who underwent PSG at our hospital from January 2015 to March 2023. The diagnosis of OSA was classified into normal, mild, moderate, and severe based on AHI, and whether arrhythmia was identified during the test was also investigated. HRV analysis performed by frequency domain analysis of the tachogram. For DLA, the tachogram was converted to a Mel-spectrogram and a Convolutional Neuronal Network (CNN) was used to confirm the confusion matrix. Results Of a total of 1,806 PSG, 1,554 cases were selected, excluding 252 cases of arrhythmia. OSA confirmed by PSG was normal in 282 patients, mild in 334, moderate in 293, and severe in 645. When comparing the results of HRV divided into AHI below 15 and above, VLF power (ms2/Hz) was 940.78 ± 763.72 vs 1132.75 ± 1104.50 (p < 0.001), LF power (ms2/Hz) was 719.26 ± 734.71 vs. 724.46 ± 945.26 (p = 0.908), HF power (ms2/Hz) was 763.61 ± 1058.92 vs 595.53 ± 1386.75 (p = 0.011), and LF/HF ratio was 1.27 ± 0.74 vs 1.63 ± 1.02 (p < 0.001). As a result of DLA, the ROC AUC Score was confirmed to be 0.7077 and the F1 Score was 0.67. Conclusion As a result of HRV using ECG from PSG, OSA patients were found to have low HF power and high LF/HF ratio, similar to previous studies. Additionally, if tachogram's DLA accuracy can be improved through preprocessing and deep learning model improvements, it is expected that it can be used as a screening tool in various place. Support (if any) This work was partly supported by Institute of Information & Communications Technology Planning & Evaluation grant funded by the Korea government No.RS_2023_00227552, Development of artificial intelligence video background removal SaaS service using domestic semiconductor 64 TOPS.
neurosciences,clinical neurology