| 研究実績の概要 |
This study aims to develop novel algorithms for SAHS screening with signals that can be collected using consumer wearable devices. We made a major break through this year as we obtained approvals to access two large sleep datasets, i.e., the OSASUD (N=30) and SHHS (N=5768) datasets. We first applied contrast set mining on the OSASUD dataset and found that heart rate and SpO2 were strongly associated with the occurrence of apnea events. We then utilized the SHHS dataset to develop apnea screening models with overnight SpO2 signals and demographic information. We proposed a novel method combining multiscale attention entropy analysis, ICA, and machine learning. Our models achieved better performance than existing ones, with the best MCC, AUC, sensitivity, specificity being 0.545, 0.823, 97.3%, and 97.2%. We also examined the effect of data resolution on model performance and found that the models could achieve good accuracy even when the resolution was as low as 1 minute. This is an important finding because consumer wearables are not able to provide high-resolution data as clinical devices do, and being able to achieve good performance at low data resolution means our models are compatible with consumer devices. Our method is statistically rigid, validated on the largest sleep dataset, and works even for asymptotic patients.
Based on the results, we made 6 publications (2 journals, 1 international conference, 2 domestic conference) and gave 1 invited talk at an international conference. We have another 2 journal and 3 conference submissions currently under review.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
1: 当初の計画以上に進展している
理由
The acquisition of two large sleep datasets shared by two overseas research institutions significantly speeded up our progress. We were able to dive directly into the model construction phase after retrieving and cleaning the two datasets, while skipping the initial data collection experiment which would have been needed should the two datasets were not available. With an epoch-wise approach, the OSASUD dataset allowed us to thoroughly explore the differences in physiological signals between sleep apnea and non-apnea states, and we were able to identify the set of signals that have strong discriminating power and are easy to measure at home using off-the-shelf wearable devices. The SHHS dataset consists of sleep data collected from more than 5000 people and is by far the largest dataset available. The SHHS dataset enabled us to rigidly develop and validate our sleep apnea screening models. By combining chaos analysis technique and machine learning, our models were able to achieve better performance than existing models, compatible with the potentially low-resolution data of consumer wearables, and could potentially be applied to more usage scenarios (e.g., for asymptotic patients).
|
| 今後の研究の推進方策 |
In the next step, we aim to further improve the model performance. We are considering several potential ways to achieve this. First, we will increase the variety of features in the time-domain, frequency-domain, as well as deriving non-linear features using other chaos analysis methods. We will perform another round of data mining on the feature set to identify the most important features that help distinguish apnea and non-apnea populations. Second, we will investigate the effectiveness of ensemble learning (e.g., voting, stacking, boosting) and deep learning (e.g., LSTM) with a more comprehensive cross-validation process. Third, we will examine the effect of feature extraction scale and data resolution. Fourth, we will benchmark the performance of our models and existing ones on the same datasets to establish a much-needed reference point for future research. Finally, we will conduct a small-scale clinical trial to validate our models off-the-shelf wearable devices.
|
