Insomnia Can Be Detected By Pulse Rate! A Proposed Predictive Model For Sleep Disorders Using Wristband Wearable Devices!
3 main points
✔️ Home sleep monitoring, especially using photoplethysmography (PPG)-photoplethysmography with a wearable device worn on the wrist, for screening for sleep disorders and health monitoring The technique has been attracting much attention
✔️ In this study, we aim to build a learning model for sleep stage classification from a large dataset of sleep-related data.
✔️ Evaluation results showed that the learning model performed better than the baseline model of PPG and ECG - an estimation model that utilized only ECG data or only PPG data
A deep transfer learning approach for wearable sleep stage classification with photoplethysmography
written by Mustafa Radha, Pedro Fonseca, Arnaud Moreau, Marco Ross, Andreas Cerny, Peter Anderer, Xi Long Ronald M. Aarts
(Submitted on 15 Sep 2021)
Comments: npj Digital Medicine
The images used in this article are from the paper, the introductory slides, or were created based on them.
background
Can wearable devices predict sleep state?
In this study, we aim to construct a learning model for classifying sleep states based on wearable devices, to improve sleep disorders.
Sleep deprivation has been linked to health problems - i.e. weight gain, systemic inflammation - and the importance of home sleep assessment and validation is increasing for the prevention of these diseases: long-term monitoring is effective for such sleep assessment, and the use of sleep polygraphic (PSG) validation, which is the gold standard for sleep assessment, can be more accurate. These measures allow for more accurate validation of polysomnography (PSG), the gold standard for sleep assessment; while these data sets can be used to identify different sleep stages (e.g., sleep onset, sleep onset, sleep onset, sleep onset, sleep onset, sleep onset, sleep onset, sleep onset, sleep onset, sleep onset, and sleep onset). On the other hand, it is difficult to distinguish between different sleep stages - rapid eye movement (REM) sleep, N1, N2, and N3 - from these data sets, and there is a need to establish a method to automatically detect these sleep states in order to improve home sleep assessment. In this context, previous studies have shown that heart rate variability (HRV) can function as a substitute for PSG - the autonomic nervous system is correlated with the progression of sleep stages, and therefore, measurement of HRV in autonomic HRV in autonomic nervous system correlates with the progression of sleep stages, and therefore, measurement of HRV in autonomic nervous system activity may enable the estimation of sleep stages. Therefore, attention has focused on ECG data and learning models that capture features with the aim of translating HRV features into sleep stages.
In this study, we aim to construct a sleep state classification model based on ECG data, trained using a time series analysis model, and introducing transfer learning based on sleep scoring indices: we use LSTM as the time series analysis model, train on a large ECG data set, and transfer learn on a small PPG We will also compare and validate our model with the base model based on these data.
What is a wearable device?
First, a brief description of wearable devices, which are the subject of this study's analysis, is given.
Wearable device is a generic term for information devices that subjects actually wear and use. Wearable means wearable or wearable, and there are wristband/wristwatch-type, eyeglass-type, and other types. Since these devices can actually be worn, they can be used in many aspects of daily life, from exercising to bathing and sleeping - they are thought to be effective in recording exercise such as jogging and swimming, and in acquiring health information by capturing heartbeat, pulse, sleep time, etc. It is thought to be effective in obtaining health information by capturing exercise records such as jogging and swimming, heart rate, pulse, and sleeping time.
technique
In this section, we give an overview of the proposed method in this study.
Dataset and proposed model
Two datasets were used in the evaluation: first, a large dataset of 292 participants with sleep scoring from ECG and PSG signals; second, a dataset of 60 participants with annotations based on PSG, PPG and AASM We perform transfer learning on this data set. In the proposed model, time series analysis based on LSTM-Long short-term memory-is used as a deep learning model for time series analysis.
The components of this model are: domain layer; Temporal layer; decision layer. The domain layer consists of a perceptron, which is used for weighting and selecting inputs and for compact coupling in the representation. This layer holds 32 vectors for each time step. This vector is generated at every time step - in the domain layer, the vector {𝑋'1, ..., 𝑋'𝑛} with |𝑋'|=32 vectors {𝑋'1, . , 𝑋'𝑛} sequences are generated; the Temporal layer consists of an LSTM stack - in this stack, the sequence {𝑋'1,... ,𝑋'𝑛} is taken and 128 features are generated at each time step. The decision layer consists of two levels of perceptrons: the first level performs dimensionality reduction on the output of the Tempral layer - reducing a vector of size 128 to size 32 The final layer generates four softmax outputs: the outputs are set to sum to 1 at each time step - these are generated at each time step, resulting in sleep stage probabilities corresponding to all epochs. probabilities are output. In order to suppress overlearning, dropout was applied in each training session. The categorical cross-entropy was used as the loss function in the proposed model.
In the transfer learning, the Temporal layer in the proposed model is introduced as a target of the transfer learning: this layer accounts for most of the model weights - 96.6% - and is likely to be difficult to train on a small data set. Therefore, by treating it as a transfer learning target, it can be trained well on smaller data sets; the domain and decision layers, on the other hand, are re-trained after transfer learning: the domain layer adjusts the input to adapt the ECG model, and the decision layer trains the mapping between the input and the mapping between inputs and labels. In these processes, the layers of the source model are frozen and trained and validated by four-fold cross-validation. In these three transfer learning models - domain model, decision model, and composite model - a name is defined by the layer to be trained.
result
In this section, we describe our evaluation of the proposed model.
The performance of the proposed method in the two base models - ECG and PPG base models - against which it is compared is described below: for ECG, the Kappa coefficient - the degree of agreement between the multiple evaluation models for the same subject - is 0.62 and the estimation accuracy is 74.83%; for PPG, the kappa is 0.57 and the accuracy is 71.88 - see the table below.
We also evaluated three learning models for transfer learning: learning on the domain layer, learning on the decision layer, and combined learning on both the domain and the decision layer. The results show that the best performance was obtained for the model that trained on both the domain and the decision layer - see table below.
The performance distributions for the three transfer learning conditions and the two baselines - see the figure below - show that the coupled learning with transfer learning performs best.
The performance of the combined approach was also compared based on F1 scores by sleep stage - see figure below.
As a result, the F1 score of the proposed model for sleep was 0.71, which was significantly higher than that of the ECG and PPG. Similarly, in REM sleep, N1/N2 and N3, the F1 score of the proposed model - domain + decision layer learning model - was 0.74, achieving higher performance than the baseline model.
consideration
In this study, based on sleep monitoring, we constructed and validated a transfer learning model of classification for sleep stages using PPG. In this approach, a deep neural network model - time series data analysis - is pre-trained using a large dataset of sleep data, and a part of the model is adapted to PPG. For pre-training, we utilized a dataset with available ECG-based HRV features. The evaluation results suggest that the three transition learning models - learning model for domain layer, learning model for decision layer, and learning model for domain layer and decision layer - perform better than the baseline model for PPG and ECG alone The results of this study suggest that the PPG-ECG baseline model is more efficient than the PPG-ECG baseline model. These results suggest that in the PPG dataset, model performance tended to be poorer for models built without prior learning - inferring that the PPG dataset is not sufficient.
The evaluation results of this study showed that the combined approach - learning model with domain layer + decision layer - achieved the best performance among the three transition learning models. The results suggest that it is necessary to include the domain layer in the learning model because, in related studies, only the soft-maximum/decision layer of the learning model was retrained - the other layers were not taken into account. By utilizing these results, it is expected that a monitoring system that can classify sleep states based on heart rate values at home will be realized.
On the other hand, challenges include: classification by sleep stage; assessment of sleep apnea; and assessment of non-health status. In the present study, we set up a 4-level classification assessment instead of the usual 5-level sleep stage: the reason is that inter-rater agreement is worst in N1 - the transitional period from wakefulness to sleep - in PSG-based assessments, and rater bias occurs. On the other hand, there are also reported cases of sleep disorders in falling asleep, so there may be cases in which these classifications are inappropriate; therefore, it is necessary to clarify the evaluation performance of the five-level classification in the proposed model. In addition, it is necessary to clarify the evaluation of sleep apnea syndrome, which is one of the most common sleep disorders. Sleep-disordered breathing has been shown to disrupt sleep architecture and is associated with various lifestyle-related diseases - i.e., angina pectoris and myocardial infarction; therefore, the sleep apnea syndrome, as measured by the apnea-hypopnea index - AHI Therefore, it is necessary to clarify the performance of algorithms for severity classification of sleep apnea as measured by the apnea-hypopnea index (AHI). In addition, the fact that the subjects analyzed in this study are healthy and do not have sleep disorders is also a possible issue: the proposed model's estimation performance for people with sleep disorders is unclear. Therefore, it is possible that the performance of the proposed model may be poor for such a case sample, and it is inferred that it is necessary to build a more robust model performance.
Categories related to this article
