Next Article in Journal
Critically Ill COVID-19 Patients Show Reduced Point of Care-Measured Butyrylcholinesterase Activity—A Prospective, Monocentric Observational Study
Next Article in Special Issue
Reading Wishes from the Lips: Cancer Patients’ Need for Psycho-Oncological Support during Inpatient and Outpatient Treatment
Previous Article in Journal
T Cell and Antibody Response Following Double Dose of BNT162b2 mRNA Vaccine in SARS-CoV-2 Naïve Heart Transplant Recipients
Previous Article in Special Issue
Deep Convolutional Recurrent Model for Automatic Scoring Sleep Stages Based on Single-Lead ECG Signal
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Deep Learning for Automatic Detection of Periodic Limb Movement Disorder Based on Electrocardiogram Signals

1
Artificial Intelligence Big Data Medical Center, Wonju College of Medicine, Yonsei University, Wonju 26426, Korea
2
Department of Medical Artificial Intelligence, College of Medical Engineering, Konyang University, Daejeon 35365, Korea
3
Department of Emergency Medicine, Wonju College of Medicine, Yonsei University, Wonju 26426, Korea
4
Department of Preventive Medicine, Wonju College of Medicine, Yonsei University, Wonju 26426, Korea
5
Department of Biomedical Engineering, College of Software and Digital Healthcare Convergence, Yonsei University, Wonju 26493, Korea
*
Author to whom correspondence should be addressed.
These authors contributed equally to this work.
Diagnostics 2022, 12(9), 2149; https://doi.org/10.3390/diagnostics12092149
Submission received: 10 July 2022 / Revised: 29 August 2022 / Accepted: 2 September 2022 / Published: 3 September 2022
(This article belongs to the Special Issue Machine Learning and Big Data in Psychiatric and Sleep Disorders)

Abstract

:
In this study, a deep learning model (deepPLM) is shown to automatically detect periodic limb movement syndrome (PLMS) based on electrocardiogram (ECG) signals. The designed deepPLM model consists of four 1D convolutional layers, two long short-term memory units, and a fully connected layer. The Osteoporotic Fractures in Men sleep (MrOS) study dataset was used to construct the model, including training, validating, and testing the model. A single-lead ECG signal of the polysomnographic recording was used for each of the 52 subjects (26 controls and 26 patients) in the MrOS dataset. The ECG signal was normalized and segmented (10 s duration), and it was divided into a training set (66,560 episodes), a validation set (16,640 episodes), and a test set (20,800 episodes). The performance evaluation of the deepPLM model resulted in an F1-score of 92.0%, a precision score of 90.0%, and a recall score of 93.0% for the control set, and 92.0%, 93.0%, and 90.0%, respectively, for the patient set. The results demonstrate the possibility of automatic PLMS detection in patients by using the deepPLM model based on a single-lead ECG. This could be an alternative method for PLMS screening and a helpful tool for home healthcare services for the elderly population.

1. Introduction

Periodic limb movement syndrome (PLMS) is a repetitive, transient movement caused by specific muscle tension during sleep. PLMS episodes are repeated four or more times in a row, with each episode in the range of 0.5–5 s in duration. The interval between two episodes is 5–90 s. A frequent occurrence of PLMS (etc., a high number of average PLMSs per hour) can lead to arousal or sleep fragmentation that can affect health and well-being [1]. PLMS is also known to be associated with several other disorders, including cardiovascular disease, hypertension, depression, narcolepsy, rapid eye movement disorders, and Parkinson’s disease [2,3,4,5,6]. Therefore, early and simple diagnosis of PLMS is clinically important to prevent other related diseases.
Nocturnal polysomnography (PSG) is the standard diagnostic test for sleep disorders, including PLMS and restless legs syndrome. PSG performs tests such as electroencephalogram (EEG), electroencephalogram (EOG), electrocardiogram (ECG), and electromyography (EMG) by attaching various types of equipment to the patient’s body to measure vital signs. In addition, PSG can accurately and objectively diagnose all sleep-related disorders based on bio-signals recorded during sleep [7]. However, PSG is expensive, inconvenient, and depends on sleep experts. Most importantly, it is a time-consuming task as the sleep expert has to manually annotate the PLMS, and the results of the annotation may vary depending on the sleep expert’s experience and proficiency.
Electromyography (EMG) is an essential or standard physiological signal for the automatic detection of PLMS. Conventional studies have proposed rule-based models, machine learning algorithms, and deep learning models for automatic detection of the PLMS using EMG signal. Wetter et al. [8] invented the rule-based model of automatic detection of PLMS using a voltage threshold and time bridging of EMG signal. Ferri et al. [9] proposed a novel rule-based approach based on a double threshold to detect PLMS after rectification. Recently, Moore et al. [10] studied a machine learning algorithm applied to ECG and EMG signals for PLMS detection. They used adaptive signal processing for enhancing the ECG interference, EMG noise filtering, and adaptive thresholds. Cavelli et al. [11] proposed a deep learning model for PLMS scoring based on the convolutional neural network (CNN) and long short-term memory (LSTM) applied to EMG signals. However, there are no studies on deep learning for the automatic detection of PLMS using ECG signals only.
Several studies have proposed alternative methods for scoring PLMS based on various sensors such as accelerometer, bands, and film sensors [12,13,14,15]. King et al. [13] presented a novel approach to measuring PLMS using Actiwatch (Cambridge Neurotechnology Ltd., Cambridge, UK). They showed the promise of continuously measuring PLMS for monitoring and treatment. Prill and Fahrenberg [14] investigated how to evaluate PLMS using multiple accelerometers. However, these studies required additional devices or multiple sensors and their performance was insufficient. Finally, Wetter et al. [8] developed an automatic scoring method based on EMG signals during PSG. They achieved an accuracy of 92.5% in detecting PLMS episodes. Portable or home PSGs are essential for home sleep monitoring to diagnose sleep disorders [15]. The portable PSG requires only four channels for physiological signals such as ECG, respiration, SpO2, and photoplethysmography. Therefore, new algorithms based on single-lead ECG signals are needed for PLMS monitoring in laboratory and home settings to provide easy and accurate healthcare solutions.
In this study, our contribution or novelty is the easy and accurate model for the automatic detection of the PLMS events. For easy implementation, a single-lead ECG-based detection will be demonstrated based on a deep learning algorithm. To achieve reliable results, a deep learning algorithm was designed with optimal architecture and trained real clinical datasets. Finally, we propose a novel approach for automatic detection of PLMS patients using a deep learning model using single-lead ECG signals. The proposed model is called deepPLM and it consists of CNN and LSTM with optimized architecture and parameters for PLMS detection from ECG signal. Clinical data sets obtained from control and patient groups with PLMS were used for performance evaluation of the training and testing phases.

2. Materials and Methods

2.1. Study Population

For this study, data from the Osteoporotic Fractures in Men (MrOS) Sleep Study, which was conducted as an auxiliary study for osteoporotic fractures, was utilized. The MrOS Sleep Study examined 5994 men aged 65 years or older between 2000 and 2002 [16]. Out of these participants, PSG was conducted on 2911 people. The following exclusion criteria was applied to all PSG subjects. The exclusion criteria were subjects with a periodic leg movement index (PLMI) of less than 15 in the patient group and a PLMI of 15 or more in the normal group were excluded. Then, the subjects were excluded who were diagnosed with sleep disorders including sleep apnea, insomnia, and narcolepsy. Finally, we excluded subjects who currently use the pacemakers (Figure 1).
Table 1 summarizes the clinical characteristics of the study subjects who passed the exclusion criteria.

2.2. ECG Dataset

The ECG signals were extracted from the PSG records of all subjects and stored at a sampling rate of 517 Hz. The analysis signal was an ECG signal with an average length of 2.8 h after each subject’s sleeping onset, and the total analysis length was 144.6 h. The entire ECG was segmented into units of 10 s and consisted of 5120 samples per segment. In the dataset, the ECG section of the normal group was indexed as 0, that of the patient group was indexed as 1, and the ratio of 0 to 1 was 1:1. For the training and evaluation of the proposed deepPLM model, the entire dataset was divided into a training group (66,560 segments from 32 subjects), an evaluation group (16,640 segments from 8 subjects), and a test group (20,800 segments from 12 subjects). The distribution of each dataset for normal and PLM events is represented in Table 2.

2.3. DeepPLM Model

In this study, the deepPLM model was constructed and optimized for the automatic detection of PLMS from a single-lead ECG since PLMSs result from several complex causes and the ECG signal contains information about movement, respiratory, and cardiac activities. We intended to develop and validate a deep learning model that can train and learn from the single-lead ECG to distinguish between PLMS and non-PLMS. Therefore, the deepPLM model is composed of a combined structure of the deep neural network that uses a CNN and LSTM, a recurrent neural network (RNN). First, a CNN was stacked in four layers and used to automatically extract a feature map by learning the pattern of the input signal [17]. Then, to control the periodic correlation or long-term dependence on the extracted feature maps, LSTM was connected in a two-layer structure. The final judgment was performed by configuring the complete connection layer and the soft-max function. Figure 2 illustrates the configured deepPLM model in detail.
The proposed deepPLM model has been implemented by one-dimensional (1D) convolutional layers and LSTM units to accurately detect the PLMSs from the ECG segments. In addition, the proposed deepPLM model was optimized using batch-normalization [18], dropout [19], and activation function [20] not only to avoid over and underfitting, but also to achieve robust detection performances.
All ECG input is applied to the batch-normalization as a pre-processor before training the configured deepPLM model. Batch-normalization is implemented by Equation (1).
x b = α ( x i μ σ 2 + ε ) + β
where ε—random noise, μ—mean of mini-batches, σ—variance of mini-batches, α—scale parameter, and β—shift parameter. Both α and β are trainable and updated in an epoch-wise manner.
Since input ECG segments are selected from the physiological signal and shaped as 1D time series, a 1D convolutional layer is appropriate to extract features by high order data abstractions [21]. A 1D convolution layer was implemented by the 1D convolution operation that is faster than two or three-dimensional convolutions, and it can be simply expressed the following equation:
x k = b k + i = 1 N w k × y i
where xk is the k-th feature map, bk is the bias of the k-th feature map, wk is the k-th convolutional kernel from all features of the k-th feature map, and yi represents the i-th feature map.
The pooling layer was used after each convolutional layer to reduce the dimensions of the intermediate feature maps. The max-pooling approach and appropriate parameters were applied in the pooling layers.
LSTM is an updated version of simple RNN with memory cells to make learning temporal associations easy over the long duration. [22]. LSTM consists of the three main memory cells which are called gates: an input gate, an output gate, and a forget gate [23,24]. LSTM is expressed as follows.
The input gate controls the flow of input activations into the memory cell.
it = σ(Wxixt + Whiht−1 + bi).
The output gate controls the output flow of cell activations into the rest of the network.
ot = σ(Wxoxt + Whoht−1 + bo).
The forget gate scales the internal state of the cell before adding it as input through the self-recurrent connection of the cell. Therefore, it adaptively forgets or resets the cell’s memory.
ft = σ(Wxfxt + Whfht−1 + bf),
gt = σ(Wxcxt + Whcht−1 + bc),
ct = ft × ct ̶ 1 + it × gt,
ht = ot × ϕ(ct),
where i, f, o, and c are respectively the input gate, forget gate, output gate, and cell activation vectors, all of which are the same size as vector h, defining the hidden value. Terms σ and τ represent nonlinear and hyperbolic tangent functions, respectively.
Lastly, some optimization techniques such as dropout and rectified-liner unit (ReLU) were applied to the configured deepPLM model. Dropout is a technique of randomly dropping out the nodes in a network to reduce overfitting by preventing complex adaptations on training data in the network [19]. ReLU is an activation function known as the robust training performance and consistent gradients, thereby facilitating gradient-based learning [25]. ReLU can be represented as
f ( x ) = max ( 0 , w x + b )
where x is the feature map, w is the weight, and b is the bias.

2.4. Implementation

In this study, data selection was performed using the statistical software R, and the pre-processing of PSG data was conducted using Python 3.7.4. The deepPLM model was designed based on the Keras [26] and TensorFlow [27] frameworks. To train and test the deepPLM model, we used hardware embedded with a graphics processing unit (GPU), the GeForce GTX 1080 Ti (11 GB, GDDR5X). The model performance was compared with that of a model built with a batch size of 16 and 500 epochs of iterative learning [28].

2.5. Evaluation Index

To evaluate the performance of the deepPLM model using a single-lead ECG signal, the following evaluation measures were used: precision, recall, and F1-score. To obtain the F1-score, two evaluation measures, precision and recall, were combined. These are defined as follows:
p r e c i s i o n = TP TP + FP
r e c a l l = TP TP + FN
where TP, FP, and FN represent the true positive, false positive, and false negative, respectively. They are determined for each sleep stage event.
The F1-score, better known as the unbalanced dataset, is computed based on the sample proportion of precision and recall as follows:
F 1 = 2 × p r e c i s i o n r e c a l l p r e c i s i o n + r e c a l l

3. Results

3.1. Performance of the Single-Lead ECG-Based Detection

This study demonstrated a single-lead ECG-based approach for the automatic detection of PLMS. The results showed very high performances for the automatic detection of PLMSs based on the single-lead ECG signal (Table 3). For the evaluation measures, we used many indexes including the precision, recall, accuracy, and F1-score. Among them, the F1-score of the deepPLM model was found to be 96% for the training set and 92% for the validation and test sets. In addition, the detection accuracy was high, at 88% for the training group, 92% for the evaluation group, and 91.5% for the test group.
Figure 3 shows the evaluation results of the proposed deepPLM model in a confusion matrix for each dataset.

3.2. Performance of the DeepPLM Model Optimization

The proposed deepPLM model has a simple design and well optimizer. ROC and AUC were used to these characteristics of the deepPLM for automatic detection of PLMS. Figure 4 shows the evaluation results via the ROC curve, which indicates that the AUC value is very high at 99% for the training group, 98% for the evaluation group, and 98% for the test group.

4. Discussion

This study proposes a novel approach based on deep learning for the automatic detection of PLMS in patients using a single-lead ECG signal. Deep learning was constructed, optimized, and named as the deepPLM model, and it was evaluated using clinical data sets measured from the control group and patients with PLMS from the MrOS database. The training group achieved high performance with an accuracy of 88%, the evaluation group achieved an accuracy of 92%, and the accuracy test group achieved an accuracy of 91.5%. The results showed a possibility of the automatic detection of PLMS from single-lead ECG using a deep learning model.
PLMS is considered prevalent among people over the age of 65. However, predictive methods for PLMS patients are not yet available [29]. Because of motion artifacts in ECG signals, the proposed approach can accurately distinguish between normal and abnormal ECGs for deep learning-based automatic detection of PLMS. ECG signals are an essential and important source of clinical information for sleep laboratory nighttime PSG as well as home sleep monitoring using portable PSG. The proposed approach for automatic detection of PLMS provides a simple deepPLM architecture that can be easily implemented with real world data. The proposed approach shows advantages such as robust performance, simultaneous usability in laboratory and home sleep monitoring, and easy implementation. First, the results have shown the possibility of the automatic detection of the PLMS from only short-term ECG segments. Since the PLMS is automatically detected, it can be used in prescreening for the associated diseases, including the cardiovascular diseases [30], cerebrovascular risks [31], and Parkinson disease [32]. Second, we studied an alternative model for automatic detection of PLMS based on deep learning from the single-lead ECG signal. The proposed deepPLM model can be used as an easy screening tool based on PSG or portable ECG that does not use the phenotypes and clinical parameters. Finally, we achieved a higher detection performance than similar conventional studies listed in Table 4.
Table 4 summarizes comparisons with previous related studies on scoring or detecting PLMS. The proposed deepPLM model achieved the best performance for the automatic detection of PLMS, which outperforms the related previous studies. Conventional studies proposed a rule-based method such as voltage thresholding and time bridging of the EMG signal [8], and double thresholding techniques after rectification [9]. However, these presented a lower performance for the PLMS detection. Another study proposed adaptive signal processing methods for PLMS detection from EMG signals. The authors used adaptive processing to the ECG interference, noise cancelation, and adaptive thresholds for PLMS detection from EMG signals [10]. Finally, Cavelli et al. [11] proposed a deep learning model for PLMS scoring using CNN and LSTM similar to the proposed model. All these studies used EMG signal for PLMS detection, and they cover the rule-based model, machine learning, and deep learning algorithms. However, the results were shown to be lower than the proposed deepPLM model which used single-lead ECG signals.
Nevertheless, this study has limitations in that it requires a small amount of data and high computational power. For this study, we used ECG signals from 52 subjects. Future studies should use a larger study population to overcome these limitations. To compute high-dimensional data abstractions, deepPLM models require relatively higher computational power than traditional machine learning methods. Therefore, in this study, a CNN–LSTM model was constructed and optimized with a simple and small structure.
In summary, we propose a novel method for automatic PLMS prediction in patients using a deep learning model with ECG signals. The deepPLM model was constructed and evaluated using datasets from control and PLMS patients. A high satisfactory performance was obtained on the training and test sets. We also propose an alternative method to predict a patient’s PLMS. Our results demonstrate the feasibility of using deep learning models with ECG signals for the automatic detection of PLMS. The results also demonstrate that the single-lead ECG signal can be used as a discriminant and alternative signal for patients with PLMS. The proposed approach is sufficient and can be a useful predictive tool for detecting sleep-related movements, including PLMS. Further studies will require a more diverse patient group and larger data sets to confirm and support these findings.

Author Contributions

Conceptualization, J.-U.P. and E.U.; methodology, J.-U.P. and E.U.; writing—original draft preparation, J.-U.P. and E.U.; writing—review and editing, J.-U.P. and E.U.; visualization, J.-U.P. and E.U.; software, J.-H.L.; formal analysis, J.-H.L.; data curation, J.-H.L.; validation, S.-B.K.; investigation, S.-B.K.; resources, S.-B.K.; supervision, K.-J.L.; project administration, K.-J.L.; funding acquisition, K.-J.L. All authors have read and agreed to the published version of the manuscript.

Funding

This research was supported by the National Information Society Agency (NIA) and funded by the Ministry of Science and ICT through the Big Data Platform and Center Construction Project (No. 2022-Data-W18). This research was partially supported by a grant of the Medical data-driven hospital support project through the Korea Health Information Service (KHIS), funded by the Ministry of Health and Welfare, Republic of Korea. This research was partially supported by the Ministry of Education of the Republic of Korea and the National Research Foundation of Korea (NRF-2020S1A5A2A03045088).

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

https://sleepdata.org/datasets/mros. All data is approved by the National Sleep Research Resource (NSRR) for the specific purpose of this study.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Scofield, H.; Roth, T.; Drake, C. Periodic limb movements during sleep: Population prevalence, clinical correlates, and racial differences. Sleep 2008, 31, 1221–1227. [Google Scholar] [PubMed]
  2. Boivin, D.B.; Montplaisir, J.; Poirier, G. The effects of L-dopa on periodic leg movements and sleep organization in narcolepsy. Clin. Neuropharmacol. 1989, 12, 339–345. [Google Scholar] [CrossRef]
  3. Lapierre, O.; Montplaisir, J. Polysomnographic features of REM sleep behavior disorder: Development of a scoring method. Neurology 1992, 42, 1371–1374. [Google Scholar] [CrossRef] [PubMed]
  4. Ancoli-Israel, S.; Kripke, D.F.; Mason, W.; Kaplan, O.J. Sleep apnea and periodic movements in an aging sample. J. Gerontol. 1985, 40, 419–425. [Google Scholar] [CrossRef] [PubMed]
  5. Wetter, T.C.; Collado-Seidel, V.; Pollmacher, T.; Yassouridis, A.; Trenkwalder, C. Sleep and periodic leg movement patterns in drug-free patients with Parkinson’s disease and multiple system atrophy. Sleep 2000, 23, 361–367. [Google Scholar] [CrossRef] [PubMed]
  6. Koo, B.B.; Sillau, S.; Dean, D.A.; Lutsey, P.L.; Redline, S. Periodic limb movements during sleep and prevalent hypertension in the multi-ethnic study of atherosclerosis. Hypertens 2015, 65, 70–77. [Google Scholar] [CrossRef]
  7. Douglas, N.J.; Thomas, S.; Jan, M.A. Clinical value of polysomnography. Lancet 1992, 339, 347–350. [Google Scholar] [CrossRef]
  8. Wetter, T.C.; Dirlich, G.; Streit, J.; Trenkwalder, C.; Schuld, A.; Pollmächer, T. An automatic method for scoring leg movements in polygraphic sleep recordings and its validity in comparison to visual scoring. Sleep 2004, 27, 324–328. [Google Scholar] [CrossRef]
  9. Ferri, R.; Zucconi, M.; Manconi, M.; Bruni, O.; Miano, S.; Plazzi, G.; Ferini-Strambi, L. Computer-assisted detection of nocturnal leg motor activity in patients with restless legs syndrome and periodic leg movements during sleep. Sleep 2005, 28, 998–1004. [Google Scholar] [CrossRef]
  10. Moore, H.; Leary, E.; Lee, S.Y.; Carrillo, O.; Stubbs, R.; Peppard, P.; Young, T.; Widrow, B.; Mignot, E. Design and validation of a periodic leg movement detector. PLoS ONE 2014, 9, e114565. [Google Scholar]
  11. Carvelli, L.; Olesen, A.N.; Brink-Kjær, A.; Leary, E.B.; Peppard, P.E.; Mignot, E.; Sørensen, H.B.D.; Jennum, P. Design of a deep learning model for automatic scoring of periodic and non-periodic leg movements during sleep validated against multiple human experts. Sleep Med. 2020, 69, 109–119. [Google Scholar] [CrossRef] [PubMed]
  12. Rauhala, E.; Virkkala, J.; Himanen, S.L. Periodic limb movement screening as an additional feature of Emfit sensor in sleep-disordered breathing studies. J. Neurosci. Methods 2009, 178, 157–161. [Google Scholar] [CrossRef] [PubMed]
  13. King, M.A.; Jaffre, M.O.; Morrish, E.; Shneerson, J.M.; Smith, I.E. The validation of a new actigraphy system for the measurement of periodic leg movements in sleep. Sleep Med. 2005, 6, 507–513. [Google Scholar] [CrossRef] [PubMed]
  14. Prill, T.; Fahrenberg, J. Simultaneous assessment of posture and limb movements (e.g., periodic leg movements) with calibrated multiple accelerometry. Physiol. Meas. 2006, 27, N47–N53. [Google Scholar] [CrossRef]
  15. Ferber, R.; Millman, R.; Coppola, M.; Fleetham, J.; Murray, C.F.; Iber, C.; McCall, W.V.; Nino-Murcia, G.; Pressman, M.; Sanders, M. Portable recording in the assessment of obstructive sleep apnea. ASDA standards of practice. Sleep 1994, 17, 378–392. [Google Scholar] [CrossRef]
  16. Blank, J.B.; Cawthon, P.M.; Carrion-Petersen, M.L.; Harper, L.; Johnson, J.P.; Mitson, E.; Delay, R.R. Overview of recruitment for the osteoporotic fractures in men study (MrOS). Contemp. Clin. Trial. 2005, 26, 557–568. [Google Scholar] [CrossRef]
  17. Kiranyaz, S.; Ince, T.; Gabbouj, M. Real-time patient-specific ECG classification by 1-D convolutional neural networks. IEEE Trans. Biomed. Eng. 2016, 63, 664–675. [Google Scholar] [CrossRef]
  18. Ioffe, S.; Szegedy, C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. Int. Conf. Mach. Learn. 2015, 37, 448–456. [Google Scholar]
  19. Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; Salakhutdinov, R. Dropout: A Simple Way to Prevent Neural Networks from Overfitting. J. Mach. Learn. Res. 2014, 15, 1929–1958. [Google Scholar]
  20. Agostinelli, F.; Hoffman, M.; Sadowski, P.; Baldi, P. Learning activation functions to improve deep neural networks. arXiv 2014, arXiv:1412.6830. [Google Scholar]
  21. Dey, D.; Chaudhuri, S.; Munshi, S. Obstructive sleep apnoea detection using convolutional neural network based deep learning framework. Biomed. Eng. Lett. 2018, 8, 95–100. [Google Scholar] [CrossRef] [PubMed]
  22. Hochreiter, S.; Schmidhuber, J. Long short-term memory. Neural Comput. 1997, 9, 1735–1780. [Google Scholar] [CrossRef] [PubMed]
  23. Sak, H.; Senior, A.; Beaufays, F. Long short-term memory based recurrent neural network architectures for large vocabulary speech recognition. arXiv 2014, arXiv:1402.1128. [Google Scholar]
  24. Chung, J.; Gulcehre, C.; Cho, K.; Bengio, Y. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv 2014, arXiv:1412.3555. [Google Scholar]
  25. Nair, V.; Hinton, G.E. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th International Conference on Machine Learning, Haifa, Israel, 21–24 June 2010. [Google Scholar]
  26. Chollet, F. Keras. 2015. Available online: http://keras.io/ (accessed on 7 June 2016).
  27. TensorFlow. Available online: https://www.tensorflow.org/ (accessed on 7 April 2021).
  28. Kingma, D.P.; Ba, J. Adam: A method for stochastic optimization. arXiv 2014, arXiv:1412.6980. [Google Scholar] [CrossRef]
  29. Ferri, R.; Novelli, L.; Bruni, O. Periodic Limb Movement Disorder. Reference Module in Neuroscience and Biobehavioral Psychology; Elsevier: Amsterdam, The Netherlands, 2017. [Google Scholar]
  30. Ferri, R.; Koo, B.B.; Picchietti, D.L.; Fulda, S. Periodic leg movements during sleep: Phenotype, neurophysiology, and clinical significance. Sleep Med. 2017, 31, 29–38. [Google Scholar] [CrossRef] [Green Version]
  31. Haba-Rubio, J.; Marti-Soler, H.; Tobback, N.; Andries, D.; Marques-Vidal, P.; Vollenweider, P.; Preisig, M.; Heinzer, R. Clinical significance of periodic limb movements during sleep: The HypnoLaus study. Sleep Med. 2018, 41, 45–50. [Google Scholar] [CrossRef]
  32. Hwang, S.R.; Hwang, S.W.; Chen, J.C.; Hwang, J.H. Association of periodic limb movements during sleep and Parkinson disease: A retrospective clinical study. Medicine 2019, 98, e18444. [Google Scholar] [CrossRef]
Figure 1. The exclusion criteria of the study population for the proposed deepPLM.
Figure 1. The exclusion criteria of the study population for the proposed deepPLM.
Diagnostics 12 02149 g001
Figure 2. Architecture of the proposed deepPLM for the automatic detection of PLMSs.
Figure 2. Architecture of the proposed deepPLM for the automatic detection of PLMSs.
Diagnostics 12 02149 g002
Figure 3. Confusion matrix of the proposed deepPLM model for the automatic detection of PLMS: (A) training set, (B) validation set, and (C) test set.
Figure 3. Confusion matrix of the proposed deepPLM model for the automatic detection of PLMS: (A) training set, (B) validation set, and (C) test set.
Diagnostics 12 02149 g003
Figure 4. ROC and AUC of the proposed deepPLM model for the automatic detection of PLMS.
Figure 4. ROC and AUC of the proposed deepPLM model for the automatic detection of PLMS.
Diagnostics 12 02149 g004
Table 1. Demographics of the study population.
Table 1. Demographics of the study population.
CharacteristicsNormalPLM
Subjects (N)2626
Age (years)76.12 ± 5.5176.08 ± 5.11
Periodic leg movement index (per hour)2.46 ± 4.1657.88 ± 30.27
Body mass index (kg/m2)27.92 ± 3.1229.15 ± 3.89
Sleep efficiency (%)74.35 ± 10.9373.00 ± 11.34
Smoking status, n (%)
    Never
    Past
12 (47.15%)
14 (53.85%)
12 (56.0%)
14 (40.0%)
Blood pressure
    Systolic
    Diastolic
127.57 ± 12.82
66.85 ± 5.66
127.35 ± 19.08
68.81 ± 7.35
Table 2. ECG dataset information.
Table 2. ECG dataset information.
DatasetsNormalPLMTotal
Training set33,28033,28066,560
Validation set8320832016,640
Test set10,40010,40020,800
Total52,00052,000104,000
Table 3. The performance of the deepPLM model for the automatic detection of PLMS patients.
Table 3. The performance of the deepPLM model for the automatic detection of PLMS patients.
DatasetsSegmentPrecisionRecallF1-ScoreAccuracy
Training setNormal0.940.970.960.89
PLM0.970.940.96
Validation setNormal0.900.940.920.92
PLM0.940.900.92
Test setNormal0.900.930.920.92
PLM0.930.900.92
Table 4. Comparison with previous related studies.
Table 4. Comparison with previous related studies.
Authors
(Year of Publication)
No.
of Subjects
SignalMethodResults
(F1-Score)
Wetter et al. (2004) [8]24EMGEMG-based analytical method0.63
Ferri et al. (2005) [9]30EMGComputer-assisted detection method0.72
Moore et al. (2014) [10]1833EMG,
ECG
Ten-step PLM detection method0.79
Carvelli et al. (2020) [11]800EMGCNN–LSTM model0.85
This work52ECGCNN–LSTM model0.92
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share and Cite

MDPI and ACS Style

Urtnasan, E.; Park, J.-U.; Lee, J.-H.; Koh, S.-B.; Lee, K.-J. Deep Learning for Automatic Detection of Periodic Limb Movement Disorder Based on Electrocardiogram Signals. Diagnostics 2022, 12, 2149. https://doi.org/10.3390/diagnostics12092149

AMA Style

Urtnasan E, Park J-U, Lee J-H, Koh S-B, Lee K-J. Deep Learning for Automatic Detection of Periodic Limb Movement Disorder Based on Electrocardiogram Signals. Diagnostics. 2022; 12(9):2149. https://doi.org/10.3390/diagnostics12092149

Chicago/Turabian Style

Urtnasan, Erdenebayar, Jong-Uk Park, Jung-Hun Lee, Sang-Baek Koh, and Kyoung-Joung Lee. 2022. "Deep Learning for Automatic Detection of Periodic Limb Movement Disorder Based on Electrocardiogram Signals" Diagnostics 12, no. 9: 2149. https://doi.org/10.3390/diagnostics12092149

APA Style

Urtnasan, E., Park, J. -U., Lee, J. -H., Koh, S. -B., & Lee, K. -J. (2022). Deep Learning for Automatic Detection of Periodic Limb Movement Disorder Based on Electrocardiogram Signals. Diagnostics, 12(9), 2149. https://doi.org/10.3390/diagnostics12092149

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop