A Novel Continuous Sleep State Artificial Neural Network Model Based on Multi-Feature Fusion of Polysomnographic Data

Figures & data

Figure 1 (A–C) Flow chart of the algorithm.

Figure 2 Multi-feature selection based on timing fitness.

Table 1 Statistical Results of the Sleep Depth Sample Values

Sleep Stages	Mean of # 3s Epochs/Patient	Std of # 3s Epochs/Patient	Percentage in Total (%)	Mean of SDV	Std of SDV
Awake	2397	1645	24.87	0.7021	0.2702
N1	803	672	8.33	0.2353	0.2709
N2	3380	1067	35.07	0.1202	0.1907
N3	1664	486	17.27	0.0396	0.0969
REM	1393	567	14.46	0.1847	0.2043

Notes: Results expressed as arithmetic mean and standard deviations. “Mean and Std of # 3s epochs/patient” and “Percentage in Total” are calculated based on the manual labeled data. “Mean and Std of SDV” are calculated based on sleep depth sample values.

Abbreviations: # 3s epochs/patient, Number of 3s epochs per patient; SDV, Sleep Depth Value; Std, Standard Deviation.

Figure 3 Comparison between sleep depth samples and sleep stages. (a) is the manually calibrated sleep staging map, with the horizontal axis representing time points (a point every 3 seconds), and the vertical axis representing sleep stages; (b) is the SDV sample values map for every 3 seconds.

Figure 4 Statistical result of sleep depth samples.

Figure 5 Error histogram of ANN model.

Figure 6 Prediction result of ANN model. (a) is the manually calibrated sleep staging map, with the horizontal axis representing time points (a point every 1 seconds), and the vertical axis representing sleep stages; (b) is the sleep depth values predicted by the ANN model.

Figure 7 Details of prediction result. (a) is the details between 500–3500 seconds of manually sleep staging result; (b) is the details between 500–3500 seconds of sleep depth values predicted by the ANN model.

Figure 8 Sleep staging result based on SDV. (a) is the manually calibrated sleep staging map, with the horizontal axis representing time points (a point every 3 seconds), and the vertical axis representing sleep stages; (b) is the SVM-classified stages result based on 3-second epochs; (c) is the result of SVM refined by HMM model.

Table 2 Accuracy Results of Sleep Stage Classification

Sleep Stages Algorithm	Awake Accuracy	N1 Accuracy	N2 Accuracy	N3 Accuracy	REM Accuracy	Total Accuracy
KNN (mean)	85.66%	17.10%	70.44%	63.02%	60.11%	67.01%
KNN (std)	0.0751	0.1031	0.136	0.2231	0.1596	0.0614
Random Forest (mean)	91.02%	25.09%	76.87%	63.69%	62.97%	71.79%
Random Forest (std)	0.0811	0.0732	0.106	0.2091	0.1329	0.057
SVM (mean)	95.63%	20.30%	87.82%	78.74%	73.29%	80.47%
SVM (std)	0.0751	0.0908	0.1302	0.236	0.1825	0.0696
KNN+HMM (mean)	90.30%	34.06%	78.69%	79.19%	78.12%	77.87%
KNN+HMM (std)	0.0795	0.2017	0.1866	0.2542	0.1912	0.0822
Random Forest+HMM (mean)	91.16%	40.47%	85.50%	79.07%	82.07%	81.55%
Random Forest+HMM (std)	0.0892	0.2433	0.1682	0.2588	0.1772	0.0925
SVM+HMM (mean)	98.12%	39.01%	95.26%	93.33%	89.12%	90.07%
SVM+HMM (std)	0.076	0.158	0.1793	0.2558	0.2008	0.0805

Notes: Results expressed as accuracy between the ground truth and the sleep stages prediction generated by K-Nearest Neighbors, Random Forest, and Support Vector Machine. The Hidden Markov Model method is used to improve the result from these three algorithms.

Abbreviations: KNN, K-Nearest Neighbors; SVM, Support Vector Machine; HMM, Hidden Markov Model.

Figure 9 Statistical result of SDV in arousal period.

Figure 10 Arousal detection based on SDV pattern recognition.

Figure 11 Comparison between two features in arousal period. (a) is the correlation between the frequency band for beta waves (feature number 23) and the arousal template; (b) is the correlation between the local entropy for theta waves (feature number 33) and the arousal template.

Figure 12 Arousal & awake probability between normal and OSA. (a) represents the probability of experiencing arousal during sleep for individuals with normal sleep patterns; (b) denotes the likelihood of arousal events occurring during sleep in individuals diagnosed with obstructive sleep apnea (OSA).