Machine Learning Techniques Applied to Data Analysis and Anomaly Detection in ECG Signals

Figures & data

Figure 1. An example of an ECG signal with marked folds.

Figure 2. The fragments of ECG signals’ records from the MIT-BIH base: (a) signal 105, (b) signal 108, (c) signal 203, (d) signal 222.

Table 1. The energy of signal’s residuum depending on decomposition level.

Decomposition level	Computational method
	MP—Gabor	MP—KSVD	OMP—Gabor	OMP—KSVD
1	28.31	34.31	35.13	39.75
2	32.15	40.15	42.08	48.66
3	47.34	51.46	55.89	61.77
4	59.67	69.04	72.42	80.54
5	78.98	83.25	88.34	92.79
6	86.26	88.94	90.66	94.03
7	88.57	90.07	92.09	94.88

Table 2. The maximum and minimum detection error for the given threshold value.

Detection error	Detection threshold value
	0.4	0.5	0.6	0.7	0.8
max	9.35	8.10	6.84	6.71	10.55
min	7.61	3.20	1.95	4.99	8.08

Table 3. The accuracy of detection in relation to the number of neurons in the hidden layer.

The number of neurons of hidden layers	The number of hidden layers
	1	2	3	4	5
20	97.68	98.12	98.21	97.32	97.34
30	97.70	98.46	98.53	97.82	97.82
40	97.78	98.57	98.66	97.70	97.70
50	97.85	98.45	98.48	97.67	97.69
60	97.67	98.24	98.27	97.56	97.54

Table 4. The accuracy of the QRS complex detection in relation to the activation function.

Type of activation function	Number of hidden layers
	1	2	3	4	5
Sigmoidal	98.36	98.55	98.63	98.70	98.71
Tangesoidal	98.22	98.44	98.79	98.71	98.72

Table 5. Comparison of average measure values of detection and sensitivity for different solutions.

Average detection measures	Type of method
	Time	Frequency	Time -Frequency	Neural Networks	SR-NN
Detection error	4.70	3.85	1.10	1.14	0.97
Sensitivity	98.60	98.44	99.48	98.99	98.98

Table 6. The comparison of detection error and sensitivity for signals: 105, 108, 203, and 222.

Detection measures	Type of signal
	105	108	203	222
Detection error	1,24	1.48	1.19	0.71
Sensitivity	99.64	98.92	99.41	98.99