Multi-scale features for heartbeat classification using directed acyclic graph CNN

Figures & data

Figure 1. Fifty randomly selected heartbeat signals for (a) Class N, (b) Class L, (c) Class R, (d) Class V, (e) Class A.

Table 1. Heartbeat classes given by the MIT-BIH database along with the regrouping defined by the AAMI standard.

MIT-BIH class	Annotation	AAMI groups
Normal beat	N	N: beats not found in the classes S, V, F and Q
Left bundle branch block beat	L
Right bundle branch block beat	R
Atrial escape beats	e
Nodal (junctional) escape beat	j
Atrial premature beats	A	S: supraventricular ectopic beats
Aberrated atrial premature beats	a
Nodal (junctional) premature beats	J
Non-conducted P-wave (blocked APB)	x
Premature ventricular contraction	V	V: ventricular ectopic beats
Ventricular escape beat	E
Ventricular flutter wave	!
Fusion of ventricular and normal beat	F	F: Fusion of ventricular and normal beat
Fusion of paced and normal beat	f	Q: paced beats or unclassified beats
Unclassified beat	Q

Figure 2. Example of CNN architecture (Mathworks online resource, 2017).

Figure 3. Visualization of the parameter setup at i-th ReLU and k-th ADD (Yang and Ramanan 2015).

Figure 4. Overview of proposed DAG-CNN method.

Table 2. Summary of the training and testing heartbeat samples.

AAMI heartbeat type	Total No.	Trainset	Testset
N	89695	45653	44042
S	2946	983	1963
V	7459	4252	3207
F	811	423	388
Q	15	8	7

Trainset (DS1) consist of heartbeats from records 101, 106, 108, 109, 112, 114, 115, 116, 118, 119, 122, 124, 201, 203, 205, 207, 208, 209, 215, 220, 223 and 230.

Testset (DS2) consist of heartbeats from records 100, 103, 105, 111, 113, 117, 121, 123, 200, 202, 210, 212, 213, 214, 219, 221, 222, 228, 231, 232, 233 and 234.

Records 102, 104, 107 and 217 were not used in trainset and testset.

Table 3. The details of back-bone CNN architecture of DAG-CNN model.

Layers	Type	No. of Neurons	Kernel size	Stride
1	Convolution	276 × 5	5	1
3	Max-pooling	138 × 5	2	2
4	Convolution	66 × 10	4	1
6	Max-pooling	33 × 10	2	2
7	Convolution	31 × 15	3	1
9	Max-pooling	33 × 10	2	2
10	Convolution	14 × 20	3	1
12	Max-pooling	7 × 20	2	2
13	Average-pooling	33 × 10	2	2
14 18–19-20	Average-pooling Fully Connected	33 × 10 25 × 15 × 5	2 -	2 -

Table 4. Different heartbeat types classification result.

		Classification result
Grand truth		N	S	V	F	Q
	N	37571	80	192	133	8
	S	427	1156	11	5	2
	V	88	2	2671	74	0
	F	60	1	2	281	0
	Q	6	0	0	1	0

Table 5. Four heartbeat types classification metrics compared to the state-of-the-art (percentage, %).

Methods	Year	Labelled	Accuracy	N		S		V		F
				Sen	Ppr	Sen	Ppr	Sen	Ppr	Sen	Ppr
Luo et al.	2017	300	97.5	99.0	98.4	71.4	94.4	93.3	93.3	82.7	58.5
Chazal and Reilly	2006	500	93.9	94.3	99.4	87.7	47.0	94.3	96.2	74.0	29.1
Jiang and Kong	2007	300	94.5	98.7	96.2	50.6	68.0	86.6	89.4	35.8	84.2
Ince et al.	2009	300	93.6	97.0	97.0	62.1	56.7	83.4	86.5	61.4	73.4
Martis et al.	2013	0	89.0	94.2	99.2	86.2	56.7	92.4	93.4	66.4	17.7
Alvarado et al.	2012	0	93.6	94.2	99.2	86.2	56.7	92.4	93.4	66.4	17.7
Ye et al.	2012	0	88.2	90.0	98.2	56.4	55.1	84.7	59.5	35.8	5.8
Zhang et al.	2014	0	88.3	88.9	99.0	79.1	36.0	85.5	92.8	93.8	13.7
Proposed methods		300	97.15	98.8	98.3	72.3	93.0	92.0	92.8	81.7	57.00

Table 6. SVEB and VEB Classification metrics compared to the state-of-the (percentage, %).

	SVEB				VEB
Methods	Acc	Sen	Ppr	Spe	Acc	Sen	Ppr	Spe
Luo et al.	98.8	71.4	94.4	99.8	99.1	93.3	93.3	99.5
Kiranyaz et al.	96.4	64.6	62.1	98.6	98.6	95	89.5	98.1
Chazal and Reilly	95.9	87.7	47.0	96.2	99.4	94.3	96.2	99.7
Jiang and Kong	96.6	50.6	68.0	98.8	97.7	86.6	89.4	98.9
Ince et al.	97.3	63.5	53.7	98.3	98.0	84.6	86.7	99.0
Martis et al.	93.3	83.2	33.5	93.7	97.4	86.8	75.9	98.1
Alvarado et al.	97.0	86.2	56.7	97.5	99.1	92.4	93.4	99.5
Ye et al.	97.4	56.4	55.1	98.6	94.6	84.7	59.5	95.4
Zhang et al.	93.3	79.1	36.0	93.9	98.6	85.5	92.7	99.5
Proposed methods	98.8	72.2	93.3	99.8	99.1	94.2	92.8	99.5

Mathworks online resource. 2017. Convolutional Neural Networks article, Natick, Massachusetts, United States. Accessed December 12, 2017. https://www.mathworks.com/content/mathworks/www/en/discovery/convolutional-neural-network/jcr:content.html.

 Google Scholar

Yang, S., and D. Ramanan. 2015. Multi-scale recognition with DAG-CNNs. Proceedings of the IEEE International Conference on Computer Vision, 1215–23

 Google Scholar