TABLE IV.

Subject-independent testing results for the multiclass classification of all models on the Pulsewatch dataset.

	Index	Model name	Rhythm	Sensitivity	Specificity	Precision	NPV	Accuracy	Macro-AUROC
Previous (retrained)	1	1D-VGG-16 [6] (PPG only)	NSR	97.40	91.25	95.83	91.65	94.48	95.82
			AF	90.21	97.67	85.73	97.34	94.76	95.82
			PAC/PVC	63.34	98.09	71.27	95.44	93.03	95.82
	2	1D-VGG-16 [6] (four channels)	NSR	96.06	91.87	96.02	87.49	93.12	98.09
			AF	97.61	98.25	89.43	99.34	97.06	98.09
			PAC/PVC	71.32	97.77	73.37	96.39	93.88	98.09
	3	2D DenseNet [8] (2D TFS)	NSR	95.31	91.33	95.73	85.97	92.38	95.86
			AF	89.21	97.86	86.70	97.09	94.83	95.86
			PAC/PVC	59.43	95.83	52.82	94.80	89.50	95.86
Ours (1D-Bi-GRU)	4	PPG only	NSR	92.45	95.66	97.75	82.33	91.96	95.84
			AF	89.65	95.83	79.81	97.14	93.03	95.84
			PAC/PVC	67.47	94.96	54.64	95.77	90.08	95.84
	5	PPG + ACC	NSR	89.65	94.61	97.10	76.50	88.86	95.91
			AF	93.63	96.84	82.80	98.23	94.55	95.91
			PAC/PVC	71.03	92.56	49.47	96.13	88.62	95.91
	6	HR only	NSR	91.01	93.16	96.42	78.91	89.74	95.29
			AF	90.16	97.25	85.76	97.33	94.76	95.29
			PAC/PVC	73.47	93.35	51.60	96.46	89.30	95.29
	7	HR + ACC	NSR	90.52	95.15	97.40	77.68	89.60	97.03
			AF	97.03	97.90	88.58	99.18	96.73	97.03
			PAC/PVC	80.90	93.31	55.09	97.44	90.46	97.03
	8	HR + magHR	NSR	92.11	93.40	96.61	81.64	91.08	96.35
			AF	93.74	97.23	87.37	98.29	95.82	96.35
			PAC/PVC	71.99	94.56	56.35	96.34	90.61	96.35
	9	HR + magHR + ACC	NSR	89.56	95.80	97.72	76.39	89.07	97.18
			AF	95.91	97.99	89.00	98.88	96.63	97.18
			PAC/PVC	81.98	92.22	51.69	97.54	89.39	97.18
	10	PPG + HR	NSR	93.15	94.38	97.14	83.73	92.25	95.15
			AF	91.49	97.60	87.24	97.69	95.38	95.15
			PAC/PVC	70.85	94.56	55.57	96.19	90.38	95.15
	11	PPG + HR + ACC	NSR	90.02	95.68	97.66	76.74	89.23	96.29
			AF	95.26	98.07	87.57	98.70	96.14	96.29
			PAC/PVC	76.43	92.12	49.43	96.80	88.59	96.29
	12	PPG + HR + ACC + magHR (best model)	NSR	91.30	96.18	97.98	79.62	90.81	97.68
			AF	97.52	98.30	90.47	99.32	97.31	97.68
			PAC/PVC	83.52	93.60	57.48	97.79	91.23	97.68