Communications Engineering

Table 3 Comparison results based on transfer learning on ECG datasets

From: Transforming label-efficient decoding of healthcare wearables with self-supervised learning and “embedded” medical domain expertise

Methods	CPSC				CinC17
	F1	Recall	Prec.	AUC	F1	Recall	Prec.	AUC
Random Init.	0.585	0.599	0.605	0.905	0.536	0.599	0.547	0.841
SimCLR⁹	0.658	0.689	0.652	0.930	0.627	0.705	0.626	0.883
BYOL³⁷	0.658	0.672	0.654	0.925	0.630	0.712	0.616	0.880
MoCo³⁸	0.640	0.663	0.639	0.929	0.651	0.719	0.625	0.881
NNCLR³⁹	0.640	0.667	0.640	0.926	0.644	0.717	0.620	0.885
TS⁴⁰	0.627	0.638	0.630	0.921	0.624	0.715	0.602	0.887
SwAV⁴¹	0.654	0.670	0.650	0.929	0.642	0.716	0.628	0.889
AMCL⁴²	0.660	0.668	0.658	0.933	0.649	0.725	0.630	0.892
CLOCS¹⁰	0.647	0.671	0.637	0.927	0.652	0.740	0.619	0.887
TFC¹⁵	0.640	0.661	0.650	0.922	0.648	0.717	0.614	0.893
SoftIns⁴³	0.657	0.670	0.652	0.935	0.645	0.730	0.629	0.890
RNC⁴⁹	0.642	0.652	0.653	0.926	0.654	0.734	0.634	0.893
Ours	0.670	0.667	0.675	0.945	0.680	0.751	0.644	0.901

Based on the pretrained model with self-supervised learning on the large-scale MIMIC-III-WDB dataset, the results on CPSC and CinC17 are reported, respectively. We compared the class-average F1, AUC, Precision, Recall, and Accuracy on the test subset based on the model with the best F1 value on the validation set. Bold indicates the best result, and Underline indicates the second-best result.

Back to article page

Search

Advanced search

Quick links