Table 4 Comparison of domain feature usage strategies on the CPSC (ECG) dataset

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

Methods

F1

Deep Regression

0.510

Domain Feature Ranking

0.550

Nearest Neighbour Search Only

0.550

Ours

0.574

  1. We evaluate four different approaches for incorporating domain features into self-supervised learning. Deep Regression: A multilayer perceptron is added to regress deep features to domain features. Ranking (RNC49): Samples are ranked using Euclidean distance in the domain feature space. Nearest Neighbour Only: Our proposed instance-level contrast is based on domain-feature-guided nearest neighbors. Ours (Full): Combines both domain-guided instance- and prototype-level contrast.
Back to article page