A Survey of Few-Shot Learning for Biomedical Time Series

TL;DR

This survey provides a comprehensive review and comparison of few-shot learning methods for biomedical time series applications and the clinical benefits and limitations of such methods are discussed in relation to traditional data-driven approaches.

Abstract

Advancements in wearable sensor technologies and the digitization of medical records have contributed to the unprecedented ubiquity of biomedical time series data. Data-driven models have tremendous potential to assist clinical diagnosis and improve patient care by improving long-term monitoring capabilities, facilitating early disease detection and intervention, as well as promoting personalized healthcare delivery. However, accessing extensively labeled datasets to train data-hungry deep learning models encounters many barriers, such as long-tail distribution of rare diseases, cost of annotation, privacy and security concerns, data-sharing regulations, and ethical considerations. An emerging approach to overcome the scarcity of labeled data is to augment AI methods with human-like capabilities to leverage past experiences to learn new tasks with limited examples, called few-shot learning. This survey provides a comprehensive review and comparison of few-shot learning methods for biomedical time series applications. The clinical benefits and limitations of such methods are discussed in relation to traditional data-driven approaches. This paper aims to provide insights into the current landscape of few-shot learning for biomedical time series and its implications for future research and applications.

Figures (3)

• Figure 1: Overview of few-shot learning setup. During the pre-training stage (optional), the model is trained to learn meaningful representations and initialize its parameters to facilitate future adaption to target tasks. Subsequently, the model learns a new task using the $N$-way-$K$-shot support set and performs inference on the query set to provide label prediction.
• Figure 2: Taxonomy of few-shot learning for biomedical time series. Different few-shot learning methods propose modifications to varying parts of the training pipeline, each highlighted using a distinct color. Data-based approaches directly address the data shortage problem by generating synthetic samples to increase the size and diversity of the support set. Metric-based approaches focus on learning the similarity between samples in a representation space. Model-based approaches design model architectures to improve generalization across tasks with few samples. Optimization-based approaches guide model convergence to parameter spaces that can be quickly adjusted by fine-tuning with a few samples.
• Figure 3: Comparison of metric-based methods in few-shot learning with a 3-way-2-shot setup. All methods employ an embedding network to project support and query samples into the representation space, in order to measure the similarity between support and query samples. a) Siamese network computes the similarity between each pair of support-query feature vectors to determine the most likely class. b) Prototypical network computes class-wise prototypes by taking the average of all support sample feature vectors from each class. Similarity is then computed between the class-wise prototypes and the query feature vector. c) Matching network directly computes the similarity between every pair of support and query feature vectors, but the support and query samples are projected using two different embedding networks. The network output is a linear combination of similarity between each support-query pair and class of the support sample.