A Few-Shot Swallowing Event Recognition Method Based on H-ProtoNet

Early screening for dysphagia is crucial to reducing the risks associated with swallowing disorders and minimizing resultant complications. Accurate recognition of swallowing events (SEs) is a key in the early screening and treatment process. Deep learning technologies for swallowing event recognition are extensively utilized. However, their implementation in clinical settings is hindered by a scarcity of labeled samples, significantly restricting their practical efficacy in medical applications. To improve the accuracy of swallowing event recognition under conditions of limited data, this paper introduces an innovative Prototypical Network, termed the Heterogeneous Prototypical Network (H-ProtoNet). Initially, the Markov Transition Field (MTF) transformation converts one-dimensional swallowing signals into two-dimensional image representations, enhancing the fine-grained representation of heterogeneous swallowing characteristics. A shallow feature expression module is incorporated into the H-ProtoNet to swiftly capture the local characteristics of swallowing signals. Additionally, a heterogeneous feature metric module is employed to effectively measure diverse swallowing characteristics, thereby enhancing the model’s precision in identifying varied swallowing features. Validated by swallowing data in daily environments, the method proposed in this paper achieved 91.24% accuracy in swallowing event recognition, which is an improvement of 5.06% compared to the baseline model. The results demonstrate that the method described in this paper significantly improves the accuracy of swallowing event recognition in a few-shot setting, effectively supporting clinical practice.