TF-LSNet: Lightweight Heart Sound Classification via Large-Small Receptive Field Modeling and Reinforced Distillation
MCML Authors
Abstract
Abstract
This study aims to develop an accurate, ultra-lightweight, and practically deployable heart sound classification framework for heterogeneous mobile devices. We propose a Time-Frequency Large-Small Network (TF-LSNet) built on a Time-Frequency Large-Small Convolution (TF-LSConv), which decouples long-range contextual modeling from local feature aggregation. Two variants are designed for different resource budgets: Large-TF-LSNet for accuracy-oriented deployment and Small-TF-LSNet for highly constrained devices. To recover the accuracy lost through extreme compression, we further introduce Physiology-Aware Dual-Teacher Knowledge Distillation with Reinforcement Learning (PADKD-RL), which adaptively assigns sample-wise teacher weights using model-state and heart-sound physiological information. The models are evaluated on PhysioNet 2016, ZCHSound, a noise-corrupted PhysioNet set, and a cross-dataset benchmark, and are deployed in an Android application. Large-TF-LSNet contains only 47.7K parameters and requires 37.7M floating-point operations (FLOPs), achieving an accuracy of 93.19±1.52% and an unweighted average recall (UAR) of 93.37±1.34% on PhysioNet 2016 and 93.58±3.15%/93.06±3.28% on the high-quality ZCHSound subset. With only 4.7K parameters and 8.9M FLOPs, PADKD-RL-enhanced Small-TF-LSNet achieves 93.36±1.76% accuracy and 92.49±1.46% UAR on PhysioNet 2016 while maintaining low on-device inference latency. These results demonstrate a favorable balance among classification performance, computational efficiency, robustness, and mobile deployability.
article MLZ+26a
IEEE Internet of Things Journal
Early Access. Sep. 2026.Authors
X. Ma • E. Li • H. Zhang • K. Qian • B. Hu • B. W. SchullerLinks
DOIResearch Area
BibTeXKey: MLZ+26a