Abstract
Abstract
This study investigates how Kolmogorov–Arnold Networks (KANs) behave when
combined with semi-supervised learning for single-lead ECG analysis. Three architectures
are compared: a standard CNN-BiLSTM-Attention model (accurate but uninterpretable),
a minimal KAN with a simple CNN front-end (ablation), and the proposed
SE-MSC-KAN-CNN. The proposed model uses multi-scale depthwise convolutions,
Squeeze-and-Excitation attention, BiLSTM, and a B-spline KAN head. Interpretability is
evaluated in three ways: visualising learned B-spline functions, applying Grad-CAM to
highlight important signal regions, and measuring how the splines change when fewer
labels are used (spline drift). On the PhysioNet 2017 dataset (8,528 recordings, four
classes), the proposed model reaches 85.2% supervised accuracy (F1=0.851),
outperforming the CNN baseline (83.1%) and matching much larger black-box models.
With only 90% of the labels, it still achieves 85.1%. After training, we applied
post-training TFLite INT8 conversion. Only the CNN-front-end KAN converted
(measured size 422 KB). The proposed SE-MSC-KAN-CNN did not convert, because
Bidirectional LSTM and custom attention lack TFLite kernels; its float32 footprint is
about 9.8 MB (theoretical INT8 size about 806 KB). Classification metrics in this paper
are float32; INT8 test accuracy was not measured. The spline drift metric summarises
how KAN coefficients change under semi-supervised training; use as an on-device
out-of-distribution detector is proposed but not validated here. These results show that
competitive accuracy and inspectable KAN maps can be obtained in one hybrid
architecture. Wearable deployment of the proposed SE-MSC-KAN-CNN is not yet
achieved: TFLite INT8 conversion failed, and the present graph still requires float32
inference.