Human-like visual computing advances explainability and few-shot learning in deep neural networks for complex physiological data

TL;DR

This work tackles data efficiency and interpretability in deep neural networks applied to complex physiological signals by using a perception-informed pseudo-colouring scheme that encodes clinically salient temporal features, notably the $QT$-interval, into ECG image representations. It combines this input encoding with $2$-way $5$-shot and $2$-way $1$-shot prototypical networks on a ResNet-18 backbone to classify drug-induced long QT syndrome cases from a dataset of ECGs, leveraging rhythm-level representations and LIME-based explainability. Results show substantial improvements in sensitivity, specificity, and accuracy under one-shot and few-shot settings, with explainability analyses indicating attention to clinically meaningful regions such as the T-wave morphology and QT duration. Aggregating multiple heartbeats into rhythm representations further boosts performance, aligning machine perception with human clinical practice and suggesting broader applicability to other heterogeneous biomedical signals.

Abstract

Machine vision models, particularly deep neural networks, are increasingly applied to physiological signal interpretation, including electrocardiography (ECG), yet they typically require large training datasets and offer limited insight into the causal features underlying their predictions. This lack of data efficiency and interpretability constrains their clinical reliability and alignment with human reasoning. Here, we show that a perception-informed pseudo-colouring technique, previously demonstrated to enhance human ECG interpretation, can improve both explainability and few-shot learning in deep neural networks analysing complex physiological data. We focus on acquired, drug-induced long QT syndrome (LQTS) as a challenging case study characterised by heterogeneous signal morphology, variable heart rate, and scarce positive cases associated with life-threatening arrhythmias such as torsades de pointes. This setting provides a stringent test of model generalisation under extreme data scarcity. By encoding clinically salient temporal features, such as QT-interval duration, into structured colour representations, models learn discriminative and interpretable features from as few as one or five training examples. Using prototypical networks and a ResNet-18 architecture, we evaluate one-shot and few-shot learning on ECG images derived from single cardiac cycles and full 10-second rhythms. Explainability analyses show that pseudo-colouring guides attention toward clinically meaningful ECG features while suppressing irrelevant signal components. Aggregating multiple cardiac cycles further improves performance, mirroring human perceptual averaging across heartbeats. Together, these findings demonstrate that human-like perceptual encoding can bridge data efficiency, explainability, and causal reasoning in medical machine intelligence.

Figures (6)

• Figure 1: Few-shot learning of single heartbeat showing normal QT-interval with no risk of TdP. With pseudo-coloring, the model focuses on both the T-wave morphology and duration (represented by the color), with more emphasis on the T-peak-T-end duration (known to be a significant clinical feature in assessing the risk of LQTS bhuiyan2015tmorin2012relationships). Without pseudo-color, the model focuses mainly on the T-wave morphology.
• Figure 2: Few-shot learning of single heartbeat showing an abnormal QT-interval prolongation at risk of TdP. With pseudo-coloring, the model focuses mostly on the T-wave area (which represents the T-wave morphology as well as the prolonged QT duration shown by the yellow-orange-red color gradient). Whereas without pseudo-color, the model incorrectly focuses on other irrelevant ECG features, in addition to the T-wave morphology.
• Figure 3: Few-shot learning of a 10-second heart rhythm image representation showing normal QT-interval with no TdP risk. Through pseudo-coloring, the model extracted the most clinically significant features (T-wave morphology and duration) and filtered out irrelevant signal features. Without color, the model extracted several clinically irrelevant features - hindering machine abstract and causal reasoning.
• Figure 4: Few-shot learning of a 10-second heart rhythm image representation showing abnormal QT-interval at risk of TdP. Using pseudo-coloring, the model extracted the most clinically significant features (T-wave morphology and duration) and filtered out irrelevant signal features. Without color, the model extracted several clinically irrelevant features - hindering machine abstract and causal reasoning.
• Figure 5: One-shot learning of a 10-second heart rhythm image representation showing normal QT-interval with no TdP risk. Through pseudo-coloring, the model extracted the most clinically significant features (T-wave morphology and pseudo-colored duration) and filtered out clinically irrelevant signal features. Without pseudo-color, the model completely failed to extract any features that were clinically significant.