Explainable ECG Time Series Classification

Introduction Deep learning models can classify electrocardiogram time series with high accuracy, yet their decision-making processes are often opaque, which poses challenges for interpretability, clinical trust, and regulatory acceptance. This thesis examines whether post-hoc explainability methods can make predictions from a one-dimensional convolutional neural network more transparent and usable in an electrocardiogram classification context. Research question The research question guiding this thesis is: How do post-hoc explainability methods, specifically Local Interpretable Model-Agnostic Explanations and Shapley Additive Explanations, manifest explainability in terms of fidelity and stability, and how do they influence perceived transparency and usability among domain experts when explaining a one-dimensional convolutional neural network for electrocardiogram time series classification? Method An explanatory sequential mixed-methods design was applied. In Phase 1, a one-dimensional convolutional neural network was trained and evaluated on a standardized electrocardiogram dataset, and instance-level explanations were generated using the two post-hoc methods. Explanations were evaluated using fidelity, operationalized as the drop in prediction confidence when masking the most influential signal segments, and stability, operationalized as the similarity of explanations under small input perturbations. In Phase 2, explanation visualizations were used as stimuli in semi-structured video interviews with four ambulance nurses to assess perceived transparency and usability. Results The quantitative results showed that the trained model achieved strong predictive performance. The explanation methods exhibited complementary strengths: Shapley Additive Explanations produced more faithful explanations in terms of prediction consistency under masking, while Local Interpretable Model-Agnostic Explanations were more stable under small perturbations. The qualitative results indicated that domain experts considered explanations valuable for verification and trust calibration when they were visually clear and clinically plausible. However, ambiguous or misleading explanations were perceived as potentially risky and may increase the likelihood of confirmation bias. Conclusions The findings suggest that post-hoc explanations can enhance perceived transparency and usability for electrocardiogram classification, but their usefulness is conditional. Effective decision support requires explanations that are visually unambiguous, coherent with the model output, and aligned with clinical reasoning. The results also indicate a practical trade-off between fidelity and stability, implying that explanation method choice should be guided by the intended clinical use and tolerance for noise sensitivity.