IKrNet: A Neural Network for Detecting Specific Drug-Induced Patterns in Electrocardiograms Amidst Physiological Variability

TL;DR

The problem is formalized as a binary drug-footprint detection task where $Y \in \{0,1\}$ indicates presence of drug effects on ECGs. IKrNet combines a multi-resolution CNN backbone for spatial feature extraction with BiLSTM temporal modeling to capture time-evolving drug-induced patterns, and is evaluated under a fairness-inspired robustness framework across protocol zones. The study demonstrates that IKrNet substantially outperforms state-of-the-art DenseNet architectures across sampling rates and datasets, with improved stability under heart-rate variability and stress conditions, supporting its clinical viability for real-world ambulatory ECG analysis. The work provides a standardized robustness evaluation approach, proven to generalize beyond clean clinical data to noisy Holter recordings, enabling more reliable TdP risk assessment in routine monitoring.

Abstract

Monitoring and analyzing electrocardiogram (ECG) signals, even under varying physiological conditions, including those influenced by physical activity, drugs and stress, is crucial to accurately assess cardiac health. However, current AI-based methods often fail to account for how these factors interact and alter ECG patterns, ultimately limiting their applicability in real-world settings. This study introduces IKrNet, a novel neural network model, which identifies drug-specific patterns in ECGs amidst certain physiological conditions. IKrNet's architecture incorporates spatial and temporal dynamics by using a convolutional backbone with varying receptive field size to capture spatial features. A bi-directional Long Short-Term Memory module is also employed to model temporal dependencies. By treating heart rate variability as a surrogate for physiological fluctuations, we evaluated IKrNet's performance across diverse scenarios, including conditions with physical stress, drug intake alone, and a baseline without drug presence. Our assessment follows a clinical protocol in which 990 healthy volunteers were administered 80mg of Sotalol, a drug which is known to be a precursor to Torsades-de-Pointes, a life-threatening arrhythmia. We show that IKrNet outperforms state-of-the-art models' accuracy and stability in varying physiological conditions, underscoring its clinical viability.

Figures (9)

• Figure 1: Clinical protocol setup (the GENEREPOL study): The heart rate is depicted for each of the 10s ECGs, recorded with a Holter device for one patient. Each ECG is depicted by a dot, which is coloured in grey when not used during the training, in green and blue respectively before drug intake (Sot-) and after drug intake (Sot+) for the Holter dataset and in orange and blue for the expert hand-picked dataset. Because of the unbalanced ECGs before and after drug intake, we selected a proportionate amount of ECGs for the training process (cf. \ref{['sec:dataset']})
• Figure 2: (a) Global architecture of IKrNet, featuring a CNN backbone with multiple branches, each capturing low- and high-frequency features through different receptive fields. (b) Detailed branch structure with inverted residual blocks, optimizing feature extraction and reducing complexity.
• Figure 3: Cross-dataset evaluation of IKrNet and DenseNet on different sampling rates (a), and across full protocol at 500 Hz (b)
• Figure 4: Results on Generepol-Holter (ECGs at 500 Hz) over protocol zones and patients. Number of patients (vertical axis) achieving accuracy above the threshold (horizontal axis).
• Figure 5: DenseNet$_{500}^{\text{HP}}$ and IKrNet$_{+}$ on single patient and influence of heart rate variability across time:DenseNet$_{500}^{\text{HP}}$ is negatively affected by sampling rate changes and stress zones where heart rate increases, whereas IKrNet$_{+}$ is more robust.