Toward dynamic graph representation of 12-lead ECG in multimodal surgical risk assessment

To the Editor, We have reviewed with great interest the recent study by Yoon et al, which introduces a multimodal deep learning model that integrates raw electrocardiogram (ECG) waveforms with clinical data for postoperative risk prediction[1]. The authors’ use of a Transformer-based architecture to capture temporal dependencies in 12-lead ECG signals represents a meaningful step forward from conventional feature-based methods. The model’s performance clearly demonstrates the value of leveraging raw physiological data in this domain. This correspondence adheres to the TITAN guidelines[2]. Limitations of the current modeling approach We would like to draw attention to a key representational constraint in the proposed framework. The model currently treats the 12-lead ECG largely as a set of independent time series, with relationships between leads inferred only indirectly through a “local lead-attention” mechanism within the Transformer. This approach does not fully account for the biophysical nature of the ECG as a spatially embedded sensor array. Each of the 12 leads provides an anatomically grounded perspective on the heart’s electrical activity. Pathological changes, such as ischemia or conduction abnormalities, are reflected not only in individual lead waveforms but also, and importantly, in altered spatiotemporal patterns of signal propagation across the entire lead network. Toward a graph-based representation We suggest that modeling the 12-lead ECG as a dynamic functional connectivity graph could yield deeper physiological insight. In this representation, leads would serve as nodes, and time-varying measures of association between them would define the edges. Such a graph would capture the integrity of the cardiac electrical network, where, for example, anterior ischemia might reduce connectivity between precordial and lateral leads, and conduction blocks could produce distinctive disruptions in network topology. Analyzing this type of data with a spatiotemporal graph neural network (ST-GNN) would enable simultaneous learning of node-level features, edge strength (inter-lead connectivity), and global network embeddings. A hybrid architecture combining graph-based and Transformer components could be particularly effective: a graph construction module could convert ECG segments into temporal graph sequences; an ST-GNN encoder could learn the evolving network structure; and a subsequent Transformer could integrate these dynamic graph representations over time before merging them with clinical variables[3]. Potential benefits of a graph-based framework This approach offers several potential advantages. By explicitly accounting for spatial relationships between leads, it may improve representational capacity and uncover novel “connectivity biomarkers.” It could also enhance interpretability, allowing clinicians to visualize “high-risk network patterns” and relate model predictions more directly to cardiac anatomy and pathophysiology[4,5]. Moreover, because information is distributed across a network, such a model may be more resilient to noise or artifacts in individual leads. Beyond these immediate gains, adopting a graph-based perspective supports a shift from static risk scoring toward dynamic system assessment. Rather than simply detecting isolated abnormalities, we could interrogate properties of the heart’s electrical network, such as its stability or resilience, under stress. Surgical intervention could be framed as a perturbation to this network, with postoperative outcomes depending on the network’s ability to withstand or recover from that perturbation. In addition, a well-defined graph structure could serve as a unifying layer for multimodal data integration. Clinical context could inform expectations about network vulnerability, while the ECG-derived graph could be related to other physiological networks, supporting a more holistic, system-level interpretation aligned with clinical reasoning. Outstanding challenges and future steps We recognize that several practical challenges remain, including how best to define connectivity metrics and how to process dense temporal graphs efficiently. Future efforts might explore adaptive graph learning techniques and develop visualization tools to support clinical translation. The broader pursuit of physiologically informed representations remains an active area in medical AI. Just as recent advances in imaging, such as the Singular Pooling method, which uses spectral decomposition to preserve critical information in sparse feature maps for ultrasound analysis[4], have refined feature extraction in convolutional networks, similar principled innovations are needed for representing relationships in waveform data. Combining explicit spatial modeling via graph networks with the strong temporal modeling capabilities of Transformers represents a promising direction forward. Concluding remarks We believe that closer collaboration across computational cardiology, graph machine learning, and perioperative medicine will be essential to develop the next generation of clinically robust, interpretable, and physiologically coherent risk prediction models[6,7].