Towards digital twin-enabled perioperative care: extending the potential of multimodal deep learning for cardiovascular risk stratification

Dear Editor, We read with great interest the recent study by Yoon et al., which presents a sophisticated multimodal deep learning model for predicting 30-day major adverse cardiac and cerebrovascular events after non-cardiac surgery[1]. The integration of raw electrocardiogram (ECG) waveforms with minimal demographic and procedural data to achieve an impressive area under the receiver operating characteristic curve of 0.902 represents a significant advancement in perioperative risk stratification. The authors’ success in leveraging a transformer-based architecture for raw signal processing underscores the profound potential of artificial intelligence (AI) to extract subclinical, prognostic information beyond human interpretation. We wish to extend the discussion on the compelling clinical translation pathways this work opens, particularly within the evolving paradigms of digital health and digital twins. The model developed by Yoon et al. can be viewed as a foundational component for a perioperative digital twin. A digital twin is a virtual, dynamic representation of a physical entity or process, continuously updated with data to simulate, predict, and optimize outcomes[2,3]. In this context, the predictive model acts as the initial “state” of a patient’s cardiovascular risk twin at the preoperative moment. The logical evolution is toward a dynamic twin that assimilates continuous data streams from wearable sensors and electronic health records throughout the perioperative journey—from the preoperative clinic to the operating room, the post-anesthesia care unit, and the surgical ward. This would enable real-time risk trajectory monitoring, allowing for pre-emptive interventions should the model predict a deteriorating trajectory, thereby moving from static prediction to dynamic, personalized risk management. For broad clinical adoption, as rightly highlighted by the authors, the next critical step involves seamless integration into clinical workflows. The minimal data requirement is a key strength for scalability. The future lies in embedding such models as silent, decision-support engines within routine clinical systems. For instance, upon scheduling a surgery, the model could automatically analyze the patient’s existing preoperative ECG and demographic data in the background, flagging high-risk patients to the anesthesiology and surgical teams via the electronic health record. This facilitates proactive optimization of medical therapy, tailored anesthetic plans, or triage to higher-acuity monitoring pathways without adding bureaucratic burden. The ultimate goal is to create an “invisible AI” that augments clinical judgment efficiently. Looking forward, the research trajectory initiated by this study could be expanded through federated learning approaches. Federated learning allows model training across multiple institutions without sharing sensitive patient data, thereby enhancing generalizability and addressing potential biases from single-center datasets. Building a robust, multi-institutional model would be a monumental step toward developing a universally applicable digital tool for perioperative care[4]. In conclusion, the work by Yoon et al. is a commendable stride towards data-driven, precise perioperative medicine. It provides a robust algorithmic foundation upon which the broader frameworks of digital twins and integrated digital health can be constructed. The transition from a highly accurate predictive model to a clinically transformative tool hinges on its evolution into a dynamic system within interoperable digital ecosystems, ultimately fostering a new standard of proactive and personalized patient care. (We ensure this article is compliant with the TITAN Guidelines[5]).