Dear Members of the ESC Working Group on e-Cardiology,
For decades, our understanding of patient trajectories has been constrained by ‘episodic’, clinic-based assessments. Today, however, we are armed with an unprecedented stream of high-fidelity data, driven to a great extent by the ubiquity of wearable technology. Yet, as this month’s curated literature demonstrates, the raw acquisition of continuous data is merely the first step. The true transformative power of this era lies in our ability to feed these rich data streams into advanced systems, nowadays powered by artificial intelligence (AI), to predict outcomes, stratify risk, and ultimately dictate precise clinical interventions. In this edition, we explore the backbone of this new paradigm through Stefan Busnatu’s sharp commentary on wearable technology, alongside three recent original research articles from the ESC journal family that illustrate the deployment of AI across the acute and chronic care continuum.
In this month’s featured commentary, Stefan Busnatu contextualises the State-of-the-Art Review by Hughes et al., providing a critical lens on the integration of consumer-grade wearables into clinical practice. Busnatu presents the core thesis: wearables now offer a highly credible, scalable alternative to traditional monitoring, capturing everything from photoplethysmography to remote dielectric sensing. However, this must not lead to uncritical adoption: AI and wearable models trained on non-representative, affluent populations can risk degrading diagnostic performance for the disadvantaged demographic subgroups that already suffer the highest cardiovascular burden. Furthermore, as highlighted by the incidental arrhythmias detected in population-scale studies, deploying these technologies without structured diagnostic pathways simply generates an “incidentalome” of unconfirmed alerts, overwhelming our clinical infrastructure. Wearables are not standalone solutions; they are facilitators that demand robust integration into routines and clinical workflow.
If wearables capture the continuous outpatient trajectory, AI models are now redefining how we handle the acute influx of patients and patient data. Emergency department (ED) overcrowding is a critical public health issue that drives up cardiovascular mortality, largely due to the error-prone nature of conventional human triage. To address this bottleneck, Bavali-Gazik and colleagues recently introduced an AI-based electronic triage model tailored for cardiac-suspected patients.
Moving beyond simple clinical algorithms, the investigators engineered a fusion model that integrates routine triage data with automated convolutional neural network (CNN)-derived ECG interpretations. This approach yielded exceptional precision (97.22%) and an area under the curve (AUC) up to 0.938 for predicting major adverse events, significantly outperforming the conventional Emergency Severity Index (ESI). While the predictive superiority of the fusion model is undeniable, its real-world clinical applicability might still face hurdles, as commonly observed with retrospective validation that may overstate the performance in a highly skewed, real-world ED population.
Another field of AI-predicted outcomes is that of the acute setting of a myocardial infarction (MI), and the long-term trajectory of MI patients. Historically, we have relied on rigid risk scores and isolated metrics, such as left ventricular ejection fraction (LVEF). Two recent studies powerfully illustrate how AI can synthesise deeper phenotypic data to stratify risk with unprecedented accuracy. Scanlon et al. investigated the prediction of cardiovascular death following MI using Gradient Boosted Cox models. Crucially, they proved that integrating comprehensive echocardiographic data provides significant incremental value over models relying solely on clinical variables and LVEF (C-index 0.861 vs. 0.792). In fact, 14 of the top 20 features driving the model’s predictions were echocardiographic variables, heavily emphasising LV size, mass, and diastolic parameters like mitral E-wave velocity.
Complementing this, Xue and colleagues introduced the AIMI model, utilising Random Forest algorithms to predict long-term all-cause mortality after an acute MI. The AIMI model incorporated 15 readily available variables (including eGFR, d-dimers, and Killip classification) and significantly outperformed traditional TIMI and GRACE risk scores. The strength of both the Scanlon and Xue models lies, at least partly, in their interpretability. By utilising SHapley Additive exPlanations (SHAP), these models discard the “black box curse”, providing clinicians with individualised risk “weightings”. For instance, the AIMI model demonstrates that the prognostic weight of a feature like diastolic blood pressure or BMI is not static; it dynamically shifts based on the individual patient’s unique physiological profile. However, the practical challenge remains data harvesting. The true clinical utility of these models will only be realised when they are embedded directly into clinical routines, automatically pulling comprehensive echo parameters and lab values without adding data-entry burdens to the clinician.
Wearables and continuous sensors are transforming our era, offering an unprecedented, high-definition stream of ambulatory biological data. This rich continuous substrate, when combined with “deep” phenotyping acts as the ideal fuel for sophisticated AI architectures. These models can distil overwhelming clinical complexity into precise, individualised risk predictions, yet with the typical caveats, mainly concerning applicability and generalisability. The technology is no longer the rate-limiting step - our clinical infrastructure and adaptability is.
Our mandate as the e-Cardiology community is to re-design interoperable EHR pathways, AI governance frameworks, and clinical workflows required to translate these algorithmic insights into decisive, equitable action at the bedside.
On behalf of the ESC Working Group on e-Cardiology,
Panteleimon Pantelidis,
Nucleus Member ESC WG on e-Cardiology 2024-2026
