Advances in the application of artificial intelligence-enhanced electrocardiogram to non-cardiovascular diseases

Abstract

The proliferation of electronic health records has prompted the development of healthcare big data and application of artificial intelligence (AI) for clinical diagnosis and treatment. AI-electrocardiogram (AI-ECG), valued for its non-invasive nature, operational simplicity, and broad accessibility, has demonstrated considerable utility in disease identification and management. Importantly, the scope of AI-ECG is not confined to cardiovascular diseases. A growing body of evidence highlights its additional value for non-cardiovascular conditions such as diabetes, liver diseases, kidney diseases, infectious diseases and electrolyte imbalances. This opinion review summarizes the research progress of AI-ECG in non-cardiovascular diseases and discusses its clinical implications to guide attention towards these novel applications.

Key Words: Artificial intelligence; Machine learning; Electrocardiogram; Application; Non-cardiovascular diseases

Core Tip: This article systematically reviews the latest advances in artificial intelligence-enhanced electrocardiogram for the diagnosis and evaluation of non-cardiovascular diseases. It summarizes its value in identifying and predicting risks across various conditions, including liver and kidney diseases, pulmonary embolism, and electrolyte disturbances. The opinion review also discusses the strengths, limitations, and future directions of this technology, aiming to inform clinicians and support clinical decision-making.

INTRODUCTION

The electrocardiogram (ECG) has long served as a fundamental tool in clinical diagnostics, primarily used to identify rhythm abnormalities such as arrhythmias and disturbances of cardiac conduction, as well as conditions such as myocardial infarction[1]. A broader physiological perspective, however, makes clear that the ECG is not merely an isolated record of cardiac activity. The tracing is shaped by the systemic internal environment because myocardial depolarization and repolarization are sensitive to changes in autonomic tone, serum electrolyte concentrations, acid-base balance, and endocrine states. In addition, many medications can modify ion channel behavior and thereby alter waveform morphology and timing[2]. Although such effects are often subtle, nonspecific, or treated as contextual variation in routine practice, they can represent measurable reflections of systemic physiology. The heart is embedded within neurohumoral and metabolic networks, and its electrical activity can register disturbances that arise well beyond the cardiovascular system[3]. Thus, ECGs have been used for the diagnosis, risk stratification and prognostication beyond cardiovascular diseases.

The rapid expansion of electronic health records and large-scale medical datasets has accelerated the application of artificial intelligence (AI), particularly machine learning, across clinical medicine, yielding the field of healthcare big data[4-6]. This data environment has also enabled the development of AI-ECG, in which models learn directly from raw voltage time signals rather than relying only on conventional, human defined measurements such as intervals and predefined morphological patterns. By analyzing the full waveform, AI-ECG approaches can integrate subtle morphology and temporal context that are difficult to capture with prespecified rules. The outputs are typically framed in forms that can support clinical decisions, including risk estimates and continuous surrogates that relate to laboratory values or other reference measures[1,7-9]. In parallel, expectations for credible progress have become more stringent. High performance in a single dataset is no longer sufficient. Contemporary studies increasingly require explicit phenotype definitions, validation in independent cohorts drawn from diverse populations, and a clear statement of intended clinical use so that model outputs can be interpreted, calibrated, and assessed within realistic care pathways[10].

Against this methodological backdrop, the scope of AI-ECG has expanded beyond classical cardiovascular endpoints to include non-cardiovascular diseases (non-CVD). In these settings, the near-term clinical application is more often supportive than substitutive. The goal is typically to aid earlier recognition, triage, and prioritization of confirmatory testing rather than to replace definitive laboratory assays or specialist assessment[11]. Early studies have reported potential utility in several non-CVD conditions, including electrolyte imbalance, chronic kidney disease (CKD), and thyroid dysfunction[12-14]. In this opinion review, we discuss the research by Karbovskaya et al[15] recently published in World Journal of Cardiology. The authors presented a model that accurately detects diabetes mellitus (DM) from a single-lead ECG. This study has contributed to the development of the AI-ECG model as a reliable screening tool for non-CVD in high-risk populations. Therefore, we summarize the latest advances of AI-ECG in non-CVD (Figure 1).

Figure 1 Potential applications of artificial intelligence-electrocardiogram model in non-cardiovascular diseases. AI-ECG: Artificial intelligence-electrocardiogram.

AI-ECG IN DETECTION AND MANAGEMENT OF NON-CVD

DM

DM is a clinical condition that causes a high amount of glucose in the blood with a simultaneous lack of insulin or insulin resistance. In 2024, approximately 589 million adults were living with diabetes, and this number is projected to rise to 853 million by 2050[16]. Given the large patient population and complications involving multiple systems, the development of diabetes management optimization tools holds great clinical value. A prior study with a small sample size proposed a machine learning model using discrete wavelet transform features extracted from ECG[17]. The findings indicated that this approach could be used to identification of patients with DM and healthy control. Decision tree reached an accuracy of 92.02%, sensitivity of 92.59% and specificity of 91.46%. Another study from Karbovskaya et al[15] demonstrates a model by identifying a specific clinical profile (high comorbidity burden with preserved cardiac function) could accurately detects DM from a single-lead ECG achieving robust performance [area under the curve (AUC): 0.880]. Patients with DM have significant fluctuations in their blood glucose levels. Both excessively high and excessively low blood glucose can lead to adverse clinical outcomes. Cichosz et al[18] developed a fuzzy reasoning method based on hybrid particle swarm optimization for detecting hypoglycemic events at night. The study included overnight ECG records of children with type 1 diabetes, correlating heart rate and QTc with hypoglycemic events. The results showed that the model was able to detect severe hypoglycemic events (with a sensitivity of 85.7% and a specificity of 79.87%). Conventional methods for diabetes testing and management, such as continuous glucose monitoring and hemoglobin A1c measurement, are invasive and often costly. A growing body of evidence suggests that ECG signals could serve as a convenient, non-invasive tool for screening and managing patients with DM.

Liver diseases

Previous studies have demonstrated the potential of machine learning in healthcare and highlighted the value of ECG in detecting systemic conditions[19]. Cardiovascular and hepatic diseases share several common pathophysiological mechanisms and pathways, and the heart and liver interact closely. For instance, in heart failure, cardiac dysfunction can lead to hepatic congestion, while liver impairment resulting in portal hypertension may further increase cardiac burden. Growing evidence suggests that AI-enhanced ECG holds promise for the prediction and management of liver disease. A machine learning model was developed using ECG features from two large-scale datasets, MIMIC-IV-ECG and ECG-View II, for the identification of liver diseases[20]. The model was trained and internally validated on over 450 thousand ECG records from MIMIC-IV-ECG, followed by external validation using more than 700 thousand records from ECG-View II. Their results demonstrated strong predictive performance and good generalizability of the model. For alcoholic liver disease, the AUC was 0.80 in internal validation and 0.76 in external validation. For liver failure, the AUCs were 0.74 and 0.75 in internal and external validation, respectively. Udompap et al[21] reported another AI-ECG model to predict metabolic dysfunction-associated steatotic liver disease by incorporating other clinical data including age, sex, coexisting diseases, and laboratory test results. Notably, the model showed consistent discriminative ability, with an AUC of 0.72 in the original cohort and 0.72 in the external validation cohort. Furthermore, the AI-ECG model demonstrates potential for risk stratification in patients with chronic liver disease (CLD)[22]. It achieved an AUC of 0.86 in distinguishing healthy controls from CLD patients, with a diagnostic odds ratio of 12.33 [95% confidence interval (CI): 11.16-13.63], indicating a strong ability to correctly identify CLD. Beyond identifying liver diseases, ECGs hold significant prognostic value for patients with advanced liver dysfunction. The AI-Cirrhosis-ECG score, a deep learning-based system utilizing convolutional neural networks[23], was developed to distinguish patients with cirrhosis. Notably, every 0.1-point increase in the AI-Cirrhosis-ECG score was independently associated with a heightened risk of liver-related mortality [hazard ratio (HR) = 1.44, 95%CI: 1.32-1.58, P < 0.001]. This collective evidence suggests that ECG data may emerge as a non-invasive biomarker for liver diseases, potentially attributable to the underlying cardio-hepatic physiological interplay. The AI-ECG represents a promising bridge between cardiology and hepatology. By enabling ECG-based screening prior to formal diagnosis, it supports targeted clinical assessment as well as early detection.

CKD

Renal dysfunction is a major determinant of poor prognosis in cardiovascular disease, associated with increased risk of cardiovascular death[24,25]. Researchers have recently developed AI-ECG models capable of detection and prediction of renal impairment, offering an innovative tool for enhancing risk stratification. In patients with CKD, AI-ECG facilitates a comprehensive evaluation, including arrhythmia detection, assessment of cardiac structural remodeling, monitoring of blood pressure, and identification of electrolyte imbalances[26]. An AI-ECG model developed by Holmstrom et al[12] was designed to both identify and stage CKD. This study incorporated 247665 ECG data from 111370 participants to develop and validate the model, and then conducted external validation using nearly 896620 ECG data from another health system. The model achieved AUC values of 0.75 for mild CKD, 0.76 for moderate-to-severe CKD, and 0.78 for advanced CKD. Furthermore, in the subset of patients under 60 years of age, high discriminatory performance was observed regardless of ECG input type, with respective AUCs of 0.84 for 12-lead and 0.82 for single-lead ECG. For patients with end-stage renal disease, kidney transplantation serves as a key therapeutic option to preserve kidney function. AI-ECG models provide significant benefits, however, it also brings clinical challenges such as infection and immune rejection, which critically influence transplant outcomes[27]. Therefore, the accurate prediction of long-term survival following kidney transplantation is vital for sustaining a balanced organ allocation system, which promotes equitable distribution and maximizes organ use. In this context, AI-ECG models might present a potential assistive tool. Preoperative ECGs from 6504 patients undergoing kidney transplant at three centers operated by Mayo Clinic were analyzed by 6 AI algorithms to predict biological age and various cardiovascular parameters, including atrial fibrillation, aortic stenosis, low ejection fraction, hypertrophic cardiomyopathy, cardiac amyloidosis. All algorithm outputs showed significant correlations with long-term all-cause mortality (P < 0.001)[28]. None of the patients had pre-existing cardiovascular disease at baseline. Consequently, AI-ECG demonstrates potential for predicting long-term cardiovascular risk and mortality in asymptomatic kidney transplant recipients, contributing to pinpoint those who would benefit from enhanced cardiac surveillance.

Pulmonary embolism

Pulmonary embolism (PE), a critical and often fatal manifestation of venous thromobosis, contributes significantly to cardiovascular morbidity and mortality in both the short- and the long-term[29]. Utilizing ECG and clinical data from over 20-thousand PE patients, Somani et al[30] developed a machine-learning model for acute PE detection. From internal validation, it achieved an AUC of 0.81. Further evaluation within the test set showed that this integrated model significantly outperformed common clinical scoring methods (Wells’ Criteria, Revised Geneva Score, PE Rule-Out Criteria, and 4-Level PE Clinical Probability Score), with an AUC of 0.84 compared to the scoring methods’ range of 0.50-0.58. Another research group developed a random forest model using only presenting ECGs to stratify PE risk[31]. The AI-ECG model accurately predicted clinical severity (low-risk vs severe PE) with an AUC of 0.83, and the identified risk groups showed significantly different 30-day and in-hospital mortality outcomes. The AI-ECG model showed superior performance to standard clinical classification in accurately screening and stratifying PE patients, while also predicting short-term mortality. It might represent a valuable tool in clinical risk assessment.

AI-ECG-BASED PREDICTION OF ELECTROLYTE ABNORMALITIES

Blood biomarkers serve as indicators of pathophysiological and metabolic homeostasis. Clinically, parameters like electrolyte levels are assessed via blood sampling to monitor disease status. Compared to invasive blood collection, surface ECG is non-invasive, quicker to perform, and more economical. Consequently, the AI-ECG-based prediction of blood biomarker levels represents a growing focus of investigative interest. Abnormalities in serum potassium can affect cardiomyocytes, and underlying acute cardiovascular disease could magnify the association between potassium abnormalities and adverse outcomes[32]. Lin et al[11] reported that blood potassium levels estimated from 12 single-lead ECGs showed a mean error of less than 0.37 mmol/L compared to laboratory measurements. The model demonstrated high diagnostic accuracy, with an AUC > 0.85 for predicting moderate-to-severe hypokalemia (Lab-K⁺ ≤ 3 mmol/L) and > 0.95 for hyperkalemia (Lab-K⁺ ≥ 6 mmol/L). Moreover, it carried significant prognostic value: Patients with abnormal ECG-K⁺ had elevated all-cause mortality risk, regardless of whether laboratory results were normal (HR = 1.4, 95%CI: 1.16-1.69 for low ECG-K⁺; HR = 1.49, 95%CI: 1.00-2.24 for high ECG-K⁺). These findings indicate that AI-ECG captures subtle physiological features that may help predict adverse clinical outcomes. An analysis[33] of more than 11000 CICU patients admitted to the Mayo Clinic demonstrated that the AI-ECG algorithm for hyperkalemia detection could consistently identify individuals with increased in-hospital and post-discharge mortality risk, thereby highlighting its potential as a prognostic biomarker derived from the ECG. The proliferation of smart wearable devices also enabled convenient collection of physiological signals including heart rate, rhythm, and blood pressure. Notably, a recent study demonstrated that an AI algorithm leveraging single-lead ECG data from smart wristbands can effectively predict blood potassium levels[34]. The model achieved AUC of 0.88 in the training set and 0.83 in the external validation set, showing performance comparable to methods using clinical-grade ECGs.

Calcium ions play key roles in cellular homeostasis and are another key electrolyte of clinical concern[35]. The Lin et al[13] extended their finding by developing an AI-ECG model to predict blood calcium levels. This model achieved a mean absolute error of 0.78-0.98 mg/dL in estimating albumin-adjusted calcium. For detecting severe abnormalities, it showed high diagnostic accuracy, with AUCs of 0.92/0.84 for hypercalcemia and 0.89/0.77 for hypocalcemia in the internal and external validation sets, respectively.

As a non-specific inflammatory marker, C-reactive protein (CRP) has been identified as a biomarker for various arrhythmias, including their prognostication[36,37]. A research team from Guangzhou has developed an AI-ECG algorithm to detect CRP levels in patients with sinus rhythm[38]. After training and validation, the AUC of model achieved 0.86 and 0.85 on validation set and test set, respectively. This work not only reveals that cardiac electrophysiological signals contain inflammation-related information but also offer a non-invasive method to screen for inflammatory status using CRP.

APPLICATIONS OF AI-ECG IN OTHER FIELDS

Syncope is a common symptom with a broad differential diagnosis, spanning from benign etiologies to life-threatening conditions. Its prognosis is highly heterogeneous, while conventional clinical prediction tools often fail to accurately stratify associated risks. This creates a critical need for rapid risk assessment to guide clinical treatment. Emerging evidence suggests that AI-ECG models, by integrating routine clinical data with ECG signals, can provide additional value for the screening of underlying causes and prognostication of syncope[39-41]. Moreover, using data from a single-center retrospective cohort, an AI-ECG model was developed for pneumothorax detection[42]. It demonstrated excellent performance, with an AUC of 0.95 using standard 12-lead ECGs input and 0.93 even when limited to a single-lead ECG. This high accuracy suggests its potential as a tool for early ECG-based screening within the medical system, which could enhance diagnostic workflows. Driven by advances in AI and the era of healthcare big data, AI-ECG models are increasingly applied beyond traditional cardiovascular management to the diagnosis, risk stratification, and prognosis assessment of diverse diseases. Their utility now extends to multiple non-cardiovascular domains, including screening for conditions such as hyperthyroidism[43], epileptic seizures[44], and obstructive sleep apnea[45], as well as assessing post-procedural mortality[46].

GAPS IN KNOWLEDGE AND FUTURE RESEARCH

This opinion review summarized the extensive application of AI-ECG in the identification and risk stratification of non-CVD, suggesting that ECG signals may become a new type of biomarker characterized by high accuracy and accessibility. Its clinical value extends beyond the traditional domain of CVD. However, several limitations should be recognized. The development and validation of AI-ECG models require large volumes of data. For certain fields with low incidence but high mortality rates, such as cancer therapy-related cardiotoxicity, further development is still needed. Second, ECG signals are influenced by various confounding factors, including patient age, gender, underlying comorbidities, medication history and individual differences during ECG acquisition. Due to the influence of a wide range of variables, the reliability of ECG signals for achieving specific and significant predictive purposes is often diminished. The application of AI-ECG in multi-system clinical management still requires verification through larger-scale, cross-ethnic studies. Moreover, like traditional AI algorithms, AI-ECG models suffer from a black-box nature, and their lack of interpretability may be one factor limiting their broader adoption. The development of explainable AI also represents a promising direction for future research. Finally, the efficacy of AI-ECG varies across different diseases, and the development and refinement of models will likely require ongoing research in the future.

CONCLUSION

AI-ECG models present a highly promising and accessible approach for screening and managing a range of non-CVD, offering particular utility in resource-limited settings. ECG-derived features, representing a novel class of biomarkers with substantial clinical relevance, can facilitate non-invasive diagnosis, enable early detection, and guide risk stratification prior to targeted clinical evaluation.