Letter to the Editor: Do we need an artificial intelligence-assisted electrocardiographic tool to diagnose diabetes mellitus or to predict its unseen cardiovascular consequences?

Abstract

Diabetes mellitus (DM) has a significant negative impact on the global health. Its burden on the medical sector is tremendous as its complications affect all body organs across all age groups. The most dramatic complications of DM include cardiovascular and neurological disorders. Microvascular damage can start years before the diagnosis of type 2 diabetes (T2DM) is made; therefore, early screening is of utmost value. Moreover, subclinical electrocardiographic (ECG) changes are common in patients with T2DM without evident cardiac disease. I have read with great interest the recent study published in World Journal of Cardiology by Karbovskaya et al, on the utility of a single-lead ECG for diagnosing DM, using machine learning and multinomial regression. The utility of single-lead ECG for predicting glycemic levels appears questionable. It should be interpreted with caution, particularly in light of the model’s limited explanatory power and elevated maximum error. A single-center, non-randomized study with a small sample size in the DM groups and misclassification bias are limitations of the study. A simple, easily accessible tool for early detection or prediction of cardiac dysfunction in DM or in people at risk is more valuable than merely distinguishing healthy from DM or type 1 diabetes from T2DM. A single-lead ECG, especially with artificial intelligence, can flag risk and help bridge some gaps in predicting event risk. However, this study needs validation with a precise aim to predict high-risk diabetics and not only to diagnose DM.

Key Words: Diabetes mellitus; Electrocardiography; Artificial intelligence-assisted; Diagnostic tool; Machine learning; Cardiovascular; Photoplethysmography; Electrocardiographic

Core Tip: Diabetes mellitus (DM) has a significant negative impact on the global health. The most dramatic complications of diabetes include the microvascular damage that can start years before the diagnosis of type 2 diabetes (T2DM) is made; therefore, early screening is of utmost value. Moreover, subclinical electrocardiographic (ECG) changes are common in patients with T2DM without evident cardiac disease. A simple, easily accessible tool for early detection or prediction of cardiac dysfunction in DM or in people at risk is more valuable than merely distinguishing healthy individuals from those with diabetes or type 1 diabetes from T2DM. A single-lead ECG, especially with artificial intelligence, can red-flagging risk and help bridge some gaps in predicting event risk. However, this study needs validation with a precise aim to predict high-risk diabetics and not only to diagnose DM.

TO THE EDITOR

Diabetes mellitus (DM) has a significant negative impact on global health. Its burden on the medical sector is tremendous as its complications affect all body organs across all age groups. The critical step in the treatment of DM is its early recognition whenever possible, to follow a strict lifestyle and to have a regular follow-up. The most dramatic complications of DM include cardiovascular and neurological disorders. Microvascular damage can start years before diagnosis of type 2 diabetes (T2DM), as the metabolic dysregulation may start a decade or more before the overt “picture” of diabetes is clinically recognized[1,2]. This early microvascular damage can lead to changes in the heart's electrical pattern and myocyte function, even before overt symptoms and at the prediabetic stage, which can be detected by subclinical electrocardiogram (ECG) changes[3]. Therefore, early screening is of utmost value.

The oral glucose tolerance test is widely considered the gold standard for diagnosing diabetes and prediabetes, especially for identifying impaired glucose tolerance. However, the glycated hemoglobin (HbA1c) test is often used for general screening and monitoring due to convenience[4]. However, it performs variably across populations. The sensitivity of HbA1c varies widely and can be low at the standard ≥ 6.5% cutoff, while others find a higher sensitivity. The specificity of HbA1c is generally high (90% or more), indicating that few healthy people are misdiagnosed as diabetic with HbA1c ≥ 6.5%. The positive predictive value of this cutoff is low (39%-40%), whereas the negative predictive value is high (> 90%), indicating that a negative result (< 6.5%) is good at ruling out DM[5,6].

While HbA1c is highly valuable, machine learning (ML) models have shown that, in conditions like anemia, HbA1c can be falsely elevated, requiring ML-based reinterpretation for accurate diagnosis[7]. These models consistently identify HbA1c as the most significant predictor for both diagnosis and complication risk in DM. Advanced models, including deep learning (DL) and ensemble techniques, are increasingly used to detect early signs of DM and predict future HbA1c trends in alignment with the subject's electronic health records and emerging, noninvasive sensors[8]. Unlike conventional statistics, artificial intelligence (AI) models can uncover complex patterns and interactions within data that are not immediately apparent, thus providing a more nuanced understanding of the disease. Models such as random forest, support vector machines, and XGBoost have achieved accuracy ranging from 77% to 96%[9-11]. Shapley Additive Explanation analysis in ML consistently identifies HbA1c as the most important predictor of DM[12]. Moreover, DL has been used for advanced pattern recognition, such as analyzing 12-lead ECG to estimate HbA1c for DM screening, which can predict outcomes independently of laboratory-based tests[13]. A study leveraged photoplethysmography (PPG) signals from smartphones or wearables to predict HbA1c, achieving a high correlation with laboratory values (r = 0.56)[14].

In this regard, I have read with great interest a recent study on the utility of a single-lead ECG for diagnosing DM, using ML and multinomial regression[15]. The registered study number (NCT04788342) corresponds to a single-center, non-randomized, observational study aiming at evaluating the potential of a single-lead electrocardiogram monitor, CardioQvark with PPG function, for assessing left ventricular systolic function. However, DM was not among the primary and secondary outcomes. Notably, this raises concern regarding the alignment of the study's registered objectives with its reported outcomes[15]. World Journal of Cardiology recently published a study by Karbovskaya et al[15] analyzed data from 743 patients to assess the presence of DM and its two types with type 1 diabetes (T1DM) and 145 with T2DM. The diagnosis of DM was based on the subject’s medical history and random blood sugar; however, HbA1c was not measured in this study. A single-channel ECG and PPG were recorded using the CardioQvark device immediately after each blood glucose analysis. Recordings were then transmitted to a cloud-based platform for analysis using proprietary ML algorithms.

The authors concluded that there is an association between AI-assisted single-lead ECG parameters and diabetes status, supporting the clinical identification of individuals who would benefit from confirmatory testing. It showed that a shorter QTc interval was associated with a higher likelihood of T2DM [odds ratio (OR) = 0.35, P < 0.001], whereas prolonged QTc was associated with T1DM (OR = 3.81, P = 0.02). However, this is unusual finding that requires further elaboration, as DM is known to prolong QTc regardless of its type[16]. Both types of DM are linked to prolonged QTc, reflecting repolarization abnormalities, possibly due to cardiac autonomic neuropathy and poor glycemic control.

The study also showed that T-wave amplitude was associated with an increased likelihood of T2DM (OR = 11.2, P < 0.001); however, another recent study by Soflaei Saffar et al[17] found that some T-wave abnormalities were more frequent in DM but were not statistically associated with the detection of T2DM. The authors attributed these abnormalities to the possible underlying myocardial strain or altered repolarization gradients. The alterations in P wave parameters implicated atrial myopathy in diabetic patients in this study[15]. Two additional findings, which may not be clinically significant at the moment, showed that the highest AUC was observed for differentiating between DM types and non-diabetic conditions, and the poorest performance was observed in distinguishing T1DM from T2DM.

The utility of single-lead ECG for predicting glycemic levels appears questionable. It should be interpreted with caution, given the model’s limited explanatory power and elevated maximum error. The model’s performance for T2DM was weak, with low specificity and a high false-positive rate. Using a single rather than a 12-lead ECG could be a limitation, along with a small sample size in the DM groups and misclassification bias (as a lack of a gold-standard diagnostic test). However, single-lead AI-ECG can predict high risk patients[18]. Table 1 shows examples of published studies on the utility of AI-assisted ECG to detect DM[19-24]. Although these studies seem promising, there are many citations in the literature; mostly abstracts or congress proceedings only.

Table 1 Examples of artificial intelligence-assisted electrocardiogram for detection of diabetes mellitus.

Ref.	Technique	Comment
Zhang et al[19]	Training deep learning models for clinical screening from retrospective data presents a fundamental challenge: Our data contains only patients where both the ECG and HbA1c were measured	AI-enabled single-lead ECG (lead I), retains much of the discriminative performance (AUC, 0.78) and could be utilized to automate the identification of patients who are likely to have DM
Koga et al[20]	SHAP was applied to assess feature importance and clinical interpretability	A LightGBM-based algorithm (DiaCardia), achieved an AUR of 0.851. It can accurately identify subjects with prediabetes from an ECG alone, with performance that is robust across different patient cohorts and independent of major clinical confounders. A version of DiaCardia using only single-lead (lead I) ECG data achieved a comparable AUROC
Kulkarni et al[21]	1D convolutional neural network (CNN)	The model diagnosed DM from a single-lead ECG with 90.3% accuracy, 92.1% sensitivity, and 88.6% specificity. It identified ECG intervals (QT, QTc) as key biomarkers
Gragnaniello et al[22]	Dimensional Convolutional Neural Network (1D-CNN) to analyze the ECG signals directly on the device	The final model achieves an accuracy of 89.52%, with an average precision and recall of 0.91 and 0.90, respectively
Yildirim et al[23]	The one-dimensional hear rate signals were converted into frequency spectrum images, which were utilized for application to well-known pre-trained models, specifically: AlexNet, VggNet, ResNet, and DenseNet	DenseNet pre-trained model yielded the highest classification average accuracy of 97.62%, and sensitivity of 100%, to detect DM subjects via hear rate signal recordings
Cordeiro et al[24]	A deep learning architecture that can identify hyperglycemia using heartbeats from ECG signals	The high sensitivity and specificity achieved by 10-layer deep neural network showed that ECG possesses intrinsic information that can indicate the level of blood glucose concentration

1D: Single dimension of the input data; SHAP: SHapley Additive exPlanations; ECG: Electrocardiogram; AUC: Area under the curve; DM: Diabetes mellitus; HbA1c: Glycated hemoglobin; AUROC: Area under the receiver operating characteristic; AI: Artificial intelligence.

A simple, easily accessible tool for early detection or prediction of cardiac dysfunction in DM or in people at risk is more valuable than merely distinguishing healthy from DM or T1DM from T2DM. A single-lead ECG, especially with AI, can red-flagging risk and help bridge some gaps in predicting event risk. This study needs validation with a precise aim to predict high-risk diabetics and not only to diagnose DM. To make the call for validation more actionable, it is essential to encourage a multidisciplinary collaboration between cardiologists, endocrinologists, and data scientists, this can ensure a comprehensive approach to developing a reliable diagnostic tool that accurately identifies high-risk diabetics.