The feasibility of early detecting coronary artery disease using deep learning-based algorithm based on electrocardiography

Introduction

Cardiovascular diseases (CVDs) continue to be the leading cause of death worldwide, accounting for approximately 33% of global mortality in 2019 [1, 2]. In China, nearly 3.5 million people die each year from CVDs or CVD-related diseases, accounting for up to 45.0% of deaths in rural areas and 42.6% in cities [3, 4]. The most common type of CVDs was coronary artery disease (CAD), which is the clinical consequence of atherosclerotic plaque formation, resulting in reduced blood supply to the distal myocardium, i.e., ischemia [5, 6]. Despite this, CAD patients may go years without a diagnosis because there are only minor or non-identifiable symptoms in the early stages [7]. There is a high prevalence of asymptomatic CAD in diabetics, particularly those with type 2 diabetes mellitus [8]. CAD continues to be a substantial cause of poor life quality and increased healthcare expenditure and as a result, there is still an urgent unmet need to develop a new strategy for early detection of CAD, which is a common challenge in the world that the clinicians are currently facing [9, 10].

Traditionally, for patients with suspected CAD, invasive coronary angiography (CAG) or coronary computed tomography angiography (cCTA) has been the preferred method for diagnosis [11]. Routine screening with CAG or cCTA, on the other hand, is not justified for patients with transient episodes of chest pain or mild symptoms due to cost, time consuming, invasive procedural risks, radiation, and potentially nephrotoxic iodinated contrast exposure [12, 13]. The conventional 12-lead electrocardiogram (ECG) is a low-cost, widely used, and non-invasive medical tool used on patients for both cardiac and noncardiac reasons [14]. In the clinical setting, electrophysiologists or cardiologists examine both rhythmic and morphologic ECG abnormalities. However, analyzing a wide range of ECGs from a variety of patients is time-consuming and highly dependent on the expertise of the individual. In addition, in the early stages of cardiac dysfunction, patients typically present mild signs and symptoms or are even asymptomatic, resulting in only minor changes in ECG [15, 16]. Until now, there have been no widely used screening strategies for patients with silent or minimally symptomatic CAD.

Deep learning (DL), which has revolutionized artificial intelligence (AI), is currently the most advanced branch of machine learning (ML) [17, 18]. It has made remarkable progress in a wide range of applications, including computer vision, speech recognition, and natural language processing [19]. DL has a powerful potential for automatically discovering the abstract representations required for prediction from the raw data without substantial information loss [20–23], which has resulted in promising achievements in medical fields, such as disease screening, diagnosis, and prediction. CNNs are one of the most successful DL architectures which have received enormous attention in the fields of medicine in the last 5 years [19]. The CNN application for digital ECG data could detect subtle changes in ECGs that were related to cardiac structural or functional abnormalities. A recent Lancet study, for example, had shown that an AI-enabled ECG could be used for recognition of patients with paroxysmal atrial fibrillation (AF) during sinus rhythm based solely on ECG [15]. In current study, we hypothesized that a novel approach based on DL methods could aid in detecting patients with CAD using only a conventional 12-lead ECG.

Results

Discussion

In this study, we developed a deep learning-based model to detect CAD using only ECG data. In the test dataset, our model using 8 leads performed moderately in detecting CAD, with an AUC of 0.75 (95% CI, 0.73 to 0.78), accuracy of 70.0% (95% CI, 69.5% to 70.1%), sensitivity of 68.7% (95% CI, 68.1% to 68.9%), specificity of 70.9% (95% CI, 70.3% to 71.0%), PPV of 61.2% (95% CI, 60.6% to 61.2%), and NPV of 77.2% (95% CI, 76.8% to 77.3%). In order to further explore the ability of the AI model to identify CAD and consider the practical application of the model, we also trained the CNN network on 8 single-lead (lead I, II, V1-V6) and also evaluated the AI model’s ability to identify CAD in the test dataset using every single-lead. However, we found that the AUC of the model using single lead to identify CAD is generally low. Although the model of V3 and V4 lead achieved an AUC of 0.73, the AUC was 0.02 higher using 8 leads of the ECG and we believe that this difference is mainly due to the following reasons: first, when CAD is in its early stages, the ECG may only show minimal alterations even normal, in other words, there are no obvious characteristics of ECG changes in the early stage of CAD. Therefore, the model may not be able to extract valuable features from single-lead ECG only. On the contrary, we believe that 8 leads can provide a richer dimension and more information, and the model can learn more valuable features, so as to produce a better recognition result. In addition, deep learning often requires a large amount of data. After increasing the amount of data, the model using single lead may extract more valuable features, and may also achieve a better recognition ability. Nonetheless, these findings demonstrated possibility and feasibility of developing a well-trained CNN model to detect CAD patients using only a standard 12-lead ECG.

CAD is a common disease that continue to be the leading cause of mortality and morbidity worldwide, imposing enormous health and economic burdens. Delayed diagnosis and treatment of CAD can have significant consequences for patients, such as myocardial infarction (MI) and ischemia-induced congestive heart failure (CHF), which is linked to a lower quality of life [24]. As a result, for CAD patients, timely diagnosis and early intervention are critical factors in improving prognosis.

Many cardiac imaging technologies, such as transthoracic echocardiography (TTE), magnetic resonance imaging (MRI) and myocardial perfusion imaging, have been widely used to detect cardiac function or structure injury. These methods, however, are time consuming and expensive, making them unsuitable for CAD screening. Coronary angiography (CAG) has long been regarded as the ‘gold standard’ for detecting CAD. However, only patients with typical clinical signs and symptoms or high risks for CAD were ever referred to CAG for further evaluation and treatment [25]. As a result, in the current study we defined CAD based on cCTA to reduce the selection bias.

ECG is a fundamental tool in the everyday practice of clinical medicine due to its wide availability and low cost and has frequently been used to non-invasively detect and diagnose cardiac arrhythmia and acute coronary syndromes. When CAD is in its early stages, the ECG may only show minimal alterations and are difficult to detect by the naked eye. We hypothesized that AI could assist us in diagnosing asymptomatic CAD by detecting minor alterations and delivering relevant information.

To test our idea, we created a novel approach based on deep learning algorithm as a reliable CAD screening tool. DL is a collection of new techniques. It has shown advantages of automatically discovering best representations needed for classification task from raw data without handcrafted feature engineering. The greatest advantage of DL is its ability to process various complex data, such as time-series data, 2D data and images. In the CNN algorithm, we used the raw ECG data (numerical matrix, 4000 × 8) rather than the ECG image itself to ensure that the complete information of the ECG raw signal was fully utilized. Although more computational power was needed to process and use the raw signal for the CNN in the model’s training process, we could use the features of ECG itself which were extracted by the CNN model over human bias. Similar to our use of ECG raw data for the diagnosis of CAD, a research from the Mayo Clinic [23] showed that a CNN model based on DL technology was able to detect patients with asymptomatic left ventricular dysfunction (ALVD), defined as ejection fraction ≤35% by echocardiography, using the ECG data alone and demonstrated its feasibility. Indeed, the ECG is an excellent substrate for AI applications, which yields reproducible and large amounts of raw data, such as portable wearable ECG device. Importantly, such large amounts of ECG data are relatively easy to store and converted to a digital format [17].

In fact, the unprecedented progress in DL technology have enabled computers to make an accurate automatic interpretation and diagnose of ECG in the early stage of the disease. Galloway et al. [26] developed an AI–ECG model to screen hyperkalemia in patients with chronic kidney disease. At an appropriate cut-off point, the model yielded values for the sensitivity and specificity of 90%, 89% respectively in a multicenter, external validation. Diagnosis delay is a clinical conundrum in acute ST-segment elevated myocardial infarction (STEMI). Base on ECG alone, Zhao et al. [22] created an AI-based STEMI autodiagnosis algorithm to identify STEMI, which achieved an AUC of 0.9954 with high accuracy, specificity, and sensitivity of 99.01%, 99.20%, 96.75% respectively, in the external evaluation. In our study, considering that the ECG showed no any specific changes and may only occur minor alterations in the early stages of CAD. Therefore, in order to improve diagnostic accuracy, we embedded the SE module into the original ResNet-50 network structure, which considered the interdependence of the model channels and could reinforce the extracted important features. To our knowledge, such model construction method is rarely seen in the study of its kind. Although compared with these similar studies, our model only yielded moderate performance (AUC 0.75). We thought it may be related to the following reasons. First, whether hyperkalemia or STEMI, the ECG features of this kind of disease are often more obvious, and the model is easier to extract such abnormal features. Second, as shown in Table 1, the ECGs we enrolled were from patients with various diseases in different departments. The underlying disease may also cause ECG changes, which might interfere with the model extracting ECG features related to CAD. In addition, non-specific changes in ECG such as ST-segment changes, flat T, T wave inversion, and so forth may affect or conceal the ECG features related to CAD, thereby affecting the effectiveness of the model to screen for CAD. Nevertheless, we still believe our research holds significant implications, because it is almost impossible for clinicians to judge whether a patient has potential CAD based only on the ECG.

The traditional automatic diagnosis of ECG mainly relies on ML algorithms and is mainly used in the diagnosis of acute coronary syndrome and arrhythmia [18]. However, the traditional ML algorithm mainly uses artificially extracted time or ECG morphological features as input, such as QT interval, RR interval or P wave amplitude, ST segment elevation et al. Advanced, it is not only time-consuming and labor-intensive, but also the selection of features is very dependent on the knowledge and experience of researchers [27]. Different from the traditional ML algorithm, the DL algorithm can directly process the ECG original signal, realize the layer-by-layer feature extraction of the input signal, get rid of the dependence and restraint on the prior knowledge and manual feature extraction, and thus realize a new approach that unifies data preprocessing, feature extraction and classification into one framework. In other words, the CNN structure can accomplish all these steps “end-to-end” without requiring class-specific feature extraction [28]. Currently, DL is mostly used to identify myocardial infarction [22, 29] or arrhythmia [30] in the automatic diagnosis of ECG, because the ECG characteristics of these types of heart diseases are more obvious, and the feature identification is relatively easy, and the results are often better. However, what we’ve done is to identify patients with coronary artery disease from the ECG that is basically normal or has no obvious specific changes. To our knowledge, our study showed for the first time to explore the application of a deep learning–based algorithm for detecting CAD using only ECG. Our model yielded moderate values (AUC 0.75, sensitivity 68.7%, specificity 70.9%) in the test dataset, which showed a comparable performance to the traditional Diamond–Forrester model [31] and the widely available CAD consortium model [32]. The Diamond–Forrester model provided an easy way for determining the pre-test CAD probability by using simple demographic data and clinical factors which included sex, age and type of chest pain [31, 33]. However, it has been shown that this model overestimates CAD risk in more contemporary populations because of the obsolescence of model development population [34]. Although other updated models such as CAD consortium model included more clinical factors and have improved performance and versatility, the multi-dimensional variables contributed to the complexity of the model and difficulty in clinical application [32, 35–37].

It is well known that CAD is related to many factors, such as sex, age, weight, body mass index, smoking status, systolic blood pressure, and so forth. Zheng et al. [38] developed a CNN model to detect CAD patients based on facial photos, which achieved moderate performance of AUC 0.730, sensitivity of 80.0% and specificity 54.0% in an independent test dataset. However, they fund it did not further improve the model’s performance when additional clinical variables were added. That is, it is easy to use this model only based on facial photos without additional demographic data or clinical variables. However, as stated by the author, privacy protection and data security were major obstacle to the application of this technology. Similarly, in the present study, our model did not use any demographic data or clinical variables but input ECG raw data only for model training, validation, testing mainly because the following reasons. First, to avoid selection bias, we included patients from all departments of the hospital and nearly 30% of the enrolled patients were from outpatient and physical examination center. For these patients, we could not get access to the clinical variables from the healthcare system of our hospital. Second, too many variables will increase the complexity of the model, which is not conducive to the practical application of the model. In addition, we believe that because the research of DL technology in ECG is still in the early stage, too many additional variables may conceal the actual screening ability of DL model based on ECG. Therefore, we are committed to improving the diagnosis accuracy of CAD based on the ECG raw data alone through DL technology. Fortunately, our model also obtained moderate performance for detecting CAD (AUC 0.75, sensitivity 68.7%, specificity 70.9%).

This study showed the possibility and feasibility of a deep learning-based algorithm for CAD screening only based on ECG. While this area of research is still in relative infancy, there is reason to be optimistic about the applications of such technology in the future, that is, the AI-ECG could be implemented into early or asymptomatic CAD screening as an effective, low-cost, and non-invasive assay. Even though the model training process were highly reliant on significant computing power and time, however, once the model was complete, only a smaller amount of computational power is needed to analyze and interpret the input ECG. In future clinic, for patients with suspected asymptomatic CAD, the computers using our algorithm can make prediction on ECG inputs and the prediction result will be sent back and displayed to the clinicians. Then, clinicians may interpret the result combining with patients’ general conditions and clinical symptoms or signs. According to the comprehensive results, physicians can help patients guide some general recommendations or further diagnostic tests. However, it is clear that we still have a long way to go before the actual clinical applications of this tool, which requires the joint effort of scientific researchers and clinicians around the globe.

Some limitations of our study should be noted. First, since the present study was based at only one hospital, the lack of external validation may result less accuracy compared with the performance in internal validation. So additional validation in patients from other hospitals even other countries and races would be necessary to examine model’s performance. However, it does not affect our main purpose of evaluating the feasibility of detecting CAD using ECG alone. In the near future, we will first plan to launch a multi-center validation of this AI algorithm in our city. Second, it is currently unknown how deep learning algorithm contribute to decision process. The lack of transparency inherent to CNN has been linked to a “black box”, which makes it difficult for clinicians to understand how exactly the AI arrives at a solution. Third, in our study, the relatively low specificity (70.9%) and up to 29.1% false-positive rate may place additional confusion and anxiety on patients. Thus, identifying a higher risk population that might benefit most from the algorithm based on basic clinical data should be explored, which is also our future study direction.

Conclusion

In this study, we developed a model based on deep learning algorithm, which brought up feasibility and possibility of predicting the presence or absence of CAD. The AI-ECG could be as a relatively low cost, noninvasive and efficient screening tool for CAD only using routine 12-lead ECG. Additional prospective studies from multicenter will be urgently needed to accurately access the screening ability of the model and determine whether the AI-ECG can relieve some economic burden on CAD patients as well as the health care system.

Materials and Methods