Addressing myocardial infarction in South-Asian populations: risk factors and machine learning approaches

Myocardial infarction (MI) is a leading cause of death globally, contributing to nearly 17.9 million annual deaths, significantly impacting healthcare systems and economic productivity¹. Certain ethnic populations^2,3, particularly South Asians, experience higher rates of MI, with onset about a decade earlier, and face more severe post-MI complications compared to Western populations^4,5,6,7. This increased risk is largely attributed to genetic predispositions, lifestyle, environmental factors, and higher rates of diabetes, dyslipidemia, and central obesity^6,7,8. The INTERHEART study, which analyzed risk factors for acute MI in 52 countries, identified smoking, psychosocial stress, physical inactivity, and poor diet as key risk factors for acute MI in South Asians⁶. Despite their disproportionately higher rates of MI and post-MI complications, research focusing on the South Asian populations is relatively scarce.

Traditional risk scores, such as the ASCVD risk estimator, Framingham Risk Score, and QRISK, often underestimate cardiovascular risk in South Asians^9,10, while the INTERHEART⁶ Modifiable Risk Score is based on case-control data¹¹. Additionally, a cohort study from Norway found that the NORRISK 2 model underestimated the CVD risk in South Asian immigrants by a twofold difference¹². Furthermore, WHO cardiovascular risk charts have been shown to misclassify high-risk South Asians as low-risk with higher prevalence¹³. These findings highlight the need for more accurate tools tailored to the unique risk profile of South Asian patients.

Given the limitations of existing risk assessment tools for South Asians, clinicians often rely on other methods and parameters. The Mediators of Atherosclerosis in South Asians Living in America (MASALA) study found the coronary artery calcium score to be a very specific marker for subclinical atherosclerosis¹⁴. Additionally, the INTERHEART study found the ApoB/ApoA-1 ratio to be a reliable biomarker for predicting CVD risk¹⁵. Thus, it is crucial to identify South Asians at risk of MI and enhance the detection of MI among symptomatic patients using advanced methods. Machine learning (ML) has shown potential in improving diagnostic and predictive capabilities. However, a significant research gap continues to exist in ML applications for CV diagnostics tailored to South Asians. This review aims to explore the risk factors for MI in the South Asian population and understand the applications of artificial intelligence (AI) and its subsets of ML and deep learning (DL) in such high-risk populations.

Cardiovascular disease (CVD) is an important cause of death among South Asians, particularly in India¹⁶, where 52% of deaths before age 70 are attributable to CVD, compared to 23% in Western populations^17,18. South Asians face a 40% higher risk of mortality from MI compared to other populations^4,19,20. The economic impact is substantial, with India projected to lose $237 billion in productivity due to CVD^4,17.

The higher CVD burden among South Asians suggests that ethnicity by itself may be a significant risk factor^21,22. South Asians tend to develop metabolic abnormalities at a lower BMI and waist circumference than their White counterparts²³, suggesting genetic and ecological influences on cardiovascular outcomes. Cohort studies in rural South India identify tobacco use, alcohol consumption, hypertension, diabetes mellitus, and obesity as significant risk factors^9,20. Low physical activity, reduced muscle mass, and central accumulation of ectopic fat also correlate with elevated risk in South Asians. While primary prevention measures, such as exercise and dietary changes, have been effective in the United States, they are underutilized in India, where evidence-based treatments are alarmingly low^17,24.

Traditional cardiovascular risk factors, such as blood pressure, cholesterol, smoking, and diabetes mellitus, are commonly used to calculate risk scores for premature MI across all ethnicities. The Framingham risk score²⁵, introduced in 1998, estimates the 10-year cardiovascular risk of individuals based on data from the Framingham heart study, a multigenerational study identifying common factors for CVD. However, this score has limitations; it underestimates CVD risk among certain populations, including South Asians²⁶. However, risk factors more prevalent among South Asians, such as insulin resistance and differential body fat distribution, are often underrecognized²⁷. The MASALA study highlighted that the typical South Asian vegetarian diet correlates with higher insulin resistance and lower HDL cholesterol levels, leading to a greater risk of MI¹⁴. The study also showed that the prevalence of CVD risk factors may differ by South Asian subgroup: hypertension was most prevalent in North Indians compared to South Indians and Pakistanis, while dyslipidemia was highest in South Indians and Pakistanis²⁸. Other studies have shown similar instances of cardiovascular risk heterogeneity within the South Asian population: Population-based analysis of five major Asian groups showed that Bangladeshis have a greater prevalence of diabetes as compared to Pakistanis and Indians, while a cohort study indicated an elevated proportional hazard of ASCVD incidence compared to Pakistani and Indian subgroups^4,29.

Developing a CVD risk calculator tailored to South Asians is crucial for addressing these unique risk factors (Fig. 1) and enhancing prevention strategies. However, research is limited by inadequate data, unreliable clinical event reporting, and insufficient information on risk factor exposure patterns, including diet, physical activity, abdominal obesity, alcohol intake, and psychosocial factors³⁰. Thus, a comprehensive framework outlining the multifactorial—genetic, economic, sociological, and dietary—causes of premature CVD and MI in South Asians is needed.

The left panel displays traditional risk factors for ASCVD, while the right panel highlights the risk factors that are more specific and prevalent in the South Asian population.

Full size image

Genetic factors

The heritability of CVD risk factors is well documented among close relatives, but recent studies have focused on the genetic contribution to population-level CVD risk^{31,32,33,34,35,36,37,38}. In South Asian populations, genetic variations manifest in various potential phenotypic traits such as increased visceral obesity, high waist-to-hip ratios, dyslipidemia (high triglyceride and low HDL levels), hyperglycemia, elevated blood pressure, and smaller diameter of coronary vessels^{39,40,41,42,43,44,45}.

Genome-wide association studies (GWAS) facilitated by next-generation sequencing have identified genetic loci and polymorphisms linked to common CVD risk factors^{33,34,37,38,46}. Although the exact mechanisms driving CVD risk in South Asians are not fully understood, population-specific genetic polymorphisms are believed to play a significant role^35,36,37. South Asians demonstrate a higher prevalence of prediabetes and diabetes mellitus, which are significant cardiovascular risk factors. The combination of these genetic traits substantially increases the incidence of CVD in this population¹⁶. Comprehensive genetic research is crucial to developing targeted prevention and treatment strategies for CVD in South Asians.

Research indicates that common disease variants account for only a modest portion of total disease heritability. There is substantial interest in the impact of rare disease variants, which are more likely to differ between ethnic groups^{47,48,49,50,51,52,53,54}. Further genetic analysis of the South Asian population is essential to elucidate the extent to which these polymorphisms contribute to CVD risk⁵⁵. Additionally, the South Asian population is genetically diverse and not homogenous, necessitating detailed subgroup analyses¹⁹.

Socioeconomic factors

Over the past fifty years, South Asia has transitioned from an agrarian-based economy to one dominated by manufacturing and service sectors, leading to significant reductions in physical activity and changes in dietary habits^56,57. This economic shift has been linked to an increase in MI among the South Asian population^57,58. Despite economic advancements, South Asia remains a significant center of global poverty, with 389 million people classified as multidimensionally poor (MPI poor) as of 2023, accounting for over 35% of the global MPI poor population^56,59. Low socioeconomic status (SES) in the region is strongly associated with increased CVD risk factors⁶⁰, including poor diet quality, substance abuse, reduced health screening, and mental health disorders^{61,62,63,64,65,66}. Low SES during childhood is also linked to higher blood pressure in adulthood⁶². The prohibitive cost of CVD care exacerbates poverty, leading to catastrophic healthcare spending and lower employment rates in affected households. This is worsened by the fact that members of CVD-affected households also experience lower rates of employment as well^67,68,69,70.

Consequently, more studies exploring the impact of SES on CVD risk in South Asian populations are necessary to understand deviation from global trends and how economics impact population-level CVD risk in South Asians. Interestingly, some studies in South Asia have revealed that high SES, rather than low SES, is associated with a higher risk of CVD and related risk factors, including obesity, physical inactivity, diabetes mellitus, hypertension, and dyslipidemia^61,63.

Incorporating economic factors into cardiovascular risk scores is crucial for providing a comprehensive understanding of an individual’s risk profile, capturing the interplay between financial stability and health outcomes. This will enable more effective prevention and management strategies tailored to the unique needs of the South Asian population.

Dietary factors

The intricate relationship between diet and MI among South Asians necessitates an in-depth examination, considering the regional diversity of dietary patterns. South Asian diets traditionally include a variety of vegetables, pulses, cereals, and potatoes rich in whole grains, fiber, and complex carbohydrates, with notable regional distinctions. However, urbanization and advancements in food processing have led to adverse lifestyle and diet changes, such as increased processed food intake and deep-frying cooking⁷¹. This refined carbohydrates-based diet, rich in sugar and saturated and trans fatty acids, and low in proteins, monounsaturated fatty acids (MUFAs), and ω-3 polyunsaturated fatty acids (PUFAs), is suggested to contribute to elevated CAD risk, obesity, and insulin resistance in South Asians^71,72. As grains are a staple to the South Asian diet, comprising 60-70% of total energy intake, micronutrient-poor refined grain intake contributes to the increased cardiometabolic risk associated with elevated triglycerides, lower HDL cholesterol, and prevalence of metabolic syndrome in South Asians⁷³. While Kanaya et al.⁷⁴ and Molina et al. ⁷⁵ did not demonstrate a direct association of HDL-to-triglyceride ratio with coronary artery calcification, the low HDL-to-triglyceride ratio was noted to be an important surrogate marker of insulin resistance in the South Asian population by Flowers et al.⁷⁶.

Regional dietary habits play a central role in understanding MI and cardiovascular risk among South Asian populations. The Pakistan Risk of Myocardial Infarction Study (PROMIS) highlighted genetic determinants, lifestyle, and dietary factors as critical determinants of CAD in South Asians, linking high-fat diets with increased MI risk⁷⁷. In Bangladesh, younger adults with acute MI showed significant associations with unhealthy dietary habits, including high rice and beef consumption and lower intakes of fruits and vegetables⁷⁸. A study in Sri Lanka also indicated that high dietary intake of meat, along with alcohol consumption and smoking, were significant risk factors for MI among young males⁷⁹. Further, in a cross-sectional analysis conducted by Daniel et al. of dietary habits across Delhi, Mumbai, and Trivandrum, regional variations in diet composition was found to have significant associations with chronic disease outcomes⁸⁰.

Specifically, South Indian diets rich in pulses and vegetables were inversely associated with hypertension and diabetes mellitus prevalence. In contrast, diets rich in high glycemic index foods such as high-fat dairy, and fried foods, were linked to increased abdominal adiposity and a high cardiometabolic risk profile. Other studies have affirmed that high-fat intake—including high sugar intake and use of ghee rich in saturated fat—and low consumption of fruits and vegetables are risk factors for premature cardiovascular events such as MI in young South Asians^5,81. Thus, dietary patterns among South Asian populations are complex and vary significantly by region. Low-fat, pulse-based diets in southern regions provide protective benefits against diabetes mellitus and hypertension, while high-fat diets in northern regions contribute to a higher cardiovascular risk profile. These findings emphasize the need to integrate dietary habits into the risk score to address the unique risk factors specific to different South Asian communities.

In summary, we propose developing a novel risk score specifically designed to predict the risk of MI among South Asians⁸². By incorporating major risk factors identified in prior research, including genetic, economic, sociological, and dietary factors, a risk score metric that accurately quantifies the odds of a South Asian individual experiencing an MI, can be obtained. Integrating AI will be beneficial for understanding and capturing the increased risk, as it can enhance predictive accuracy and identify complex patterns in the data. This approach requires thorough validation and depends on data quality.

AI is an interdisciplinary field encompassing computer science, statistics, psychology, neuroscience, philosophy, control theory, and mathematics. It aims to perform tasks that typically require human intellect, such as problem-solving, building intelligent agents, navigating complex environments, knowledge representation, inference, planning, visual perception, decision-making, and language translation⁸³.

ML, a subset of AI, involves developing algorithms and models to identify implicit patterns or relationships in data with minimal human intervention. ML consists of a model, a learning algorithm, and a task⁸⁴. In contrast to traditional programming, ML algorithms learn from datasets and are able to explore both linear and non-linear relationships between variables. ML is categorized into supervised and unsupervised learning. Supervised learning algorithms use labeled datasets to map inputs to outputs, while unsupervised algorithms uncover the underlying structure of datasets without labeled data and is used to explore hidden relationships and patterns among data.

DL, a subset of ML, employs artificial neural networks to represent multiple levels of abstraction through computational models⁸⁵. It can be used in both supervised and unsupervised fashions and consists of feature extraction and classification. Artificial Neural Networks (ANNs) comprise an input layer, one or more hidden layers, and an output layer, with initial training weights and tuning hyperparameters.

ANNs utilize activation functions, allowing DL to model non-linearity. DL algorithms are trained using backpropagation through gradient descent, where backpropagation computes the gradient of the loss function concerning weights and biases, and gradient descent iteratively updates these parameters to improve feature extraction and prediction accuracy.

DL has been employed for cardiovascular image interpretation and the development of models combining clinical and imaging parameters to predict cardiovascular outcomes⁸⁶. In our review, we explore the ML algorithms, their applications in diagnostic methods for CVDs, and their possible use for more accurate and efficient MI risk prediction and diagnosis in the South Asian population.

Leveraging ML for improved MI diagnosis in South Asian populations

ML holds significant promise for enhancing MI diagnosis in South Asians, who experience higher incidence and earlier onset compared to other ethnic groups⁵. Traditional diagnostic tools often fail to account for the unique genetic, environmental, and lifestyle factors contributing to this elevated risk. ML algorithms, particularly convolutional neural networks (CNNs)⁸⁷, can analyze complex data such as electrocardiograms (ECG) and cardiac biomarkers with high precision, identifying subtle patterns that may be missed by clinicians⁸⁸. Table 1 provides a comprehensive overview of various ML models and techniques, along with their advantages and limitations.

MI diagnosis traditionally relies on clinical history, ECG changes, and biomarkers like troponins⁸⁹. However, these methods can yield inconclusive results, particularly in atypical cases. Gupta et al. noted that South Asian patients with MI may take longer to seek medical attention after the onset of symptoms, which can complicate timely diagnosis and management⁹⁰. ML models can enhance diagnostic accuracy by integrating multimodal data from ECG, and biomarkers, providing a comprehensive assessment in a shorter time frame⁹¹. ML algorithms have shown promise in diagnosing myocardial infarction (MI) and predicting short-term outcomes, demonstrating significant improvements over traditional methods, as detailed in Tables 2 and 3. These advancements highlight the potential for tailored approaches to enhance cardiovascular care and management in diverse patient populations. This integrative approach could be crucial for the South Asian population, where ML can improve the identification of atypical presentations⁹⁰.

Personalized diagnostic approaches enabled by ML consider individual risk factors unique to South Asians, including higher rates of diabetes and central obesity, tailoring assessments to the specific risk profile of each patient⁹⁰. This personalization enhances the relevance and accuracy of diagnoses. Additionally, ML reduces diagnostic errors by providing consistent and objective analyses of diagnostic data, minimizing the risk of human error⁹². The complexity of integrating and interpreting multiple diagnostic tests often leads to errors; however, ML algorithms can automatically interpret ECG signals and imaging results, ensuring more reliable diagnoses.

Moreover, ML optimizes healthcare resources by prioritizing high-value diagnostic tests and reducing unnecessary procedures, which is particularly beneficial in resource-limited settings in South Asian countries⁹³. This efficiency ensures patients receive appropriate and timely care. Incorporating ML into MI diagnosis also addresses health disparities in access and quality of cardiovascular care. ML-driven mobile health applications can facilitate early detection and continuous monitoring of cardiovascular health in remote or underserved areas⁹³. It must be noted that Machine Learning for risk prediction has limitations when applied in clinical settings. Traditional ML models may not be able to overcome confounding factors or biases that are inherent to the training data. Careful consideration of the quality of training data is necessary, and additional strategies by the clinician must be made to mitigate these biases to ensure accurate outcomes in clinical decision-making.

In summary, ML could significantly enhance MI diagnosis in the South Asian population by improving diagnostic accuracy, integrating multimodal data, and identifying atypical presentations⁹⁴. ML may provide personalized diagnostic approaches, reduce errors, and optimize healthcare resources. These algorithms can generate differential diagnoses, suggest high-value tests, and minimize repeated testing. Beyond diagnostics, ML may be able to predict patient outcomes and assist in clinical decision-making by providing insights based on large datasets, helping clinicians tailor treatment plans effectively⁹⁵. These advancements are crucial for addressing the unique challenges faced by South Asians in cardiovascular health, ensuring better clinical outcomes for this high-risk population.

Electrocardiogram

Electrocardiograms (ECG) are non-invasive, low-cost tools essential for detecting abnormal cardiac activity through analysis of electrical impulses, providing important information for MI diagnosis⁸⁹. However, traditional ECG interpretation is limited by the complexity and variability of cardiac signals, which can obscure subtle but clinically significant patterns⁸⁸. ML algorithms can enhance ECG analysis by segmenting and classifying signals using supervised and unsupervised learning, thereby improving diagnostic accuracy and clinician productivity^88,96.

Supervised learning algorithms, such as logistic regression, support vector machines, ANNs, and random forests, are widely used for ECG classification, aiding disease diagnosis and risk stratification^92,94,97. Unsupervised learning methods, including principal component analysis (PCA), help discover hidden patterns in ECG datasets by reducing dimensionality and managing large volumes of signals. This approach can identify phenotypic subtypes in conditions like hypertrophic cardiomyopathy and cluster biomarkers associated with MI^97,98. Nonetheless, overfitting remains a challenge in ML, where models perform well on training sets but poorly on unseen data.

DL architectures, such as autoencoders, deep belief networks (DBNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs), mitigate overfitting and improve ECG classification through hierarchical feature extraction^98,99. RNNs perform recursion by directing the sequence of nodes connected in a chain, where the output of each node is used as input for the next. RNNs are suitable for time-series data like ECG signals but face limitations in learning long-term dependencies for long sequence inputs. Long Short-Term Memory (LSTM)¹⁰⁰ networks address this issue and are popular algorithms for ECG diagnosis. By combining forward and backward LSTMs, bidirectional LSTM¹⁰¹ networks analyze ECG signals using information from both the future and the past, enhancing the accuracy and robustness of ECG signal analysis.

Advanced DL techniques for ECG classification

Advances DL approaches have demonstrated exceptional performance in ECG classification, addressing challenges such as overfitting and the need for hand-crafted features. Stacked autoencoders, comprising multiple layers of autoencoders and a classifier, learn robust representations through complex relationships¹⁰². For ECG classification, denoising and sparse autoencoders can enhance feature detection¹⁰².

In addition to stacked auto-encoders, DBNs, composed of stacked restricted Boltzmann machines (RBMs), are generative neural networks utilizing unsupervised learning modules to predict probability distributions over inputs. During pre-training and fine-tuning of the ECG dataset for classifying MI, each RBM is trained sequentially using contrastive divergence to learn initial weights, followed by backpropagation to optimize discriminative performance with labeled data. DBNs perform well even with corrupted data, although achieving high classification accuracy can be challenging⁹⁸.

CNNs, a form of feedforward neural network characterized by a hierarchical structure, are predominantly used for image-based datasets in DL. Unlike fully connected neural networks, where every layer is connected, CNNs learn filters that apply operations to each sub-region of the input. Architecturally, CNNs consist of convolutional filters, pooling modules, and fully connected layers. Convolutional layers perform the convolution of each sub-region of the input with a filter kernel, extracting features from the input provided by previous layers¹⁰³. Pooling layers reduce the dimensions of feature maps, retaining essential information while lowering computational complexity. Fully connected layers integrate features extracted by convolutional and pooling layers for final classification.

Several advanced CNN architectures, like AlexNet and VGG-Net, improve performance by using deeper networks and smaller convolutional filters^103,104. A deep CNN developed by Acharya et al.⁹⁴ for detecting MI achieved an accuracy of 93.53% using ECG⁹⁴. Xiong et al.¹⁰⁵ reported the highest accuracy for MI localization using ResNet¹⁰⁶, achieving 99.99% accuracy, followed by 99.95% using a convolutional neural network.

Despite these accomplishments, CNNs face several limitations in MI detection. Access to high-quality and large datasets is a major challenge. The performance of DL algorithms often lacks standardized measurement and comparison across different experimental settings, many of which are arbitrary. Additionally, robustness is an issue, as ECG waveforms in clinical settings can be deformed by external forces, potentially affecting DL performance in real-life environments. In summary, advanced ML and DL techniques significantly enhance ECG classification and cardiac disease diagnosis, though challenges such as data quality, model robustness, and interpretability remain.

Cardiac biomarkers

Cardiac biomarkers are proteins released into the blood following cardiac injury, crucial for diagnosing acute MI¹⁰⁷. Aydin et al. identified myoglobin, fatty acid-binding protein (FABP), and glycogen phosphorylase BB (GPBB) as early biomarkers for MI, while troponin T (TnT) and troponin I (TnI) are definitive biomarkers for MI diagnosis¹⁰⁷. Troponin I (TnI) is currently regarded as the gold standard for diagnosing acute MI, with elevated troponin levels at 0, 3, and 6 hours, offering critical insights into the severity and timing of myocardial injury.

The myocardial-ischemic-injury index (MI3), an ML algorithm developed by Than et al.¹⁰⁸, incorporates age, sex, and cardiac high-sensitivity troponins measured in paired samples. This model was trained on a cohort of 3,013 patients and validated on 7,998 individuals presenting with suspected MI. MI occurred in 10.6% of patients in the test set and 13.4% in the training set. The MI³ index effectively identified low and high-risk patients using a gradient boosting algorithm, allowing for individualized and objective assessments of the likelihood of MI, thus enhancing patient care. The MI³ algorithm computes a value between 1 and 100, reflecting an individual’s likelihood of an MI diagnosis¹⁰⁹. Despite its potential, gradient boosting is prone to overfitting and is computationally expensive. Current diagnosis of MI relies on fixed troponin thresholds, which does not account for varying troponin levels based on age, sex, or time.

Doudesis et al.¹⁰⁹ validated the MI3 algorithm through an exploratory analysis using a multicenter randomized trial conducted in Scotland. Among 20,761 patients, 3,272 (15.8%) had an MI. The MI3 algorithm demonstrated a clear area under the receiver-operating characteristic curve of 95%, identifying 12,983 (62.5%) patients as having a low probability of MI at a prespecified threshold, with a negative predictive value of 99.8%. Additionally, 2,961 patients were identified as having a high probability, with a positive predictive value of 70.4%. At one year, subsequent MI or cardiovascular deaths occurred more frequently in high-probability patients than in low-probability patients (17.6% vs. 0.5%; p < 0.001). These results suggest that the MI3 algorithm is a potential tool for improving the accuracy of MI diagnosis by accounting for varying troponin levels, age, gender, and time, thus allowing for more personalized and accurate risk assessments. Multilayer Perceptrons (MLPs) are another ML algorithm that can be used to classify cardiac biomarkers. MLPs comprise multiple layers of neurons, each fully connected to the next, allowing them to capture complex patterns in the data. These advanced algorithms can enable a more nuanced analysis of cardiac biomarkers.

In summary, leveraging cardiac biomarkers such as troponins with innovative ML algorithms like MI³ offers a more personalized and accurate approach to diagnosing and managing MI. These advancements hold promise for significantly improving clinical outcomes in patients with suspected acute MI. ML algorithms such as VGG-Net and CNNs utilize image datasets, achieving high accuracy in diagnosing MI¹¹⁰. However, there are limitations, such as a lack of access to quality datasets and challenges in evaluating dataset quality. Significantly less research has focused on assessing the information quality in datasets¹¹¹, and high-quality datasets are essential for developing and training ML algorithms, particularly in MI detection. To address these challenges, a user-centric, ML-based information quality evaluation tool for assessing the quality of MI datasets across three dimensions—consistency, relevance, and accuracy—could prove useful. Consistency measured using edit distance and cosine similarity, relevance assessed by Jaccard similarity and distance, and accuracy evaluated using KL divergence, the Kolmogorov-Smirnov test, and Pearson correlation can enable us to quantify the dataset’s information quality.

Transformer-based models in MI

Based on the Transformer model, the advanced ML algorithm categorizes MI risk into low, medium, and high probability levels. Transformer models have achieved advanced performance in various tasks, including machine translation, language generation, image classification, and object detection^111,112,113. By providing a probabilistic classification, the Transformer model aids in patient stratification and tailored interventions.

Transformer-based models, originally introduced for neural machine translation, have revolutionized natural language processing (NLP) by significantly improving sequence translation accuracy through the attention mechanism¹¹⁴. This addresses fixed-length vector problems by allowing a joint soft search of the source sentence, thus improving translation accuracy. Encoder-decoder architectures have been the state of the art for translation tasks. However, while translating sequences, the decoder could not gain the context of previous vectors in a sequence of sentences. Due to this, translation accuracy was low.

Attention mechanisms overcame this limitation, enhancing translation performance by allowing a joint soft search of the source sentence without facing bottlenecks of a fixed vector, thereby improving translation accuracy. Transformers, introduced by Vaswani et al.¹¹², are solely based on the attention mechanism, dispensing with recurrence and convolution entirely, achieving state-of-the-art results in machine translation¹¹². Subsequently, Transformers have been applied to computer vision^113,115 and speech recognition, achieving advanced performance across numerous ML tasks^111,116,117.

Based on transformers, subsequent generations of language models were developed. Devlin et al.¹¹⁸ introduced Bidirectional Encoder Representations from Transformers (BERT), enabling models to gain a contextual understanding of sentence sequences in machine translation tasks and improving downstream tasks such as text classification, question answering, named entity recognition, and text summarization.

Additionally, Generative Pre-trained Transformers (GPT) by Radford et al.¹¹⁷ utilized only the decoder from Transformers to address natural language generation tasks. GPT-2 increased the capacity and size of parameters, further enhancing natural language generation¹¹⁹. GPT-3 achieved excellent results in many natural language processing tasks, while GPT-4 further increased parameter size and introduced multi-modality, handling both image and text inputs. Recent work by Bubeck et al.¹²⁰ suggests the emergence of Artificial General Intelligence using a GPT-based ML model.

Despite the significant technical progress of language models in ML, their utilization in the field of cardiology is still in its early stages, with considerable gaps in research and application of language models. However, language models hold great potential for diagnosing and managing MI. An innovative vision transformer, HeartBEiT, developed by Vaid et al.¹²¹, demonstrates superior diagnostic performance for ECG analysis compared to traditional CNN architectures, particularly in low-data regimes. This model, pre-trained on a vast body of ECGs, enhances explainability and improves diagnosis by highlighting important regions of the ECG.

Selivanov et al.¹²² demonstrated the potential of language models in medical image captioning by combining radiological images with organized patient data from textual records, generating summaries that aid in diagnosis and treatment planning. Moreover, Bubeck et al.¹²⁰ demonstrate the capabilities of GPT-4 in mastering medical natural language processing tasks, such as clinical question answering, note generation, and treatment plan generation. Training medical datasets using GPT-4 can significantly improve the prediction of MI. Beyond prediction, GPT-4 can identify at-risk patients and generate tailored treatment plans for those likely to suffer from acute MI, thereby enhancing patient care and outcomes.

In summary, leveraging advanced Transformer-based models alongside innovative ML algorithms like HeartBEiT and GPT-4 offers a more personalized and accurate approach to diagnosing and managing MI. These advancements hold promise for significantly improving clinical outcomes in patients suspected of acute MI.

The integration of ML and DL algorithms in the risk prediction, diagnosis, and short- and long-term prognosis of MI can have important clinical implications and offer significant improvements over traditional methods. ML algorithms, particularly CNNs and transformer models, have demonstrated superior performance in analyzing complex data from ECGs and cardiac biomarkers and can detect subtle patterns often missed by clinicians, improving diagnostic accuracy⁹⁴. This helps prioritize resources, particularly in resource-limited settings in South Asia, ensuring timely and appropriate care or referral. ML algorithms specifically can help in triaging patients in the emergency department¹⁰⁸. ML can offer clinical decision support by simultaneously and rapidly analyzing vast amounts of data, offering insights into treatment plans¹²². Thus, the integration of ML and DL holds promise for improving risk prediction, enhancing accuracy, and enabling efficiency in MI diagnosis and management, specifically for the South Asian population. Generative artificial intelligence (AI) powered tools, such as virtual co-pilots, can support clinicians by offering real-time, personalized health recommendations that account for individual genetic, lifestyle, and environmental factors. Furthermore, AI platforms can deliver culturally tailored health education, improving patient understanding and engagement. Integrating these advanced AI technologies into clinical practice has the potential to significantly improve secondary cardiovascular prevention strategies, particularly in high-risk populations.

Despite significant advances in ML and DL, several knowledge gaps remain. One of the main challenges is the lack of high-quality and standardized datasets. Many ML models are trained on datasets with poor consistency, thereby impacting their generalizability and robustness in clinical settings¹⁰⁵ The current ML models in myocardial infarction do not capture the unique genetic, environmental, and lifestyle factors of South Asian populations, leading to underestimation of MI risk in these groups, thereby noting the need for risk assessments tailored to different ethnic groups¹² More comprehensive ML models integrating multimodal data are required. Establishing global standards to improve the quality of datasets will enable proper training of machine learning models. The most important step is translating ML algorithms into actual clinical practice. Collaborative efforts are needed between researchers, clinicians, and policymakers to ensure the accessibility of ML tools to diverse populations.

In conclusion, MI significantly impacts global health, particularly in South Asian populations with higher incidence and earlier onset. Traditional risk scores often fail to capture the unique genetic, environmental, and lifestyle factors contributing to this elevated risk. Advanced ML and DL techniques, such as CNNs and transformer-based models, show substantial promise in enhancing MI detection, prediction, and management by integrating multimodal data and identifying subtle patterns missed by clinicians. Developing tailored MI risk scores for South Asians, incorporating genetic, economic, sociological, and dietary factors, is essential for effective prevention and early interventions. Ensuring high-quality datasets is crucial for these ML models to be clinically applicable. Leveraging these advanced tools can significantly improve cardiovascular outcomes and quality of life for at-risk populations.