Machine learning is a subset of artificial intelligence, aimed at developing ‘intelligent’ algorithms that harness data to execute tasks with optimal performance (1). Machine learning algorithms can be broadly split into four categories: Supervised, semi-supervised, unsupervised and reinforcement-learning (2). In supervised learning, the algorithm is given a dataset known as ‘training data’, where each training sample corresponds to one or several inputs and the desired output (3). Through an iterative process, the algorithm determines a function which can correctly predict the desired output from a set of new, previously unknown inputs. Supervised machine-learning tasks include classification, regression and forecasting (4). Unsupervised learning, conversely, is carried out on unlabeled datasets with the aim of extracting patterns and information without external supervision (5). Semi-supervised learning, as indicated by the name, falls somewhere between supervised and unsupervised techniques, as only a portion of the training data is labeled (6). Lastly, during reinforcement learning tasks, an intelligent agent takes actions in a set environment (7). The actions return a reward, while also influencing the environment and the state of the agent. The goal of the agent is to ‘learn’ the policy which maximizes the reward function, or more generally, maximizes the reinforcement signal that is generated by the rewards (8).

Machine learning approaches, as a whole, have become increasingly relevant in the current era of ‘big data’. Data technologies are rapidly evolving, with data storage sizes entering petabytes, cloud services enabling high data transfer speed and computational systems shifting towards high performance cluster computing (9-11). This has allowed the implementation of machine learning in a range of diverse fields, including healthcare. A trove of biomedical data is generated daily throughout the process of medical practice and patient care. Examples of such data include imaging results (ultrasounds, magnetic resonance imaging, computed tomography scans), laboratory test results (cell cultures, biological material analyses and sequencing), patient medical history, drug effects and interactions and patient health outcomes (12). These can be regarded as attractive targets for the implementation of machine learning algorithms, wherein the desired objective is tailored to the respective challenge.

The prediction of disease complications constitutes a key aspect of patient treatment and management (13). The study published in 2022 by Ghosheh et al (14) investigated the use of a predictive system to predict the risk of developing complications in patients diagnosed with coronavirus disease 2019 (COVID-19), trained on data of >3,000 patients in the United Arab Emirates. The prediction of disease complications, overall, can be interpreted as a predictive classification problem, thus opening the door to the application of supervised machine learning algorithms. The current status of the patient can be analyzed by diagnostic classification models, while potential outcomes can be predicted by prognostic classification models (15). Predictive classification algorithms have been implemented in the context of various diseases within the past years. During the peak of the recent severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) pandemic, classification models were designed to tackle various aspects of the disease, such as the detection of viral infection through X-ray imaging and CT scans or the prediction of outcomes of patients with COVID-19 using their recorded characteristics as input (16-18). This is even more important under the scope of personalized medicine, as the individual patient profile, which includes features, such as comorbidities, age, ethnic background etc., may affect disease progression and clinical manifestations (19). In a number of cases, several conditions, or labels, may be assigned to a single patient; for example, an individual may be suffering from COVID-19, while also exhibiting cardiovascular issues and high cholesterol. In such a case, the challenge of building predictive models can be regarded as a multi-label classification problem.

This type of classification task falls under the supervised machine learning ‘umbrella’ and constitutes a modeling problem where the class, also known as target or label, is predicted for a data point, otherwise known as input sample (20). In single-label classification, from a collection of discrete L (L>1) labels, a single label ‘l’ is assigned to each input sample (21). If the number of labels L is 2, the task is considered a binary classification problem, whereas if the number of labels L>2, it is considered a multiclass classification problem (22). In the case where each input sample is associated with a set of non-exclusive class labels, the task is called multi-label classification (23).

One of the common challenges in classification and even more so in the case of multi-label, biomedical datasets, is the presence of imbalanced data (24). In the multi-label setting, imbalance can be traced to three distinct levels, imbalance within labels, imbalance between labels and imbalance within label sets (25). In inter-label imbalance, the label column contains a disproportionate ratio of negative vs. positive samples, effectively obscuring the signal of the particular label (26). In intra-label imbalance, there is difference in frequency between labels, with frequency most often termed as the number of positive instances (27). Labels with an abundance of positive samples are considered in majority, while the rest are minority labels (28). In multi-label classification, there is a possibility for a sample to simultaneously contain a label in minority and another label that is in majority. Labels in majority are termed head labels and labels in minority tail label (29). The last source of multi-label imbalance is the existence of label sets, where some sets may be more frequent in the dataset than others (30-32).

Myocardial infarction (MI), or more generally known as a ‘heart attack’, is caused by a decrease or total interruption in blood flow to part of the myocardium (33). As per the World Health Organization, more than four out of five cardiovascular disease-related deaths are due to heart attacks and strokes, while a third of these deaths concern individuals aged <70 years (34). Thrombosis has been established as the most prominent driver of acute MI, itself stemming from events of atherosclerosis and inflammation (35). Atherosclerotic lesions emerge as thickening of the coronary inner walls of the artery or as fatty streaks, ultimately leading to thrombus formation (36). During their hospitalization, patients with acute MI often face severe complications, such as shock, stroke, atrioventricular block, respiratory failure and cardiopulmonary arrest (37-39). Shock and posterior cerebral artery, in particular, constitute prevalent causes of mortality in patients with acute MI (40). Establishing a robust plan of action is inextricably linked to the successful management of MI-related complications in the hospital setting. Therefore, the application of a classification model to predict potential MI complications may prove useful in informing the decision-making of medical professionals.

In the present study, to demonstrate the potential of leveraging biomedical patient data in the context of predictive classification, a multi-label classification model was trained on a recently released, public dataset of MI patient data.

At the end of the data preprocessing steps, the dataset contained 1,700 rows corresponding to 1,700 patient samples, 100 columns of patient measurements/features serving as the input (X) data and 18 columns of patient outcomes, serving as the target label (y) data. As visible in the histogram plot illustrated in Fig. 2A, there is notable imbalance across the instance of labels. The number of labels per sample is also varied, with most samples assigned one or two labels and very few samples exhibiting more than three labels (Fig. 2B). This can be traced back to the challenge of data collection that was touched upon in the introduction segment; the limited number of patients affects the number of labels (outcomes) that happened to be present among them and were thus recorded.

The output label portion of the dataset was used as input to generate information regarding label interactions and relationships. In the case at hand, the potential outcomes and complications of hospitalized patients with MI are regarded as labels. The results of the exploration of the label space can be summarized in the circular graph of Fig. 3. Each label, i.e., each complication, is represented as a node in the graph, with an edge existing when there is co-occurrence between the labels, weighted by the frequency of co-occurrence. By using the Louvain algorithm, a popular community detection method, three clusters are reported, denoted by purple, yellow and light blue colour.

The atrial fibrillation (FIBR_PREDS) complication (non-lethal) exhibits strong relations with progress of congestive heart failure and asystole, both tagged within the dataset as lethal complications. Notable relations are also reported between atrial fibrillation-relapse of the myocardial infarction, and atrial fibrillation-pulmonary edema (OTEK_LANC). It is also noteworthy that, as regards lethal outcomes, cardiogenic shock, asystole and ventricular fibrillation have been clustered together, as have pulmonary edema, progress of congestive heart failure and thromboembolism. The lethal complication of myocardial rupture, on the other hand, has been clustered with- and exhibits relations with-the myocardial rupture (RAZRIV) complication label, potentially indicating that patients with myocardial rupture were assigned to both labels up to the lethal outcome.

The candidate algorithms were evaluated according to their performance on the dataset and the results are summarized in Fig. 4. The highest HL was exhibited by the Multi-label k-Nearest Neighbor (MLkNN) classifier, whereas the lowest HL (0.053) was exhibited by the OnevsRest classifier with XGBoost. This intuitive classifier strategy, also known as the one-vs.-all, trains a binary classifier independently for each label, and can be applied to both multi-class and multi-label problems.

Predictive classification algorithms have been implemented in the context of various diseases over the past years. During the peak of the recent SARS-CoV-2 pandemic, classification models were designed to tackle various aspects of the disease, such as the detection of viral infection through X-ray imaging and CT scans or the outcome prediction of COVID-19 patients using their recorded characteristics as input (16-18). This is even more critical under the scope of personalized medicine, as the individual patient profile, which includes features, such as comorbidities, age, ethnic background etc., may affect disease progression and clinical manifestations (19). In a number of cases, several conditions, or labels, may be assigned to a single patient; for example, an individual may be suffering from COVID-19, while also exhibiting cardiovascular issues and high cholesterol levels. In such a case, the challenge of building predictive models can be regarded as a multi-label classification problem, as explored herein.

The development and testing of classifiers is a complex task, particularly in the case of multi-label classification. The comparative evaluation of various algorithms on the myocardial infarction dataset elected the OneVsRest heuristic as the best-performing one. Furthermore, the use of XGBoost as the base classifier enabled the fine-tuning of the model and accelerated the learning process. XGBoost was identified as the best performing algorithm in a study published in 2022 for a similar predictive task. That study aimed to develop and validate a machine learning-based model to predict regional lymph node metastasis in osteosarcoma using data from 1,201 patients, identifying T and M stage, surgery and chemotherapy as significant risk factors and XGBoost as the best performing predictive algorithm for that task (64).

The use of disease datasets is also employed by other frameworks; for example, Tang et al (65) described a Gaussian randomizer-based system for early fundus screening with privacy preserving and domain adaptation, employing a multi-disease dataset.

It would be of interest, as an extension of the proposed framework, to evaluate a set of different base classifiers in the context of the OneVsRest strategy and observe the effect that their substitution may have on the classification performance. The present study focused on a subset of the available algorithms and strategies; therefore, there exist other potential machine-learning components and techniques to be evaluated on this task. In terms of the dataset itself, graph exploration highlighted shared label instances that potentially contain information relevant to myocardial infarction pathophysiology. The label imbalance that marks the dataset constitutes an interesting point in terms of handling a multi-label classification problem. Methods to address imbalanced datasets in multi-label classification have been reviewed elsewhere and include, but are not limited to, random oversampling and undersampling, heuristic oversampling, cost-sensitive learning, and ensemble approaches (26,31). Oversampling is the process of increasing the rate of minority class instances within an imbalanced dataset to compensate for the occurrence of common classes (66). Modern and widely-used techniques, such as the Synthetic Minority Over-sampling Technique (SMOTE) create synthetic data points by using the feature space of the minority class and k-nearest neighbors; however, applying the k-nearest neighbor approach on binary input data such as the dataset at hand would serve no purpose (67). Furthermore, the existence of binary input data excludes the use of SMOTENC, the SMOTE extension for numerical and categorical features (67). Therefore, if an added oversampling step were to be applied, a custom oversampling function would need to be created to increase tail label samples based on the calculated oversampling ratio of the labels. Lastly, the concept of errors constitutes an important facet of developing accurate and reliable biomedical classification models. A classifier is subject to two main types of errors, false positives, also known as type I errors, and false negatives, also known as type II errors (59). In the case of false positives, the classifier predicts a label which is not present in the test set, while in the case of false negatives, a label that should have been predicted is missing. Similarly, true positives are results where the classifier has correctly predicted a label presence and in true negatives, the classifier has correctly predicted the absence of label, i.e., the absence of the positive instance. In the case of disease complication predictions, we are greatly invested in limiting the rate of false negatives, where the classifier fails to predict a label (a complication) which in truth exists. One could argue that in the hospital setting, it would be less damaging to monitor a patient in anticipation of a complication that turns out to be a false positive, than failing to catch a complication that may be lethal. Therefore, the selection of performance metrics and the penalties enforced on the errors of the model are greatly affected by the nature of the disease which we wish to interpret as a classification task.

In conclusion, MI constitutes a highly frequent phenomenon in the subset of the population suffering from cardiovascular problems. The development of accurate and scalable systems to support the decision-making process of the medical professionals in the hospital can alleviate the weight of patient management and may potentially increase the odds of survival for myocardial infarction patients. The use of predictive systems for disease-related challenges has been garnering attention the past years with the increase in computational power and novelty of algorithms.

The data-driven approach presented herein and the obtained results underline the potential of machine learning applications in risk predictions, in particular for the challenge associated with MI. As demonstrated through the evaluation, high-performance algorithms, such as the Extreme Gradient Boosting algorithm can be employed as base classifiers in the context of machine-learning model development, while disciplines such as graph theory can shed light into the elaborate networks underlying myocardial infarction progression. Public dataset repositories can provide the large-scale quantities of biomedical and patient data that are required to build efficient and reliable predictive classification models. This data-driven approach can be further scaled and enhanced; there exists promise in the use of ensemble models, made up of different classifiers with different aptitude towards predicting specific labels. Overall, the classifier-based pipeline holds the potential to support the decision-making process of healthcare professionals and aid a proactive approach to patient care.