1 INTRODUCTION

Stroke, a leading neurological disorder worldwide, is responsible for over 12.2 million new cases each year. It primarily occurs when the brain's blood supply is disrupted by blood clots, blocking blood flow, or when blood vessels rupture, causing bleeding and damage to brain tissue.^1-3 Deprivation of cells from oxygen and other nutrients during a stroke leads to the death of cells, ending in permanent disability and even death.⁴

Ischemic and hemorrhagic stroke are two types of strokes. In ischemic stroke, the blood clot blocks the cerebral vessels, and in hemorrhagic stroke, bleeding occurs inside the brain.^{5, 6} High blood pressure, obesity, physical inactivity, and smoking are among the most critical risk factors in stroke sufferers.⁷

Experts predict that both the death rate from strokes and the number of people affected by this condition will rise alongside global population growth. However, these fatalities can be significantly reduced through early detection and treatment.⁸

In the past decades, with the emergence of evidence-based approaches to stroke prevention, acute stroke management, and stroke recovery, there has been a significant shift in the field of stroke, and the mission of the World Stroke Organization (WSO) is to reduce the global burden of stroke through prevention, treatment, and long-term care.⁹

Today, technology's role in healthcare has grown significantly, capturing the interest of medical professionals. Machine Learning (ML) algorithms, in particular, are being leveraged to produce precise data for diagnosing, prognosis, and predicting various diseases.^10-12 Furthermore, developing predictive tools powered by Artificial Intelligence (AI) can potentially prevent and decrease stroke occurrences.¹³ As a branch of AI, ML offers an array of models capable of detecting intricate patterns, understanding the connections among them, and utilizing this knowledge for prediction or decision-making purposes.^14-16

AI techniques differ from traditional prediction models because they utilize and combine a vast array of variables to describe the ambiguous and complex nature of human physiology.⁵ In addition to helping with prevention, diagnosis, and patient monitoring, they can play a crucial role in managing a vast amount of data accurately and practically.¹⁷

AI algorithms can accurately classify, diagnose, and segment the lesions in the brain tissue. With the help of AI, doctors can diagnose intracranial bleeding, microbial bleeding, and acute ischemic stroke more efficiently.¹⁸ AI uses unique ML algorithms to “learn” features from large data sets and recognize patterns that are often invisible to the human eye.¹³

ML technology employs a range of techniques for automated data analysis, including linear and logistic regression models, Support Vector Machines (SVM), Random Forests (RF), classification trees, and discriminant analysis.^{19, 20} ML is a multivariate approach that can be used to analyze complex and heterogeneous data types and include them in risk prediction models, making it a promising solution for stroke risk prediction.¹³

The purpose of this study is to systematically review published papers on stroke prediction using machine learning algorithms and introduce the most efficient machine learning algorithms and compare their performance. The papers have published in period from 2019 to August 2023.

2 METHODS

The authors conducted a thorough search following specific inclusion criteria they developed. All articles that used ML to predict the occurrence of stroke were reviewed. A comprehensive search was conducted from 2019 to August 2023 using selected keywords in PubMed, Scopus, Web of Science, and IEEE databases. The search strategy is illustrated in Table 1.

2.1 Eligibility of articles

Specific criteria were defined for the inclusion and exclusion of articles. Inclusion criteria were articles that used ML algorithms to predict stroke, articles written in English, available full-text articles, and articles published between 2019 and August 2023. Articles related to other diseases, articles published in languages other than English, articles that used deep learning algorithms, review articles, meta-analyses, books, letters to the editor, or conference papers were excluded from the systematic review. The selection process of articles is depicted in the PRISMA flow diagram Figure 1.

In this systematic review, the data extraction process included a thorough review of previous articles to gather information about their methods and results. Articles were extracted using standardized table formats. Extracted information was the first author of the article, the countries included in the study, the year of publication of the article, the characteristics and dimensions of the studied data set, the types of ML algorithms used and the best algorithm, the evaluation criteria of ML algorithms and selected features in stroke prediction.

3 RESULTS

In the systematic search to identify articles related to stroke prediction with ML algorithms, 5422 articles were identified in the first step. Then, 5360 articles were screened based on the title and abstract, and 42 other articles were removed based on the full text. Ultimately, 20 full-text articles were included in the final analysis (Table 2).

3.1 Study features

All articles included in this study date from 2019 onward (Figure 2). These studies have been conducted in different geographical areas, including four studies in China,^{24, 26, 36, 37} four in India,^{21, 29, 30, 34} four in Saudi Arabia,^{25, 27, 31, 32} three in Japan,^{21, 35, 39} three in Bangladesh,^{8, 21, 32} two in Malaysia,^{29, 38} two in Egypt,^{25, 28} two in the United States,^{25, 30} one in Greece,²² one in Ireland,³³ one in Germany,²³ one in Chile,³⁹ one in England,²⁴ one in Canada ²³ and one in Singapore ³³ (Figure 3). Among these studies, one article was conducted in the form of a prospective cohort study on half a million Chinese adults,²⁴ and the other article was conducted in the form of a 2-year longitudinal cohort study in Southeast China.³⁶ Another study developed a web page and mobile application to improve the display of results.⁸

3.4 Machine learning modeling

In various research in the field of stroke prediction, several algorithms were used to create models. In five articles, the RF algorithm was introduced as the best and most efficient algorithm for stroke prediction^{21, 25, 29, 30, 32} Two studies^{8, 23} recommended the SVM algorithm, and two other studies^{22, 28} chose the Stacking approach as the best approach for building ML algorithms. Two studies determined the XGBoost algorithm as the most efficient algorithm,^{26, 37} and in one study, the combination of XGBoost and RF algorithms was used as one of the algorithms of the new proposed model with high-performance.³¹

However, a study that compared the performance of three algorithms (LR, RF, XGBoost) showed that all three algorithms obtained almost identical results with an accuracy of 65%.³⁵ Another study that compared the performance of NB, SVM, RF, KNN, DT, Stacking, and Majority Voting algorithms showed that the performance of these algorithms was similar, except that the NB algorithm showed the lowest performance.²⁷ The present study indicated that various algorithms such as RF, ANN, NB, SVM, DSGD, GBT&COX, Stacking, and XGBoost are the most efficient algorithms in this field, with a broad review of various studies in the stroke prediction area. (Figure 4).

4 DISCUSSION

A stroke is a critical medical emergency that can be life-threatening or cause irreversible damage. Hence, accurately diagnosing and preventing strokes is crucial. Currently, the application of ML algorithms in healthcare is rapidly increasing.⁴⁰ These algorithms support physicians by leveraging their powerful processing capabilities for clinical decision-making, prognosis, and, notably, forecasting the likelihood of a stroke. Unlike conventional prediction models that rely on calculations, ML models utilize various variables to accurately represent human physiology's complexities.⁴¹

In today's world, machine learning is emerging as a powerful tool in modeling complex and hidden relationships between clinical variables and physiology⁴² and traditional methods are not able to detect and predict stroke in the early stages. Consequently, a systematic review was conducted, focusing on articles that met specific research criteria. This review examined 20 articles to identify the most efficient ML algorithms for predicting strokes.

The review of all studies showed that ML has been an effective and positive method in predicting stroke, and most of the studies conducted in recent years indicate the use of ML in stroke prediction as an emerging tool in the healthcare area. In addition, among the articles reviewed in this research, most of them were conducted in Asian countries such as India, China, Saudi Arabia, Japan, Malaysia, Singapore, and Bangladesh, seeming to be due to the high rate of stroke in these countries and the importance of its prediction.

The articles included in the research had different data sets and modeling records, which showed the dynamics of ML models in predicting stroke. Moreover, the Kaggle data set was used more frequently in the studies. Matthew Chun et al. used the largest data set among the articles included in the study, which had 512,726 records.²⁴ On the other hand, the smallest data set was in the study of Vodencarevic et al., who used the registry data set with 384 records.²³

The research reviewed utilized a variety of algorithms to develop and present the model, employing different criteria to evaluate the ML algorithms' efficiency. In most studies (n = 17), accuracy served as a primary metric for gauging the effectiveness of the ML algorithms. Essentially, an algorithm's higher accuracy indicates superior performance and efficiency in predicting strokes, according to these studies.

In 10 studies, the accuracy of the stroke prediction algorithm was above 90%.^{8, 21, 22, 25, 27-32} Among these 10 studies, five recommended the RF algorithm as the most efficient algorithm in stroke prediction.^{21, 25, 29, 30, 32} Although the RF algorithm has a high accuracy of 90 in all studies, the highest accuracy recorded was in the study of Biswas et al.⁸ in 2022 in Bangladesh, in which the SVM algorithm is the most efficient and best stroke prediction algorithm with an accuracy of 99%.

The variations in selecting the most efficient algorithms and their accuracy appear to stem from differences in sample size, data set, and data type. Consequently, further research must be conducted using consistent data sets, sample sizes, and data types to obtain more reliable outcomes and identify the most efficient ML algorithm model.

Choosing a suitable algorithm and data set can affect the performance of ML models in predicting stroke. Most articles used conventional ML algorithms to build prediction models, and a small number combined models to achieve more stable and robust models.

In some studies, using multiple models, combined methods, and the combination of these models in ML algorithms improved the accuracy of the final model compared to other models and approaches.⁴³ For example, in the study of Alruily et al., the highest AUC value was achieved using the combination of XGBOOST, GBM, and RF models to improve the accuracy of their final “RXLM” model.³¹

Chun M et al.,²⁴ in a study in 2021, found that when they combine Cox and GBT models, they have higher accuracy, specificity, and PPV for predicting stroke than when they use these models separately.

Among the 20 articles studied in our research, the highest degree of sensitivity is related to the study by Alanazi et al.²⁵ and Xiao Zhang et al., in these studies, RF and XGB algorithms are known as the best models, respectively.

The swift advancement of artificial intelligence allows healthcare providers and decision-makers to leverage ML models to pinpoint and understand risk factors associated with strokes. This aids in early prediction and minimizes the severe complications of strokes.²² In a 2021 study by Alanazi et al,²⁵ using Random Forest (RF), Decision Tree (DT), Naive Bayes (NB), and Bayesian Network (BN) algorithms, Alanazi et al.²⁵ demonstrated that nine laboratory tests, alongside age and gender, significantly correlate with stroke likelihood.

Ivanov et al.,⁴⁴ emphasizing that data quality and pre-processing play an important role in the development of reliable models, presented a detailed stroke data optimization model to improve stroke prediction, and in this the research of SVM algorithm with 98 percent accuracy and 97% recall has achieved a high score.

In the study of Dritsas et al.,²² the most critical and relevant risk factor for stroke is age. This is consistent with the findings of this systematic review, which confirms that the incidence of stroke after 45 years is twice as high, and 70% of all strokes occur after the age of 75.⁴⁵

The findings of Sharma et al.⁴⁶ study using a comparison of five algorithms (RF, JRip, NB, MLP, and DT) for early prediction of stroke showed that although lifestyle changes cannot prevent the inevitable occurrence of stroke, they can Significantly reduce the risk of stroke.

One of the challenges of machine learning in managing unbalanced classes is neglecting minority classes. The most popular technique used in the studies of this systematic review to manage unbalanced classes is the SMOTE technique, so that 35% of studies have suggested the use of the SMOTE technique. This technique helps in creating a more balanced data set without duplicating the minority class samples, leading to better generalization.

The under-sampling technique has been used in studies^{30, 33, 34} to manage unbalanced classes, this technique Randomly removes instances from the majority class until the classes are balanced and this causes Simple to implement and reduces the size of the data set, making it faster to train models.

In the study of Biswas et al.,⁸ to manage unbalanced classes, the use of the Random Over-sampling technique has been proposed and used, and the advantage of this technique is simple implementation and increasing the representation of the minority class.

In the study of Vodencarevic et al.,²³ four techniques (COST, SMOTE, under-sampling, and Anomaly Detection Techniques) were used for the balanced distribution of unbalanced classes.

Each of these techniques offers a way to improve model performance on imbalanced datasets, helping to ensure that the minority class is adequately represented and accurately predicted. The choice of technique depends on the specific characteristics of the data set and the problem being addressed.

5 CONCLUSION

This research revealed that AI algorithms could help doctors and other healthcare professionals by predicting stroke across all examined texts. By moving beyond traditional methods that typically lack accuracy, are error-prone, and consume considerable time and resources, the development of ML models enhances their accuracy and efficiency in disease prediction, including strokes. This advancement enables prompt interventions, potentially lowering the mortality rate and complications associated with strokes. However, despite the obvious advantages of ML algorithms over classical statistical approaches, there is a pressing need to establish standards and protocols to improve the accuracy and sensitivity of data modeling in ML.

AUTHOR CONTRIBUTIONS

Farkhondeh Asadi: Conceptualization; Methodology. Milad Rahimi: Data curation; Formal analysis. Amir Hossein Daeechini: Writing—original draft; Writing—review and editing. Atefeh Paghe: Writing—original draft; Data curation.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflict of interest.

ETHICS STATEMENT

This research was approved by the research ethics committee of Shahid Beheshti University of Medical Sciences with ethics code IR.SBMU.RETECH.REC.1402.849.

TRANSPARENCY STATEMENT

The lead author Farkhondeh Asadi, Amir Hossein Daeechini affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained.