Early Prediction of Stroke Based on Deep and Machine Learning by Applying Medical Imaging and Tabular Data

In the past 20 years, stroke has become one of the top causes of mortality and lifelong disability worldwide. Only in China, there are 2 million patients diagnosed with stroke annually, and the mortality rate is 11.48%. Stroke is divided into ischemic kinds, which account for 85% of cases, and hemorrhagic types, which account for 15% of cases. Patients who have had a stroke are subject to long-term moral and financial responsibilities if they recover from the condition while others pass away as a result of a stroke surgery [1].

Stroke victims have a better chance of recovering from strokes if they seek medical attention as soon as possible. Any stroke victim should visit a doctor as soon as possible. Otherwise, death, long-term disability, and brain damage will occur. Various factors can cause strokes in patients. A stroke is caused primarily by diet, inactivity, alcohol, tobacco, personal history, medical history, and complications, according to the National Heart, Lung, and Blood Institute [2].

Due to the development of the field of artificial intelligence with its multiple branches of machine learning and deep learning, which includes neural, convoluted, and recurrent networks, as well as the field of data science and its analysis, in anticipating many health problems and detecting diseases in the early stages of the development of these diseases, stroke was and still is the subject of interest and research by many researchers Those who focused their efforts on this disease and the early detection of it because of its unfortunate consequences for people and society [3, 4].

Artificial intelligence research needs data analysis tools in performing tasks for large data, i.e., the sample approved by the responsible authorities is of a large size, and therefore, recently, solid sites have begun to publish data from approved and licensed agencies by the World Health Organization for research purposes, and the data volume may reach up to millions of people, and this is what makes the need for advanced and efficient algorithms with a high capacity to process all this data [5, 6].

This data includes a lot of information about patients, including personal information such as name, identification number, gender, and others related to the clinical symptoms shown in patients, while some of them include percentages of blood analyses, thyroid, associated infections, or medical history, in addition to other research that uses radiological images, scans, or magnetic resonance [7, 8].

Stroke is one of the diseases that has occurred recently. It poses a great danger due to the number of deaths that the disease causes in society, in addition to the fact that the person being treated is exposed to symptoms that accompany it, which may last a lifetime, such as loss of speech, paralysis, or other damages. The case of sudden cessation of blood flow constitutes a burden on people. Medical personnel are a case without introductions, so artificial intelligence models have been designed to predict stroke. In our proposed work, we emphasize early detection of stroke, which contributes to reducing the harm of this deadly disease, according to the World Health Organization [1, 2].

This paper focuses on predicting stroke with machine learning algorithms as well as with deep learning and comparing which of the results is better. Then, apply the principle of basic data analysis and enter it for processing again in those algorithms, and also compare the difference in the results obtained by the two methods, as this is a contribution to the field of practical application. For those algorithms, moreover, in both machine and deep learning algorithms, it will make changes to the parameters used, for example, the number of hidden layers and the type of optimization function.

In this work, we provided early diagnosis of stroke based on data, whether it was in the form of an Excel sheet in which several data were collected for a group of people, including one who was affected and the other was healthy, or it was magnetic resonance images, which contributes to the possibility of reaching good results by diversifying methods and databases. To change the parameters of each model to achieve the desired results. This study presents a model for diagnosing stroke with Our experiment showed good results, and it is possible to use these models in the future in health institutional contexts and integrate them with the mechanisms used in hospitals to produce rapid and accurate diagnostic results. An interactive programming environment Anaconda was used with a graphical interface and a textual interface supported by python-newt, which is freely available on the Internet and can also be easily downloaded to a personal laptop. Jupyter Notebook application: It is easy to use and follow the steps. It is worth noting that Jupyter Notebook is a programming application in the Python language.

The rest of the paper is organized as follows: The second section includes the works related to this paper; the third section describes the data used; the fourth section includes the presented methodology; the fifth section includes prediction algorithms; the sixth contains measures of achievement and results; the seventh compares previous works and the current work; and finally, conclusions and recommendations for future work are presented in the eighth section of this paper.

Table 1. Dataset 1 information used in proposed work

Table 2. Dataset information

1.png

Figure 1. Dataset 2 (normal and stroke brain) images

In fact, we present in the paper the algorithms that resulted in the best results in order not to explain all the algorithms without meaning to their results. Therefore, the presentation here is of the best among all machine learning algorithms.

As for neural networks, we still obtained good results through three layers and quarter layers, so there is no need to waste time and effort by implementing neural networks with more layers. As for the deep learning network, which is the convolutional neural network, specifically the VGG 16 convolutional neural network, which is distinguished in terms of performance because it has depth. Almost as good, as it only has 16 depth layers that have weights instead of relying on a large number of redundant parameters. Moreover, the selection of parameters for training of the compiler contains specifications of: optimizer and loss.

We find that it showed the best results in the other path because the training was done by adjusting the weights, and thus we obtained a high accuracy rate and a very low data loss rate to reach the highest level of control without deficiency or over-fitting.

5.1 Machine learning methods

There are various methods for detecting disease cases, whether they affect humans, plants, or animals, but machine learning methods remain the main field of research in this topic, as they have established the foundations for these studies [16, 17]. The following are the machine learning algorithms that were implemented in this paper:

(1) K-Nearest Neighbors (KNN): K-NN is a form of slow learning where no pre-processing stage is specified and all calculations are saved for classification. This data categorization technique makes decisions based on the nearby training data points on the feature map. The target class is predicted by the K-NN classifier using the Euclidean distance metric. The best value for parameter k, which governs how well the classifier performs, is determined by the dataset. After examining the impacts, the optimal value is then established [18].

(2) Support Vector Machine (SVM): The support vector machine is one of the supervised machine learning techniques that categorize unknown input using labeled data. Decision limits are expressed using the notion of decision levels, or hyperplanes. It is important to note that a set of data objects can be divided into many classifications using the superlevel. The supported vector machine was employed in this study to categorize the data by developing a function that, with the least amount of error and maximum percentage of accuracy, labels each data point with the proper label [19, 20].

(3) Decision Tree (DT): Since this study focuses on stroke prediction, decision trees can be utilized as both classification and regression methods until all features, including data, are finished, each stage of the contract is concluded with a question-and-answer session [21, 22].

(4) Random Forest (RF): The so-called random selection forest, sometimes known as the random forest, is a classification and regression procedure that, in reality, consists of a lot of decision trees. Each tree offers a solution to a particular issue. The best choice is to classify each tree according to its output, which is the best solution from the set of tree solutions. In order to improve the performance of classifiers and the accuracy of classification or prediction, group learning provides a more accurate method. In order to achieve this, a random forest classifier is used, in which many decision trees are grouped on different subsets of data and averaged to improve the accuracy expected [4, 5].

(5) Logistic Regression (LR): The logistic regression technique is used in classification problems, where the output is in the form of probability, which is then translated into one of the categories. Upon hearing about this algorithm, the initial assumption is that it is utilized for solving regression issues. It is a method that uses the sigmoid function to estimate the relationship between pairs of dependent variables as well as at least one individual variable, and it is amenable to supervised training. It is used to resolve regression issues in addition to classification issues. It might fall under one of the three categories in terms of classification [7, 19].

• Binary classification
• Multiple classifications
• Ordinal classification

(6) Naïve Bayes (NB): One of the methods of supervised machine learning, which relies in its work on the principle of Bayes' theory, is the probability theory that predicts the occurrence of an event based on a prior set of conditions associated with that event [20, 21].

(7) XGBoost: XGboost is a fantastic example of a gradient augmentation technique. The gradient gain alternative can be carefully planned for accuracy and optimization, even though there may not be any ground-breaking mathematical breakthroughs in this scenario. It consists of a linear version, and the newborn tree may be a technique that uses various AI algorithms to determine whether a flimsy newborn would produce a reliable newborn in order to improve the correctness of the version. For instance, parallel learning and impulsive learning can both be used to learn random forest [19].

(8) CatBoost: Hancock and Khoshgoftaar's [23] gradient boosting technique known as CatBoost works well with categorical characteristics by reducing information loss. CatBoost is distinct from other gradient-boosting techniques. It initially uses ordered boosting, a useful variant of gradient boosting methods, to handle the problem of target leakage. This technique also does well with small datasets. Thirdly, CatBoost can control category features. This handling is typically completed in the preprocessing stage and comprises converting the original category variables to one or more numerical values [23, 24].

5.2 Deep learning methods

Deep learning is learning by means of artificial neural networks when it includes several layers called by this name, and deep learning is learning using neural networks and it has three types according to the type of network and how to process it [25-27].

• Deep learning using neural networks
• Deep learning using convolutional networks
• Deep learning using iterative networks

Deep neural networks with several layers are used in order to obtain better results and high accuracy. As the number of layers of the neural network increases, the complexity increases, and the possibility of obtaining high accuracy in training and examining data increases [28, 29]. The artificial neural network consists of several layers starting from the input layer and the middle layers known as the hidden layers and the output layer, the big change is usually made in the middle layers where more than one layer is used and the parameters used are changed in it from the number of neurons and the function used, as well as what is known as cutting neural connections with the function known as dropout. This paper presents deep learning for both types of data used for such purposes. Once the data is trained in the form of Excel sheets, and another time the data is trained in the form of magnetic resonance images (MRI) by using convolutional neural networks (CNN), especially VGG16, as shown in Figure 3, then the results obtained from all methods were compared [30].

3.png

Figure 3. VGG-16 components

Table 6. Comparison results with previous works