Published Paper Details:

Stroke is still a serious health issue and affects a large number of people every year. In many situations, it does not just lead to death but also leaves patients with long-term problems such as reduced mobility, memory issues, or emotional stress. These effects make daily life difficult and also increase the load on healthcare systems. Because of this, predicting stroke risk at an early stage becomes important, as it gives some time for preventive steps to be taken. In this work, a system is developed to estimate the risk of stroke using artificial intelligence techniques. The main idea is to provide some level of early support to doctors by using patient data, rather than depending only on symptoms that appear later. While working on this, one practical issue was the lack of proper datasets. Most datasets that were available were either incomplete or did not match the required parameters. So instead of relying on limited data, a synthetic dataset was created based on medically accepted ranges. While doing this, an attempt was made to keep the data realistic, but it is clear that it may not fully capture real-world variations. The dataset includes common risk factors such as age, gender, hypertension, heart disease, glucose level, BMI, and smoking habits. Before using the data, some preprocessing steps were necessary. Missing values had to be handled, categorical data was converted into numerical form, and scaling was applied where needed. During this step, it became clear that some models, especially Logistic Regression, were sensitive to how the data was scaled. Without proper normalization, the results were not stable. Three models were tested: Logistic Regression, Decision Tree, and Random Forest. These were selected mainly because they are widely used and easier to understand in healthcare-related problems. While testing, the Decision Tree model gave good results on training data but did not perform as consistently on test data, which indicated overfitting. Random Forest, on the other hand, gave more stable results across both training and testing. This is likely because it combines multiple decision trees, which helps reduce overfitting. Among all the models, Random Forest performed better overall, giving an accuracy close to 92%. It also showed better balance between precision and recall compared to the other models. Based on these observations, it was selected for the final system. The system itself is divided into stages such as input, preprocessing, prediction, and output. It provides a simple indication of risk, which can help in early-level decision-making. In future, this system can be connected to hospital records or wearable devices for continuous monitoring. However, one limitation of this work is that the model is trained only on synthetic data, so its performance with real patient data may vary. Overall, this work shows that machine learning can be useful as a supporting tool for early stroke risk prediction. At the same time, more testing with real clinical data is required before it can be used in practical situations.