Stroke_Prediction

• Author : Shehani Wetthasinghe
• Last modified: 01/04/2023

Source : https://www.kaggle.com/datasets/fedesoriano/stroke-prediction-dataset

The goal of the project

To predict the chance of having a stroke based on common risk factors using machine learning techniques.

Overview

According to CDC;

“A stroke, sometimes called a brain attack, occurs when something blocks blood supply to part of the brain or when a blood vessel in the brain bursts. In either case, parts of the brain become damaged or die. A stroke can cause lasting brain damage, long-term disability, or even death.”

Data Features

Feature	Description
Gender	Male / Female / Other
Age	Age of the patient in years
Hypertension	Yes(1) / No(0)
Heart Disease	Yes(1) / No(0)
Ever Married	Yes / No
Work Type	Children / Never worked / Government job / Private job / Self employed
Residence Type	Rural / Urban
Average Glucose Level	Average Glucose Level in Blood (mg/dL)
BMI	Body Mass Index (kgm -2 )
Smoking Status	TFormerly smoked/Never smoked/Smokes
Stroke (Target)	Yes(1) / No(0)

In the original dataset;

• contains 5110 rows and 11 columns
• 11 columns = 10 input features + 1 target variable
• From the total number of patients;
  • ~95% detected as negative for brain stroke (normal)
  • ~5% detected as positive for brain stroke (stroke)

Distribution of Key features

• Patients who got positive for brain stroke is approximately above 30 years old
• The risk of having a stroke is increasing with age

According to the CDC, patients can be categorized as;

• BMI is less than 18.5: Underweight range
• BMI is 18.5 to 24.9: Healthy Weight range
• BMI is 25.0 to 29.9: Overweight range
• BMI is 30.0 or higher: Obese range

• There is a risk of having a stroke if the BMI beyond the overweight range
• But there is also a chance of having a stroke even your BMI falls in healthy weight range

According to CDC the average glucose in blood can be catergorized into;

• Diabetes: above 126 mg/dL
• Prediabetes: 100 – 125 mg/dL
• Normal: Below 99 mg/dL

According the above plot, the risk of having a stroke is not much dependent on the average glucose level in blood

Correlations between key feartures

• There is a high risk on the people above ~45 years who suffering from diabetes to have a brain stroke
• There is a higher risk on the people above ~45 years who suffering from overweight to have a brain stroke

Machine Learning Models

Here, I used classification ML models such as;

• Logistic Regression
• K-Nearest Neighbor
• Random Forest
• XG Boost

Developments

• Reduce the number of features using PCA
• Synthetically generated new data to balance the dataset using SMOTE
• Optimize the default parameters of the models

NOTE: In this problem, the number of false negatives required to minimize to reduce the risk on the patient. Therefore, the recall score is the most important metric.

Performence of pre-optimized Models

• For all tested models, the best recall score obtained when they are incorporated with both PCA and SMOTE.
• Therefore, I optimized the paramaters of those models to enhance the performence further.

Comparison between pre and post optimized Models

Model	Precision_train	Recall_train	F1 Score_train	Accuracy_train	ROC AUC Score_train	Precision_test	Recall_test	F1 Score_test	Accuracy_test	ROC AUC Score_test
Logistic Regression	0.139	0.786	0.236	0.751	0.846	0.128	0.758	0.22	0.739	0.84
Optimized Logistic Regression	0.138	0.797	0.236	0.748	0.846	0.131	0.806	0.225	0.73	0.841
KNN	0.3	1.0	0.462	0.886	0.997	0.076	0.29	0.12	0.793	0.635
Optimized KNN	0.336	1.0	0.503	0.903	0.997	0.075	0.242	0.115	0.818	0.614
Random Forest	1.0	1.0	1.0	1.0	1.0	0.089	0.161	0.115	0.879	0.749
Optimized Random Forest	1.0	1.0	1.0	1.0	1.0	0.074	0.081	0.077	0.906	0.755
XGBoost	0.209	0.904	0.339	0.828	0.94	0.128	0.597	0.211	0.783	0.779
Optimized XGBoost	0.22	0.882	0.353	0.842	0.932	0.135	0.613	0.222	0.791	0.788

Conclusions

• Out of the 4 models, the optimized logistic regression model can be selected as best model according to the recall score and ROC score.
• Especially in this project, we are trying to maximize recall score to minimize the risk on patietns.
• The logistic regression models are the ones identified as the minimal overfitting models because their training scores are approximately similar with testing scores.
• The optimized logitic regression model not only gained the highest testing recall score, it also performed very well on testing data over training data according to the recall score.
• Therefore, optimized logistic regression model is the best model to predict the risk of stoke of patients.

Recomendations

• As per the findings of this project, the optimized Logistic Regression Model can be used as an early detection of brain stroke and direct to the treatments
• This model is developed to minimize the inaccurate predictions which may increase the severeness of the brain strokes
• The risk of have a brain stroke is increasing with age, BMI and average glucose level in blood
• Therefore, it is recommended to keep track on your glucose level in blood and BMI when you are getting older.