🩻🩺 Stroke Prediction Model 🩺🩻

📋 Abstract

This research implements machine learning methodologies to develop a predictive model for stroke risk assessment utilizing clinical and demographic features. The study addresses the critical healthcare challenge of early stroke detection through advanced statistical modeling and machine learning techniques.

📝 Methodology

Data Characteristics

The dataset comprises heterogeneous patient records with continuous and categorical variables. The target variable exhibits class imbalance with a 1:10 ratio of stroke to non-stroke cases, as illustrated in Figure 1.

Figure 1: Distribution of target variable demonstrating class imbalance

Feature Analysis

The dataset includes both continuous and categorical variables:

• Continuous Variables: age, average glucose level, BMI
• Categorical Variables: gender, residence type, work type, marital status, smoking status

Distribution analysis of key continuous variables revealed significant patterns, as shown in Figure 2.

Figure 2: Distribution of average glucose levels across the patient population

Feature Engineering

• Continuous Variables:
  • Standardization applied
  • Missing value imputation via median substitution
• Categorical Variables:
  • One-hot encoding implementation
  • Mode-based imputation for missing values

Model Development

Multiple classification algorithms were evaluated:

1. Logistic Regression with balanced class weights
2. Random Forest with class-weight optimization
3. XGBoost with gradient boosting
4. Synthetic Minority Over-sampling Technique (SMOTE) combined with Logistic Regression

📊 Results

Performance Evaluation

Models were assessed using multiple metrics, with the Random Forest classifier's performance illustrated in Figures 3 and 4.

Figure 3: ROC curve demonstrating model discrimination capability

Figure 4: Precision-Recall curve highlighting model performance on imbalanced data

Classification Results

The confusion matrix (Figure 5) provides detailed insight into the model's classification performance.

Figure 5: Confusion matrix showing classification outcomes

Model Calibration

Probability calibration analysis (Figure 6) revealed slight underconfidence in probability estimates.

Figure 6: Calibration curve showing probability estimation reliability

Feature Attribution

Analysis revealed age, blood pressure, and glucose levels as primary predictive indicators, consistent with established clinical literature (Figure 7).

Figure 7: Feature importance rankings from Random Forest classifier

👨‍💻 Conclusion

The Random Forest classifier demonstrated superior performance in balancing sensitivity and specificity. However, identification of certain stroke cases remains challenging, as evidenced by the confusion matrix metrics (14 true positives versus 37 false negatives). The study's findings contribute to the growing body of literature on machine learning applications in preventive healthcare.

The research highlights the potential and limitations of machine learning in clinical prediction tasks, suggesting areas for future investigation in feature engineering and model optimization.