Stroke-Prediction

A Web based Hybrid Machine Learning and Deep Learning Approach for Stroke Prediction. It addresses data imbalance, compares classical and neural models and applies hyperparameter optimization.

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

---

Libraries used for dataset processing

• Numpy
• Pandas

Libraries used for graphical representation

• Matplotlib
• Seaborn

Libraries used for Scaling and Oversampling

• Sklearn.preprocessing
• Imblearn

---

PREPROCESSING

---

• Removed the id column – decreasing the dimension – did not add to insights in the data analysis.

df = df.drop(['id'],axis=1)

• Count for NULL values are checked among the attributes of the dataset

print(df.isna().sum())

• Only BMI-Attribute had NULL values
  
• Plotted BMI's value distribution - looked skewed - therefore imputed the missing values using the median.
  
• Didn’t eliminate the records due to dataset being highly skewed on the target attribute – stroke and a good portion of the missing BMI values had accounted for positive stroke

• The dataset was highly biased because there were only few records which had a positive value (4.97% of 100%) for stroke-target attribute

• In the gender attribute, there were 3 types - Male, Female and Other. There was only 1 record of the type "other", Hence it was converted to the majority type – decrease the dimension

• Most of the attributes in the dataset were binary values – converting the numeric bin values into string bin values for dummy encoding.

  • Dummy encoding similar to one-hot encoding – Values in the binary ecoded columns are 1/0 – Additional attributes/columns created.

• SMOT done on the dataset to balance the skew in the target attributes.

  • Boosting the number of records in the minority class – records

---

EDA - Exploratory Data Analysis

---

• Plotted plots of each attribute - Analyse trends if any – plots: pie, histogram.
• Plotted relation of target attribute to other attributes to find any correlation.
• Plotted the heatmap – correlation plot between the attributes.
• Heatmap showed very less correlation between the attribute values.
• A hybrid Random Forest + Artificial Neural Network (RF+ANN) model was implemented to combine tree-based feature learning with neural representation capacity. The hybrid approach achieved 98.51% accuracy, demonstrating competitive performance alongside the optimized Random Forest baseline.

---

MODEL BUILDING

---

• Creating a train and test split of the oversampled dataset. (80-20)

Applied various Machine learning models for predictive analysis

1. Logistic Regression
2. Decision Tree
3. Random Forest
4. SVM
5. Naive Bayers
6. ANN
7. RF+ANN

Analysed the results generated using confusion matrix - accuracy, precision, recall and plotting the ROC plot and generating the AUC scores.

Accuracies calculated:

1. Logistic regression : 76.5%
2. Decision Tree : 96.97%
3. Random Forest : 98.62%
4. SVM : 77.3%
5. Naive Bayers : 76.2%
6. ANN : 78.04%
7. RF+ANN : 98.51%

Chosen model - RANDOM FOREST

The accuracy of random forest model increased from 98.62% to 99.02% using hyperparameter tuning with GridSearchCV