A Classification Machine Learning Model to Predict Heart Disease Heuristics
The blog post describes a predictive modeling approach using Supervised Learning
MOTIVATION: In order to accurately classify Heart Disease patients with the presence or absence of the pathological condition, a predictive modeling framework is entailed. Here in this blog post I present a Supervised Machine Learning approach which is developed in order to group patients with and without the disease. The data set pertains to that of "Processed Cleveland Heart Disease" and is made publicly available from the UCI ML repository: (https://archive.ics.uci.edu/ml/datasets/heart+disease).
WORKFLOW & QUESTIONS OF INTEREST: Exploratory data analysis is first performed on the chosen data set mentioned above, followed by the necessary data preprocessing and data transforms. The final data set is studied using various supervised algorithms. Using K-Fold Cross Validation various algorithms were evaluated to better predict the accuracy of the final model. Overall, the goal of this blog post is to demonstrate the end-to-end approach for a supervised classification model.
METHODOLOGY: The data set consisted of 302 records of patients with 14 attributes. The attributes comprise of the following variables - (1) Age, (2) Sex, (3) Type of Chest Pain, (4) Resting Blood Pressure, (5) Serum Cholesterol, (6) Fasting Blood Sugar, (7) Resting ECG results, (8) Maximum Heart Rate, (9) Exercise induced Angina, (10) ST Depression, (11) Slope of the exercise peak ST segment, (12) the number of major vessels colored by flouroscopy, (13) Thal, and (14) Diagnosis of the Disease. The following table summarizes the meaning of each variable.
|Age||Numerical||Age in years|
|Type of Chest Pain||Categorical||
1= Typical Angina
|Resting Blood Pressure||Numerical||in mm Hg on admission|
|Serum Cholesterol||Numerical||in mg/dL|
|Fasting Blood Sugar||Categorical||(>120 mg/dL) (1=true 0=false)|
0 = Normal
|Maximum Heart Rate||Numerical|
|Exercise Induced Angina||Categorical||1=yes 0=no|
|Slope from ST Segment||Categorical||
1 = Up Slope
|Major Vessels colored||Numerical||0-3|
3 = Normal
|Diagnosis of Disease||Categorical||1,2,3,4=yes 0=no|
- MISSING DATA & PREPROCESSING: Few records of missing data were noted for the variables of "Major Vessels Color" and "Thal". Since the number of records were less than 5 these values were assigned the respective numerical/categorical variables randomly. Further, since the objective of the model is to classify patients into two groups, the various groups of patients with heart disease (i.e., with response of 1, 2, 3 and 4) are grouped into a single one.
- The code can be found in the following link: (https://github.com/uppulury/ClevelandHD)
- DESCRIPTIVE STATISTICS: The statistics of each feature attribute is studied and figures below depict the distribution for each one of them.
- Quite clearly the continuous variable have a Gaussian like distribution. Among the categorical variables there is class-imbalance in some variables such as Thal, ST Slope, Resting ECG and Fasting Blood Sugar. Further, the map of correlation among the variables evince strong correlations such as ST Depression and ST Slope.
- SPOT-CHECK ALGORITHMS: The preprocessed data was split into training and validation sets using Python's train_test_split in the ratio of 80/20. Classification algorithms were applied to the training set using 10-Fold Cross Validation technique to evaluate the model accuracy. Logistic Regression produced the highest accuracy where as SVM produced the lowest accuracy on the data set.
- In order to improve the accuracy of the model the data set was standardized using Python's Standard Scalar. The same set of classification algorithms were applied and the model accuracy overall improved across all algorithms very likely owing to the underlying Gaussian structure in the variables. The SVM algorithm produced the best fit to the data set with a slightly higher accuracy and low variance compared to the other classification algorithms.
- FINE-TUNING THE MODEL:
- (1) LR Fine Tuning - no improvement
- (2) kNN Fine Tuning - only a very small improvement
- Overall, the biggest jump in model improvement seems to be for SVM
- Ensemble Methods were implemented: Gradient Boosting and Extra Trees Classifiers produced the highest accuracy. These accuracy scores are similar to the SVM employing the Standardized data set.
- Ensemble Methods were Predicted and Scored on the Training and Validation Data Sets: not much improvement is seen here from the 10-fold CV approach.
- Ensemble Methods were Predicted and Scored on the Standardized Training and Validation Data Sets: not much improvement in accuracy is seen for Ensemble algorithms on standardized data.
- FINE-TUNE MODEL USING SVM: A GridSearchCV was performed on a range of c_values and various kernel_values using SVM algorithm. The optimal performance was achieved for C=0.1 and a Sigmoid kernel. These parameters were used in the Finalized model to predict and score the algorithm on the Validation Data Set. The accuracy of the final model is 80%.
CONCLUSIONS: Most algorithms including the Ensemble methods produced high accuracy for the model. The best case accuracy was achieved using an SVM model on the standardized data set. The final results herein presented are for the SVM model and corresponding c_value=0.1 and a 'sigmoid' kernel. Following is the Confusion Matrix from the finalized model: