In this program I implement two machine learning models (Neural Network & KNN) which I then train on a Cleveland heart disease database in order to predict whether or not a patient has Heart Disease.
The Cleaveland database has 76 attributes however most experiments use 14 of them. My Neural network and KNN models are trained on the data with the most commonly used 14 attributes. It gives a binary classification (Heart Disease present or not present). After testing the models and tweaking the parameters I got on average an accuracy of 91% for the Neural Network and 89% for KNN.
1.Age in years
2.Sex
3.Chest pain type
1 if typical angina
2 if atypical angina
3 if non-anginal pain
4 if asymptomatic
4.Resting blood pressure (in mm Hg on admission to the hospital)
5.Serum cholesterol in mg/dl
6.Fasting blood sugar > 120 mg/dl
7.Resting ECG results
8.Maximum heart rate achieved
9.Exercise induced angina
10.ST depression induced by exercise relative to rest
11.The slope of the peak exercise ST segment
12.Number of major vessels (0-3) colored by flourosopy
13.Thal
14.(Target Value) diagnosis of heart disease (angiographic disease status)
Target value =
0 if < 50% diameter narrowing
otherwise 1
The target 🎯 values are [0,1,2,3,4]. 0(min-value) corresponds to no presence of heart disease in a patient and 4(max-value) corresponds to the highest level of certainty of the presence of heart disease.
I changed this multiclass classification into a binary classification because of how the data is used in experiments.
"Experiments with the Cleveland database have concentrated on simply attempting to distinguish presence (values 1,2,3,4) from absence (value 0)."
The following code demonstrates how I did this:
new_targets = np.where(old_targets != 0, 1, old_targets)
There were not too many missing values in the original dataset however in order to get rid of the few there were I first had to convert them to NaN because they were marked as (?).
The following code demonstrates how I did this:
dataSet.replace('?', np.nan, inplace=True)
Then I replaced all the NaN values with the mean of each column using SimpleImputer. I chose to replace them with the mean since it would be the best guess as to what the values would be.
The following code demonstrates how I did this:
replace_missing= SimpleImputer(strategy='mean')
clean_dataset = pd.DataFrame(replace_missing.fit_transform(dataSet), columns=dataSet.columns)
The 84% accuracy is for the validation set. The table is the classification report and the value 0.89 in the row labeled accuracy means that the model has an 89% accuracy on the testing set.
The first graph measures the Loss over epoch which essentially helps visualize how well the Neural Network along with indicating whether the model is overfitting. For example if the validation loss increases over epochs while the training loss decreases then there is a possibility that the model is overfitting.
The second plot shows the accuracy of the model on the validation and training sets. There should not be a large discrepancy between the two since that would mean the data is not properly split.
93% test accuracy was the most common percentage I received for the Neural Network:
Python🐍, Tensorflow, Matplotlib, Scikit-learn, Numpy, Pandas🐼, Imblearn
Janosi,Andras, Steinbrunn,William, Pfisterer,Matthias, and Detrano,Robert. (1988). Heart Disease. UCI Machine Learning Repository. https://doi.org/10.24432/C52P4X.