The main objective of this project is to build a binary classifier model, which predicts whether a person is at a 10-year risk of future coronary heart disease (CHD) on the basis medical history and other information.
The dataset provides the patient's information about following features.
- Sex : male or female ("M" or "F") (Nominal)
- Age : Age of the patient
- Education : Education level of the patient ( 1: lowest to 4 : highest) (ordinal)
- is_smoking : whether or not the patient is a current smoker ("YES" or “No”) (Nominal)
- CigsPerDay : the number of cigarettes that the person smoked on average in one day.
- BPMeds : whether or not the patient was on blood pressure medication (0 or 1) (Nominal)
- prevalentstroke : whether or not the patient had previously had a stroke (0 or 1) (Nominal)
- prevalenthyp : whether or not the patient was hypertensive (0 or 1) (Nominal)
- Diabetes : whether or not the patient had diabetes (0 or 1) (Nominal)
- TotChol : total cholesterol level (Continuous)
- sysBP : systolic blood pressure (Continuous)
- diaBP : diastolic blood pressure (Continuous)
- BMI : Body Mass Index (Continuous)
- HeartRate : heart rate (Continuous)
- Glucose : glucose level (Continuous)
- tenyearchd : 10-year risk of coronary heart disease CHD (binary: “1”, means “Yes”, “0” means “No”) - DEPENDENT VARIABLE
- Total number of rows in data : 3390
- Total number of columns : 17
- No duplicate rows were present in dataset.
- Null values were found in following columns
Education,BPMeds- Categorical featurescigsPerDay,totChol,BMI,heartRate,glucose- Numerical features
To remove null values I have used SimpleImputer using
- mode for categorical features
- median for numerical features
- Used Label encoding to transform categorical features -
sex,is_smoking.
- Used square root transformation to reduce skewness of
cigsPerDay. - Used logarithmic transformation to reduce skewness of
totChol,sysBP,BMI,heartRate,glucose.
- Used
StandardScalerfor rescaling of features.This step scales data into a uniform format that would utilize the data in a better way while performing fitting and applying different algorithms to it.
Visualization of data were mainly performed using Matplotlib and Seaborn libraries and the following graph and plots had been used:
-
Bar Plot.
-
Histogram.
-
Scatter Plot.
-
Line Plot.
-
Heatmap.
-
Box Plot
-
Pair Plot
-
Out of 3390 given patients records almost 500 patients had cardiovascular disease. Hence, the dataset found to be imbalanced.
-
Out of 87 diabetic patient 35 patients had cardiovascular disease.
-
Out of 100 patients on BP medications 34 had cardiovascular disease.
-
Out of 22 hypertensive patients 10 had cardiovascular disease.
-
With increasing age risk of cardiovascular disease found to be increasing.
-
diaBPandsysBPfound to be stongly correlated. -
sysBPfound to be moderately correlated withprevalenthyp, i.e. prevalent hypertension. -
glucoselevel also found to be moderately correlated to whetherDiabetes. -
Risk of cardiovascular disease found to be not strongly depending on eduacation level.
-
Men are found to be at higher risk of cardiovascular disease.
-
Smoking increases risk of cardiovascular disease.
Feature Selection is crucial step before model training as it improves the performance of model and also to reduce multicollinearity in dataset.
Feature selection increases the predictive power of machine learning algorithms by selecting the most important variables
and eliminating redundant and irrelevant features. For seelcting features I have used `Chi- Square test` for categorical features and `corraltion heatmap` for numerical features.
(a) Chi-Square test (for categorical features):
(i) Found p-value of `is_smoking` very high. Therfore dropped it.
(b) Correlation heatmap (for numerical features):
(i) Dropped 'diaBP` due to high correlation with `sysBP`.
Also `id` column is dropped as it is unqiue for each row and therefore does not provide any information in training model.
Input features used for model building are as follows:
- Sex
- Age
- Education
- CigsPerDay
- BPMeds
- prevalentstroke
- prevalenthyp
- Diabetes
- TotChol
- sysBP
- BMI
- HeartRate
- Glucose
Imbalanced dataset can make model highly inaccurate. Therefore handle imabalanced dataset we have used SMOTE boosting.
Since we have to predict whether person has a 10-year risk of future coronary heart disease (CHD) , so this is a binary classification problem. Based on the problem statement which is basically to predict whether person has a 10-year risk of future coronary heart disease (CHD) I decided to develop predictive models using following algorithms and have compared there performance.
- Logistic Regression
- SVM Classifier
- Random Forest Classifier
- Gradient Boosting Classifier
- KNN Classifier
Following necessary steps have been employed for better model performance.
(i) train test split for model evaluation.
(ii) Hyperparameter tuning for best learning parameters that developed a better model. Hyperparameter tuning was done with the help
of RandomizedSearchCV.
(iii) Since the dataset is inbalanced, therefore default value of threshold(0.5) can result in inaccurate model. Therefore, I have tuned threshold value also to make models more accurate.
To evaluate the models I have used Recall, F1-score and roc_auc_score.
(i) All models had lower recall score due to imbalanced data and using threshold = 0.5 (default)
(ii) Logistic regression was most accurate as compared to other models
Recall for test dataset = 0.6773
roc- auc score for test dataset = 0.6803
f1- score = 0.3554
(ii) Using lower threshold value (< 0.4) improves recall significantly and f1-score also improves slighlty.
Recall for test dataset = 0.79
f1- score = 0.37
(i) Handling imbalanced data was a tedious task.
(ii) Building models using small size dataset was difficult.