0. Heart Failure Prediction in Patients

1. Model Description

In South Africa, cardiovascular disease is the second-highest cause of death behind only to HIV/AIDS. Heart failure contributes significantly to this, so what causes heart failure?

This project's aim is to answer that question. So by using a dataset that has 12 features, ranging from the patient's age to their exercise angina and their maximum heart rate, this project starts by examining feature importance with the desire to see the factors that contribute more to a positive heart failure likelihood. It then moves on to testing different scikit learn models, so that it can find a model that more accurately describes the dataset. It eventually ends with an XGBoost model that has an roc score of at least 94%, that is a model that will 94/100 times correctly predict whether a patient is likely to suffer heart failure or not.

2. Dataset Features Description

The dataset has 918 records, with 12 features:

• age: age of the patient (in years).
• sex: sex of the patient (m: male, f: female).
• chest pain type: chest pain type (ta: typical angina, ata: atypical angina, nap: non-anginal pain, asy: asymptomatic).
• resting bp: resting blood pressure (mm Hg).
• cholesterol: serum cholesterol (mm/dl).
• fasting bs: fasting blood sugar (1: fasting bs > 120 mg/dl, 0: fasting bs < 120 mg/dl).
• resting ecg: resting electrocardiogram results (normal: normal, st: having ST-T wave abnormality, lvh: showing probable left ventricular hypertrophy by Estes' criteria).
• max hr: maximum heart rate achieved (in the range 60 and 202).
• exercise angina: exercise-induced angina (y: yes, n: no).
• old peak: old peak = extent of ST depression induced by exercise relative to rest.
• st_slope: the slope of the peak exercise ST segment (up: upsloping, flat: flat sloping, down: down sloping).
• heart disease: Target Variable (1: heart disease, 0: normal heart).

3. Exploratory Data Analysis

In this part:

• column headers are converted to lowercase alphabets and whitespaces are replaced by underscores.
• features are classified into 2 classes: categorical_variables and numerical_varibles.
• attributes in the dataset are then also converted to lowercase alphabets, and whitespaces replaced by underscores.
• the dataset is then checked for missing values.
• descriptive statistics performed on the dataset using pandas' .describe() method.

4. Splitting The Dataset

The dataset is split into 3 sets of size 60/20/20 using scikit-learn's train_test_split() function.

• 60% training set.
• 20% validation set.
• 20% test set.

5. Feature Importance

In this part:

• global rate is calculated to check for class imbalance.
• mutual information: is calculated to measure importance of categorical variables.
• correlation: is calculated to measure importance of numerical variables.

6. Feature Matrix

In this part, the DictVectorizer() function is used to create 3 matrices from 3 dictionaries.

• the training set is first converted into a dictionary using pandas' to_dict() function, and then converted into the matrix X_train.
• the validation set is first converted into a dictionary using pandas' to_dict() function, and then converted into the matrix X_validation.
• the test set is first converted into a dictionary using pandas' to_dict() function, and then converted into the matrix X_test.
• the feature matrix is then extracted from dictVectorizer using DictVectorizer()'s feature_names_ function.

7. Models Evaluated on The Dataset

In this part, models are evaluated and then their performance is compared using their roc_auc_score.

• Logistic Regression: first on the baseline model, and then parameter tuning on the C and solver parameters.
• Decision Tree: first on the baseline model, and then parameter tuning on the max_depth and min_samples_leaf parameters.
• Random Forest Tree: first on the baseline model, and then parameter tuning on the max_depth and n_estimators parameters.
• XGBoost: first on the baseline model, and then parameter tuning on the eta, max_depth, and min_child_weight parameters.

8. Final Model

The best-performing model is selected to be the XGBoost model with the following parameters, this model is tested on the test set:

• eta = {0.3}.
• max_depth = {3}.
• min_child_weight = {1}.

9. Python Scripts

• notebook.ipynb script is exported into train.py.
• the final model is saved as heartfailure_model.bin using pickle.
• predict.py uses FastAPI to predict the patient's heart failure probability.