Heart disease deaths have been rising in Singapore (Singapore Heart Foundation, n.d.) and detecting cardiovascular disease (CVD) early can help to manage its entailing consequences (Sarrafzadegan & Mohammmadifard, 2019).
Our model finds its main uses in 2 applications.
First, for more targeted identification of candidates for health screening. Our model can be integrated into HealthHub, a mobile application by the government, to allow users to do a self-assessment for their Cardiovascular Diseasee risk. Users who have been found to be at high risk of Cardiovascular Disease can then be recommended to go for health screening for further follow-up actions to be taken. The data for this can be updated into HealthHub through the integration of Health applications that are already available and widely used for regular monitoring on both iOS and Android platforms.
Second, for medical triaging at healthcare institutions so that patients who are at higher risk of cardiovascular disease complications can be attended to faster. This could come in the form of earlier medical appointment slots or priority queues thereby reducing the chances of a severe event from occurring.
Our video can be accessed here: https://youtu.be/Wj-rIJu2epg
Before processing the data, it should be noted that there are a couple different types of inputs:
- Objective: Factual information;
- Examination: Results of medical examination;
- Subjective: Information given by the patient.
The data is organised into multiple columns representing various factors. These factors and their datatypes are:
- Age | Objective Feature | age | int (days)
- Height | Objective Feature | height | int (cm) |
- Weight | Objective Feature | weight | float (kg) |
- Gender | Objective Feature | gender | categorical code |
- Systolic blood pressure | Examination Feature | ap_hi | int |
- Diastolic blood pressure | Examination Feature | ap_lo | int |
- Cholesterol | Examination Feature | cholesterol | 1: normal, 2: above normal, 3: well above normal |
- Glucose | Examination Feature | gluc | 1: normal, 2: above normal, 3: well above normal |
- Smoking | Subjective Feature | smoke | binary |
- Alcohol intake | Subjective Feature | alco | binary |
- Physical activity | Subjective Feature | active | binary |
- Presence or absence of cardiovascular disease | Target Variable | cardio | binary |
After importing the data, we first removed all the duplicates by comparing ID values. This removed 24 values.
After that, we modified the gender column to represent male and female as well as the age column to represent age in terms of years instead of days.
From here we moved to some basic EDA and Data Cleaning. Visualising the correlation between variables in the form of a heatmap gave us this:
From here, we moved on to cleaning the data to better represent some of these correlations. The steps taken were:
-
Adjusting for outliers of the height and weight variables by comparing the variables to each other and to the age of the individual. Further justification through violinplots revealed that there are 322 outliers which were removed.
-
Adjusting for blood pressure, we compared blood pressure against the world average and clinically understoof ranges of optimality for systolic and diastolic blood pressure. After that, through the verification using both box and violin plots, we were able to identify 1479 values that were considered to be outliers which were also removed.
First, by plotting out the categorical variable of "cardio" in the form of a pie-chart. The result showed that the presence of cardiovascular disease was 50.6% of the dataset and the absence was 49.4% of the dataset which reduces classification bias.
Second, by analysing the new dataset, various statistics such as the mean age of patients, gender ratio, mean height, mean weight etc. were able to be derived. Interestingly, these values are reflective of normal everyday figures of large population groups.
Third, by comparing the new data's correlation with the old through the use of the heatmap, we get this:
This highlights the presence of Systolic and Diastolic blood pressure playing an important role in determining if an individual would have a cardiac episode or not. However, neither of the two variables guaranteed it.
Therefore, the need for a model which can take into account these various factors in tandem to be able to create a prediction for an individual.
We evaluated 5 models - Logistic Regression, Classification Tree, Random Forest, K-Nearest Neighbour and Neural Network.
A supervised machine learning model
It is suitable for our problem because we have a categorical dependent variable (cardiovascular disease/no cardiovascular disease).
This model makes use of a Sigmoid function and is very efficient to make predictions
A supervised machine learning model.
Good for helping to find data that can be omitted to further optimise predictions, based on gini coefficient.
2 layers of classification were used.
Greedy strategy of classification which may not be optimal.
This model gave us the highest accuracy and this is what we chose.
K-Nearest Neighbour classification (KNN) is a supervised machine learning model that is based on assumption that data points with similar characteristics are close to one another.
The model allows for dealing with variables that do not necessarily have a linear relationship with cardiovascular disease. A clear evidence of this is the finding that smoking does not have a linear relationship with cardiovascular disease. (Kondo et al, 2019).
The model can also offer a high level of accuracy provided that the optimal K value is used. It must be noted that by having a very small number of neighours, it could lead to high variance and low bias, which denotes overfitting. By having a large number of neighbours, it could lead to high bias and low variance, which denotes underfitting.
The model calculates the distance from the unknown data point to its neighbours and K number of the data points nearest to the data point will be used to determine the classification of the unknown data point. The majority class of the neighbouring (K) data points will be used as the designated class for the unknown data point.
In our project we calculated the accuracy of the KNN model with different K values from K=1 to K=199. The highest accuracy was achieved at K=103 at 73.2%.
After cleaning the data, we split the data into 30% for test data and 70% for training. We then trained the model, then used it on the test data. The model outputs a list of 0s and 1s with 0 denoting no cardiovascular disease and 1 denoting cardiovascular disease.
The KNN classifier function we used is from the sklearn library.
- Sensitive to irrelevant features The accuracy of the model can be affected if irrelevant features are used.
The dataset should be thoroughly cleaned and the correlation between each feature and cardiovascular disease should be assessed prior to using the feature for training the model.
- Computationally Intensive The model has to store all the data that was used for its training. This can cause finding the optimal K value to take a very long time.
As a mitigation, there should be a cut-off threshold established when searching for optimal K value. For example, if K=103 gave the highest accuracy and from K=104 to K=109, the accuracies are no higher than the highest, the search can cease and use K=103 as the optimal value.
This can help to save time when using the model, which could prove crucial in the context of real world application. A patient who is found to be at risk can be quickly attended to by medical professionals.
We evaluated the 5 models and we chose KNN as our model. We firmly believe that it can achieve good accuracy in predicting cardiovascular disease.
Clone the repository and open the ipynb file on Jupyter Notebook.
MIT
Singapore Heart Foundation. (n.d.). Heart disease statistics. Singapore Heart Foundation. Retrieved April 13, 2023, from https://www.myheart.org.sg/health/heart-disease-statistics/
Kondo, T., Nakano, Y., Adachi, S., & Murohara, T. (2019). Effects of tobacco smoking on cardiovascular disease. Circulation Journal, 83(10), 1980-1985.
Sarrafzadegan, N., & Mohammmadifard, N. (2019). Cardiovascular disease in Iran in the last 40 years: prevalence, mortality, morbidity, challenges and strategies for cardiovascular prevention. Archives of Iranian medicine, 22(4), 204-210.