IV. Insights / Evaluation

Insights from Overall Approach:
Methods of replacing missing values may not be ideal.
For continuous data, missing values were replaced with its mean value for the following:
- `resting_blood_pressure`
- `cholesterol`
- `maximum_heart_rate`
- `ST_depression`
The missing data were replaced with the our sample mean. There was a possibility that it messes up the correlation of the attributes with having heart disease. The missing data could have been from healthy people as well as from people who have heart disease. Taking `cholesterol` for example, if a large proportion of people with missing data for `cholesterol` actually have heart disease, the expected `cholesterol` average for the missing data would be higher than the sample mean. This will skew the data for `cholesterol` and subsequently, the correlation of `cholesterol` and having heart disease.
For categorical data, missing values were replaced with "0" for the following:
- `fasting_blood_sugar`
- `exercise_induced_angina`
- `resting_electrocardiographic_results`
- `slope_peak_exercise` The issue is much worse for categorical data as it directly inserts data points that are either strictly healthy people or strictly with heart disease. For continuous data, when the replacement was done with the sample mean, it could somewhat capture the effects of both healthy and unhealthy people. Using `resting_electrocardiographic_results` as an example, having a value of 0 means that the `resting_electrocardiographic_results` are normal. People who actually have heart disease with a missing data point for the `resting_electrocardiographic_results` will have a replacement value of 0, indicating that the results are normal. However, this should not be the case. Clearly, it can be seen that the results will be skewed. This will decrease the reliability of these categorical attributes as a whole. A way to circumnavigate this issue is to obtain more data points. By having more data points, the sample mean that is used to replace the missing data for continuous variables will be closer to the actual mean and more accurately reflect the attribute's correlation with having heart disease. Similarly for categorical data, the same issue will still be present. However, having more data points will dilute the effect of the inaccurate data, causing it to skew our results to a smaller extent. In addition, it unlocks the option for us to drop these data points entirely which is currently not possible due to the already limited dataset available. Insights from Multivariate Analysis:
This model suggested that all attributes were weak indicators of predicting heart disease. With all attributes having weak correlations with each other and `diagnosis_of_heart_disease` (< 0.7), it also suggests that we should increase our sample size and experiment with a greater variety of more attributes (outside of the 14 original attributes) for further studies.
Insights from Decision Tree Analysis:
Through feature importance, both trees consistently predicted `chest_pain_type` and `cholesterol` as the top predictors of heart disease. However, we expected the Pruned Tree to provide more reliable assessments in metrics performance than the Full Tree. Given that both trees produced similar results in reliability (precision & accuracy), it is likely that there is insufficient data for a Pruned Tree to provide more precise results. Insights from Logistic Regression:
From Logistic Regression, we noted that `sex` (coefficient = 1.46) was the highest predictor of heart disease, suggesting that females were more likely to have heart disease. However as identified earlier during data visualisation, the dataset was poorly balanced with 79% of our dataset from females, and 21% from males. Future research should include a more even distribution between males and females.
Additionally, for `chest_pain_type` which was split via One-Hot Encoding, we identified that `CPT_4.0` (chest_pain_type 4: asymptomatic) had one of the highest coefficients (coefficient = 1.19). Realistically however, a person without heart disease could also be asymptomatic. Thus, using asymptomatic chest pains to indicate heart disease could detect alot of False Positives (actual no heart disease, predict heart disease) which would not be recommended.
Lastly, `CPT_2.0` (chest_pain_type 2: atypical angina) was found to have the most negative correlation (coefficient = -0.96) with heart disease, thereby having the least risk of heart disease. However our research suggested that `CPT_3.0` (chest_pain_type 3: non-anginal pain) was found to be the least related with heart disease. This suggests that our dataset may potentially be incorrectly labelled. It also suggests that `chest_pain_type` in this dataset may not be a helpful attribute in predicting heart disease, and should be dropped for future analysis despite being a top indicator in Decision Tree Analysis.