Heart Disease Estimation with Logistic Regression: R Shiny App
The question behind this work is whether we can use the health information at our disposal to predict heart disease more accurately than we have been able to so far. To provide some context and highlight the scope of the problem, here are some statistics from the CDC about heart disease in the United States:
- Heart disease is the leading cause of death for men, women, and people of most racial and ethnic groups in the United States.
- One person dies every 34 seconds in the United States from cardiovascular disease.
- About 697,000 people in the United States died from heart disease in 2020—that’s 1 in every 5 deaths.
- Heart disease cost the United States about $229 billion each year from 2017 to 2018. This includes the cost of healthcare services, medicines, and lost productivity.
To sum it up, there is a tremendous annual loss of life, money, and productivity every year due to heart disease.
Not surprisingly, it turns out that this is not an easy problem to solve. A 2020 study at UT Southwestern showed that using sophisticated genetic testing does not greatly improve predictions based on traditional risk factors like high blood pressure, cholesterol levels, diabetes, and smoking status (1). A 2019 study of about 423,000 UK biobank records only achieved an area under the curve for the ROC of .774 (2). In looking for research using more sophisticated methods like neural networks and machine learning, the articles I found that achieved models with accuracy closer to 100% tended to be ones based on small data sets with only about 300 observations (3, 4). Based on these studies, unfortunately, we should not expect high accuracy even if we do all the modeling right.
I challenged myself to make an app that estimates the probability of heart disease with user input. Making a binary outcome variable of heart disease versus no heart disease meant that I would need to use logistic regression as machine learning is currently beyond the scope of my training. To make an interactive app, I chose to use an R Shiny interface.
Originally, the dataset came from the CDC and is a major part of the Behavioral Risk Factor Surveillance System (BRFSS), which conducts annual telephone surveys on the health status of U.S. residents. BRFSS collects data in all 50 states as well as the District of Columbia and three U.S. territories and completes more than 400,000 adult interviews each year, making it the largest continuously conducted health survey system worldwide.
The most recent dataset (as of February 2022) includes data from 2020. I downloaded a filtered version of it from Kaggle that contained 320,000 rows and 17 columns. Most columns are questions asked to respondents about their health status, such as "Do you have serious difficulty walking or climbing stairs?" or "Have you smoked at least 100 cigarettes in your entire life?".
Here is a look at the raw data. Notice that HeartDisease on the left is binary in the form of "Yes" or "No", while other variables like BMI and PhysicalHealth were numeric. AgeCategory is a good example of something that appears numeric but is actually an ordinal categorical variable. Most of the variables were of either numeric or character type. I had to convert most of them to factors, often binary or ordinal. I also had to do some re-coding, as in the case of recategorizing gestational diabetes as "No".
- Binary: df$Diabetic = as.factor(df$Diabetic)
- Ordinal: df$GenHealth = factor(df$GenHealth, levels = c("Poor", "Fair", "Good", "Very good", "Excellent"))
- Recoding: df$Diabetic[df$Diabetic == 'Yes (during pregnancy)'] <- "No"
The dataset was very skewed, with less than 10% of the sample having heart disease. This made it necessary to use more advanced sampling techniques to balance the dataset. The two most relevant options were oversampling, which randomly resamples with replacement from the underrepresented group, and the ROSE, or Random Over-Sampling Examples, a method that generates synthetic data to balance out the underrepresented group.
I created training and testing subsets from the data, 80% of the sample and 20% of the sample, respectively, and used them to determine that the oversampling approach yielded a better area under the ROC curve result, suggesting that it had higher overall predictive accuracy. As you can see, the two sampling approaches were ultimately quite similar in terms of their ROC curves, but there is even more smoothness (and therefore, area) under the oversampling one.
The AUC (area under the curve) for the ROC curve for the ROSE method was .778
The AUC for the ROC curve for the oversampling method was .789
Using the dataset generated with oversampling yielded a McFadden pseudo-R2 of .288, suggesting a good degree of model fit.
Here are the model coefficients. I highlighted the biggest risk factors in red and the biggest protective factors in green. A history of having a stroke and being above 50 are the biggest sources of increased risk for heart disease and being in very good or excellent health are the biggest sources of reduced risk of heart disease.
Here I show the confusion matrix for the oversampled data. You can see that the incorrect predictions from the model are in the tens of thousands, which illustrates how hard it is to make accurate predictions even from a large dataset. Using that matrix, I can provide standard metrics of model performance. The accuracy, or total correct predictions over total predictions, was .75. The precision, or proportion of positive identifications the model got correct, was only .22. So we see the model is essentially trigger-happy with positive identifications. The recall, or proportion of actual positives the model identified correctly, was 0.78.
Here are some key features I considered while designing the app:
- Allow users to enter responses to health queries
- Include a button to calculate after users make a selection
- Repeat button presses will perform new calculations with new menu selections
- Display caveats about accuracy and precision
Here’s a look at how the app turned out. Once you run the Preprocessing file in the GitHub repository, the use of the app is straightforward: you simply answer the health questions on the left and press the Calculate button when you are ready to view the result.
So what did we discover from this? Unfortunately, nothing too profound. The biggest risk factors identified by the model of stroke and age over 50 are unsurprising. The same goes for the biggest protective factors of reporting being in very good or excellent health. Perhaps more interesting is the finding that it appears difficult to surpass 75% accuracy in predictions with this approach. This highlights how much science still does not know about the specific causes of heart disease across genetic and lifestyle variations.
With more time, I would like to revise the output of the app to provide a confidence interval to give users a better idea of the range of risk associated with their specific health data. I would like to study machine learning and neural net techniques to make more sophisticated classifiers. I would also like to see if it is possible to get details about family history and diet and genetic markers to see if it is possible to enhance the predictions with richer data.