Diabetes Classification

Tools:

• Python
• Pandas
• Numpy
• Scikit-Learn
• Seaborn
• Matplotlib

Installation requirements declared in pyproject.toml file.

Processes:

• The exploratory data analysis notebook documents the process of analyzing the data, exploring variable correlations and the cleaning processes that lead to data processing functions in the process_data.py file.

Models:

• K Nearest Neighbor
• Support Vector Classifer
• Decision Tree Classifier
• Random Forest Classifier
• Logistic Regression Classifier

Motivations:

Diabetes is a chronic disease that interferes with the body's ability to digest sugar, causing sugar levels to build up in the blood stream. When left untreated, a build up of sugar in the blood stream can cause blockage in the arteries, block bloodflow to the extremities, and eyes. Further complications from the blockages may cause heart disease/heart attacks, loss of extremities and/or vision loss.

There are three different types of diabetes:

Type 1: Also known as Juvenile diabetes, is chronic and occurs to younger patients due to the pancreas not being able to produce any or enough insulin. Type 2: Happens to predominantly adult patients, onset from lifestyle habits such as poor diet, or sedentary lifestyle. In type 2, the body makes less insulin or becomes resistant to insulin. Gestational: Occurs when women aren't able to produce enough insulin during pregnancy. To prevent complications from diabetes, patients must monitor sugar levels in the blood through bi/annual or annual A1C tests and daily pricks, adapt better diet and lifestyle habits to manage weight, and use the hormone insulin to digest sugar if their bodies stop producing it or don't produce enough.

Why is this a problem?
Currently, diabetes affects 37 million people a year, and costs the US $327 million dollars in healthcare costs, and loss of time costs. In addition, Diabetes is the 7th leading cause of death in the United States and it could become an even bigger problem as obesity rates increase. Diabetes can be difficult to manage but it is not impossible.

How can this project help? This project uses healthcare data to train models to classify diabetes outcome in patients based on other health outcomes. In many healthcare settings, existing patient data can be used to help prioritize at-risk patients with the preventative care that they need. For example, patients that the model identifies as diabetic can be given priority lab appointments to confirm their A1Cs, or have priority for nutritional counceling appointments.

Data:

The dataset for this project was obtained from Kaggle. It was sourced from the National Institute of Diabetes and Digestive and Kidney Diseases, and it consists of female patients that are 21 years or older and of Pima Indian descent.

This dataset is interesting because it captures diabetes and health outcomes of minority women of color, a group that has a long and complicated history with modern healthcare practices, from being victims of unethical practices in the United States, as was the case with the birth control trials on unsuspecting Puerto Rican women in the 1930s , to lacking the resources to access modern healthcare [4]. Healthcare needs of women are different from men, and as the FemTech industry continues to grow, it is not only important to research how to improve womens' health outcomes, but it is also more profitable since women have higher needs of care [5], [6], [7]. Furthermore, women's health is one of the biggest determinants of population health, and currently infant mortality rate is one of the strongest measures of a country's well being [8]. Still, infant mortality rates in the United States are the worst among minority women of color [9].

While this project focuses on identifying diabetes among women to improve health outcomes and administer early intervention efforts, it is also a commentary on the healthcare system's need to gain the trust of women of color, and adapt to the needs of minority women.

The data columns are: Pregnancy Glucose Blood Pressure SkinThickness Insulin BMI DiabetesPedigreeFunction Age Outcome We'll be using the other columns to predict feature 9: Diabetes diagnoses (0, 1) where 0 = No, 1 = Yes

Conclusion:

The best performing model that we experimented with in this project was KNN with a k value set to 18. The classification report:

Our initial goal was to create a model that performed well at classifying positive diabetes cases. As we can see, all of our models performed better at detecting patients who did not have diabetes (those in class 0).

If we revisit the distribution of our two classes, we can see that there were a lot more rows consisting of patients who did not have diabetes vs those who did, meaning that our model had better data/information on patients who did not have diabetes.

Improvements:

The biggest shortcoming for this project is its data simplicity. In order to research diabetes and build a good classification model, we need a lot of robust data sourced from a reliable source, for my research purposes ideally representing minority women patients from the United States. Therefore, further work in this project would start with allocating better quality and higher quantity data.

Run project:

cd <directory>
git clone <repository link>
poetry install
jupyter notebooks