Early Diabetes Prediction based on patients' symptoms

Logistic Regression, Decision Tree, & Random Forest Classification Algorithms

• This project aims to train Machine Learning classification models to predict whether a patient is at risk of developing diabetes.
• Key indicators (symptoms) correlated to diabetes will be identified.
• This project will be useful to clinicians, as it will help them understand how to better avoid missing the False Positive cases who may go undiagnosed.

Data Souce:

• Collected data from UCI repository - https://archive.ics.uci.edu/dataset/529/early+stage+diabetes+risk+prediction+dataset

 

PART 1: Data Collection & Cleaning

PART 2: Analysis of data using Vizualization and Statistics

Questions to answer:

1. Is obesity related to diabetes status? or is it independent

   <p>Chi2ContingencyResult(statistic=2.3274739583333344, pvalue=0.12710799319896815, dof=1, expected_freq=array([[166.15384615,  33.84615385],
           [265.84615385,  54.15384615]]))<p>

• The P-Value of 0.127 shows there is no significant evidence that there is a relationship between obsesity status and diabetes.
• The expected freq array shows the values if everything was independent. Any deviation from the independent cross tab will be deemed dependent.
• Therefore array shown is similar to the chi2 cross tab numbers, and can say that obsesity and class are independent.

 

2. Is gender related to diabetes status?

   Chi2ContingencyResult(statistic=103.03685927972558, pvalue=3.289703730553317e-24, dof=1, expected_freq=array([[126.15384615,  73.84615385],
   [201.84615385, 118.15384615]]))

• The p-value is 3.29e-24, meaning that there is significant evidence that there is a relationship between a postitive diabetes status and gender.
• The Independent array values are significantly differnt from the cross tab values therefore shows that obsesity status and gender are dependednt (there is a relationship)
• There is significantly more females with diabetes compared to those without diabestes. (173)
• There is significantly more males with out diabetes (181) than those with diabetes.

 

3. Is polyuria related to obesity status?

   Chi2ContingencyResult(statistic=227.86583895496773, pvalue=1.7409117803442155e-51, dof=1, expected_freq=array([[100.76923077,  99.23076923],
   [161.23076923, 158.76923077]]))

• The P-Value of 1.74e-51 shows that there is strong significant evidence that there is a relationship between Polyuria status and diabetes.

 

4. Is gender related to polyuria status?

          Chi2ContingencyResult(statistic=36.49184228561174, pvalue=1.5330652930649977e-09, dof=1, expected_freq=array([[165.26153846,  96.73846154],
          [162.73846154,  95.26153846]]))

• P-Value of 1.53e-09 means that there is a strong evidence that there is a relationship between being a female and having a positive polyuria status.

 

5. Is there a relationship between age and diabetic status?

• Box plots don't seem to show a big difference between the means and medians of patients with and without diabetes, as there is a lot of overlap.

 

Correlation heatmap

For gender, if negative, it is female, if positive, male

1. Diabetes is positively correlated (0.449) with females, therefore females are 44.9% more likely to have diabetes than males.
2. Patients with polyuria are 66.6% more likely to have diabetes than those who do not have polyuria.
3. Patients with polydipsia are 64.9% more likely to have diabetes than those who do not have polydipsia.
4. Patients who have sudden weightloss are 43.7% more likely to have diabetes.
5. Patients with partial paresis are 43.2% more likely to have diabetes.
6. Female patients are 32.8% less likely to have alopecia than male patients.

 

Top 5 important features to look up for obsesity based on the decision tree model are:

1. Polydipsia
2. gender (Female)
3. Polyuria
4. age
5. alopecia

feature importance 3 polydipsia 0.424208 1 isfemale 0.131796 2 polyuria 0.115480 14 alopecia 0.074900 0 age 0.066944 9 itching 0.049823 11 delayed healing 0.042698 6 polyphagia 0.033854 10 irritability 0.022087 7 genital thrush 0.016927 13 muscle stiffness 0.010640 8 visual blurring 0.006771 12 partial paresis 0.001938 4 sudden weight loss 0.001935 5 weakness 0.000000 15 obesity 0.000000

 

DummyClassifier model to establish a baseline

• 26 False positive patients - patients do not have diabetes but are predicted to have - An issue because they would have to go through additional tests and anxiety.
• 24 False negative patients - patients have diabetes but were predicted not to have it- An issue because the disease is left undiagnozed and dangerous for patient's health.
• Accuracy of the Dummy classification model is is 52%, which is not high enough and therefore can try more models.

 

Logistic regression model

• lower False positives (5) and False Negatives(5) in the logistic model

• Accuracy score has gone up from 52% in dummy model to 90% in logistic model

• Precision, recall and f1-scores have all increased as well

• This means that the logistic model a better model to predict diabetes than the dummy classifier model

 

Decision Tree model

• The Decision Tree model identified 4 False Positive and Zero False Negative Values
• The Accuracy Score of the Decision Tree model is the highest at 95% compared to logistic regression model at 90%

 

Random Forest Model

• The Random Forest model identified 1 False Negative and 4 False Positive cases.
• The Random Forest Classification model has an accuracy of 95%
• This is slighly lower compared to the Decision tree regression model accuracy of 96%

The overall best model in this analysis is the Decision tree at 95% prediction accuracy.

 

Reference:

• Early stage diabetes risk prediction dataset.. (2020). UCI Machine Learning Repository. https://doi.org/10.24432/C5VG8H.