Pima Indians Diabetes Dataset

Classification

Minjae Chung (01715597)

1

Pima Indians Diabetes.ipynb - Colaboratory

Introduction to dataset

• 768 Subjects

2

Pima Indians Diabetes Database | Kaggle

Introduction to dataset

3

Introduction to dataset

4

Why is it worth to analyze?

• Association between diabetes and other health features

​

• Predict diabetes after health medical examination with AI

​

• Prevention and treatment

5

EDA (Exploratory Data Analysis)

Outlier Detection

• 17 Subjects with
• at least 2 outliers

​

6

EDA (Exploratory Data Analysis)

• Checking Null datatype

-> 0 Null Values

7

EDA (Exploratory Data Analysis)

• Summary and statistics

8

EDA (Exploratory Data Analysis)

• Correlation matrix between variables: Heatmap (1/2)

9

EDA (Exploratory Data Analysis)

• Correlation matrix between variables: Heatmap (2/2)

10

EDA (Exploratory Data Analysis)

1. Glucose: 0.47
2. BMI: 0.29
3. Age: 0.24
4. Pregnancies: 0.22
5. Pedigree 0.17
6. Insulin 0.13
7. Skin Thickness 0.07

7. Blood Pressure 0.07

11

EDA (Exploratory Data Analysis)

• Density curve of Glucose distribution
• Subject with diabetes show higher glucose concentration

12

EDA (Exploratory Data Analysis)

• Density plot of Age distribution
• Outcome 0: No diabetes
• Outcome 1: Diabetes
• Subject with diabetes tends to have higher age

​

​

​

13

EDA (Exploratory Data Analysis)

• Bar plot of Pregnancies
• Higher chance of having diabetes with higher number of pregnancy

​

14

EDA (Exploratory Data Analysis)

• Density curve of BMI distribution
• Subjects with diabetes showed more overweight and obese state
• Zero BMI values?

​

15

Healthy (18.5 to 24.9)

Overweight (25.0 to 29.9)

Obese (30.0 or higher)

​

​

​

(Centers for Disease Control and Prevention)

Mean BMI: 30 35

EDA (Exploratory Data Analysis)

16

• Zero values are null values for particular features

Feature Engineering

• 17 Subjects
• Outliers deleted
• 768 subjects

-> 751 subjects

​

17

Feature Engineering

• Delete 2 features with least correlation

​

• 1. Skin Thickness
• 2. Blood Pressure

18

Feature Engineering

• Converting zero values into null values
• Glucose, Insulin, BMI

19

Feature Engineering

• Replacing Nan to Mean value of the feature

20

Feature Engineering

• Data Scaling
• Equalizing the scale of features
• -> Improving the performance of models

​

• StandardScaler()
• = scales features to have zero mean and unit variance.

21

Feature Engineering

• Standardization

22

Artificial Neural Network 1

23

Input layer

Output layer

Hidden layer1

Hidden layer2

Hidden layer3

Neurons: 8 7 5 3 2

Artificial Neural Network 2

24

Input layer

Output layer

Hidden layer1

Hidden layer2

Hidden layer3

Hidden layer4

Hidden layer5

Neurons: 8 10 7 7 5 5 2

• More Neurons and hidden layer with same number of neurons of the previous layer are added

Artificial Neural Network 3

25

Input layer

Output layer

HL1

HL2

HL3

HL4

HL5

HL6

HL7

HL8

HL9

HL10

HL11

8 15 10 10 7 7 7 7 7 7 5 5 2

Neurons

• More Neurons and hidden layers added

Network Modelling / Optimizer = Adam

26

• ANN1 < ANN2 < ANN3

Model Evaluation (Original Data)

27

Confusion Matrix

28

Model Evaluation (Data modified)

29

Summary

• Dataset:

Pregnancies, Glucose, Blood Pressure, Skin Thickness

Insulin, BMI, Diabetes Pedigree, Age, Outcome

​

• Feature Engineering
• Deleting outliers
• Deleting features “Blood Pressure” & “Skin Thickness”
• Replacing null values with mean values
• Standard data scaling

​

• Classifiers

ANN, SVC, Random Forest, Extra Trees, Logistic Regression, Ensemble

​

Evaluation metric -> Recall

​

30

Conclusion

• ANN – Better evaluation score, when more neurons and hidden layers

​

• ANN had best accuracy, precision and f1-score and Logistic Regression model had best Recall among classifiers

​

• Data modification showed improvement on Recall except for Logistic Regression

​

​

31

Interpretability/ Explainable AI (XAI)

• Machine learning -> Black box

​

• Difficult interpretation

​

• How each features affects the prediction of the model

​

​

32

Interpretability/ Explainable AI (XAI)

• SHAP (SHapley Additive exPlanation)
• Black-Box Model : f
• Surrogate Model : g

33

SHAP value

• Force Plot
• Subject 0:

Positive SHAP value vs. Negative SHAP value

Glucose contributed the most

34

Prediction

SHAP value

35

• Force Plot
• Subject 300:

Positive SHAP value vs. Negative SHAP value

Glucose contributed the most

Glucose Dependence Plot

36

• High values of Glucose variable have a high positive contribution on the prediction

Age Dependence Plot

37

• High values of Age variable tends to have a high positive contribution on the prediction until age of 50.

SkinThickness Dependence Plot

38

• Value of Skin Thickness variable are showing no correlation with SHAP value

Summary Plot

39

• Order of global feature importance
• Glucose contributed the most and Skin Thickness contributed the least

Summary Plot

• 1. Glucose
• 2. BMI
• 3. Age
• 4. DiabetesPedigree
• 5. Pregnancies
• 6. Insulin
• 7. Blood Pressure
• 8. SkinThickness

40

Summary Plot

• 1. Glucose
• 2. BMI
• 3. Age
• 4. DiabetesPedigree
• 5. Pregnancies
• 6. Insulin
• 7. Blood Pressure
• 8. Skin Thickness

41

1. Glucose: 0.47
2. BMI: 0.29
3. Age: 0.24
4. Pregnancies: 0.22
5. DiabetesPedigree 0.17
6. Insulin 0.13
7. Skin Thickness 0.07

7. Blood Pressure 0.07

Thank You for listening

42

SVC classifier (Support Vector Machine)�

• Supervised Learning Algorithm
• Support Vector
• Hyperplane

​

• Create decision boundary that

segregate n-dimensional space into

classes

43

Support Vector Machine (SVM) Algorithm - Javatpoint

Random Forest Classifier

• Supervised Learning Algorithm
• A number of decision trees on various subsets of given dataset and takes
• Ensemble learning (majority voting)
• Bagging (Boot strap + aggregation)

44

Machine Learning Random Forest Algorithm - Javatpoint

Extra Trees Classifier

• Supervised Learning Algorithm

​

• Variety of Random Forest

​

• Subset of decision tree

​

• Majority voting

​

• No Boot strap sample = uses whole original sample

​

• Random split points (RF=optimum split points)

45

What is the difference between Extra Trees and Random Forest? | Quantdare

Logistic Regression Model

• Supervised Learning Algorithm
• Predicting the categorical dependent variable using a given set of independent variables.
• Probability between 0 ~ 1
• Classification problems
• Threshold

46

Logistic Regression in Machine Learning - Javatpoint

Ensemble Model

• Process of combining multiple classifiers to solve a complex problem and to improve the performance of the model.

47

Feature Engineering

• Histogram: Data distribution

48

Glucose

BMI

Insulin

Feature Engineering

• Histogram: Data distribution

49

Glucose

BMI

Insulin

BMI Dependence Plot

50

• High values of BMI variable have a high positive contribution on the prediction

Pregnancies Dependence Plot

51

• High values of Pregnancies variable tends to have a high positive contribution

Optimizer

• Gradient Descent – slow, local minimum

​

• Gradient Descent + momentum – faster, saddle point

​

• Adagrad (Adaptive Gradient algorithm) – slow, escape saddle point

​

• RMSProp (Root Mean Square Propagation) – faster, decaying factor

​

• Adam – Momentum + RMSProp

52