Skip to content

Analysis and Prediction of Diabetes Outcome with machine learning and deep learning approach

Notifications You must be signed in to change notification settings

aashay15/Diabetes-Outcome-Prediction

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

12 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

title author date output
Diabetes Relation Analysis
Aashay Sharma
24/07/2020
html_document
keep_md
true

In this Particular analysis we will try to find out relations between different variables given in the data set and we will try to fit a model to predict the diabetes outcome.

Reading the Data and Performing some Exploratory Analysis

Data is already clean and does not recquire some specific cleaning but we will just convert the Outcome Variable to a factor variable to perform some plots.

data <- "/Users/aashaysharma/Desktop/RStudio/diabetes/diabetes.csv"
diabetes <- read.csv(data)

head(diabetes)
##   Pregnancies Glucose BloodPressure SkinThickness Insulin  BMI
## 1           6     148            72            35       0 33.6
## 2           1      85            66            29       0 26.6
## 3           8     183            64             0       0 23.3
## 4           1      89            66            23      94 28.1
## 5           0     137            40            35     168 43.1
## 6           5     116            74             0       0 25.6
##   DiabetesPedigreeFunction Age Outcome
## 1                    0.627  50       1
## 2                    0.351  31       0
## 3                    0.672  32       1
## 4                    0.167  21       0
## 5                    2.288  33       1
## 6                    0.201  30       0

Now we have many variables but our outcome is discrete that is it is binary data 1 or 0.

Now I will perform aov analysis to get the variance distribution of the data so we can see that what variables account for what amount of variance.

variance_analysis <- aov(Outcome ~ . , data = diabetes)
summary(variance_analysis)
##                           Df Sum Sq Mean Sq F value   Pr(>F)    
## Pregnancies                1   8.59    8.59  53.638 6.16e-13 ***
## Glucose                    1  34.02   34.02 212.406  < 2e-16 ***
## BloodPressure              1   0.12    0.12   0.771 0.380213    
## SkinThickness              1   0.86    0.86   5.393 0.020481 *  
## Insulin                    1   0.26    0.26   1.594 0.207108    
## BMI                        1   6.78    6.78  42.331 1.40e-10 ***
## DiabetesPedigreeFunction   1   1.82    1.82  11.349 0.000793 ***
## Age                        1   0.46    0.46   2.865 0.090922 .  
## Residuals                759 121.57    0.16                     
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Now we can see that the variables which account for high variance and lowest p-Values are :

  • Pregnancies
  • Glucose
  • BMI
  • DiabetesPedigreeFunction
  • and SkinThickness but it accounts for a bit higher p-Value compared to others in this list.

Converting the outcome to factor variable :

diabetes$Outcome2 <- as.factor(diabetes$Outcome)

Some Exploratory analysis :

a <-ggplot(data = diabetes, aes(x = Outcome2, y = Pregnancies)) + geom_boxplot()
b <-ggplot(data = diabetes, aes(x = Outcome2, y = Glucose)) + geom_boxplot()
c <-ggplot(data = diabetes, aes(x = Outcome2, y = BMI)) + geom_boxplot()
d <-ggplot(data = diabetes, aes(x = Outcome2, y = DiabetesPedigreeFunction)) + geom_boxplot()
e <-ggplot(data = diabetes, aes(x = Outcome2, y = SkinThickness)) + geom_boxplot()
grid.arrange(a, b, c, d, e, nrow = 3, ncol = 2)

Inference from the GRAPHS

Okay so after looking at the graphs we can infer that Pregnancies and Glucose have a significant mean difference with least outliers, BMI and PedigreeFunction have a lesser mean difference but have many outliers which can account for lesser accurate fit and the last graph SkinThickness has a lesser mean difference but one outlier.

So we can fit 3 Different models :
  • Pregnancies and Glucose as features
  • BMI and PedigreeFunction as feature along with the first 2
  • and all the feautres including SkinThickness.

We will use caret package for fitting and plotting.

Models :

First we will separate the data into a training and testing set.

set.seed(1234)
inTrain <- createDataPartition(y = diabetes$Outcome2, list = FALSE, p = 0.65)
train <- diabetes[inTrain,]
test <- diabetes[-inTrain,]

Model 1 ( Features : Pregnancies + Glucose)

Random Forest :
RF_model <- train(Outcome2 ~ Pregnancies + Glucose, method = "rf", data = train, ntree = 100)
## note: only 1 unique complexity parameters in default grid. Truncating the grid to 1 .
RF_predict <- predict(RF_model, test)
confusionMatrix(test$Outcome2, RF_predict)
## Confusion Matrix and Statistics
## 
##           Reference
## Prediction   0   1
##          0 139  36
##          1  36  57
##                                          
##                Accuracy : 0.7313         
##                  95% CI : (0.674, 0.7835)
##     No Information Rate : 0.653          
##     P-Value [Acc > NIR] : 0.003716       
##                                          
##                   Kappa : 0.4072         
##                                          
##  Mcnemar's Test P-Value : 1.000000       
##                                          
##             Sensitivity : 0.7943         
##             Specificity : 0.6129         
##          Pos Pred Value : 0.7943         
##          Neg Pred Value : 0.6129         
##              Prevalence : 0.6530         
##          Detection Rate : 0.5187         
##    Detection Prevalence : 0.6530         
##       Balanced Accuracy : 0.7036         
##                                          
##        'Positive' Class : 0              
## 
Logistic Regression:
LR_model <- train(Outcome2 ~ Pregnancies + Glucose, method = "glm", family = "binomial", data = train)
LR_predict <- predict(LR_model, test)
confusionMatrix(test$Outcome2, LR_predict)
## Confusion Matrix and Statistics
## 
##           Reference
## Prediction   0   1
##          0 155  20
##          1  40  53
##                                           
##                Accuracy : 0.7761          
##                  95% CI : (0.7214, 0.8246)
##     No Information Rate : 0.7276          
##     P-Value [Acc > NIR] : 0.04112         
##                                           
##                   Kappa : 0.4798          
##                                           
##  Mcnemar's Test P-Value : 0.01417         
##                                           
##             Sensitivity : 0.7949          
##             Specificity : 0.7260          
##          Pos Pred Value : 0.8857          
##          Neg Pred Value : 0.5699          
##              Prevalence : 0.7276          
##          Detection Rate : 0.5784          
##    Detection Prevalence : 0.6530          
##       Balanced Accuracy : 0.7604          
##                                           
##        'Positive' Class : 0               
## 

Model 2 ( Features : Pregnancies + Glucose + BMI + DiabetesPedigreeFunction)

Random Forest :
RF_model2 <- train(Outcome2 ~ Pregnancies + Glucose + BMI + DiabetesPedigreeFunction, method = "rf", data = train, ntree = 100)
RF_predict2 <- predict(RF_model2, test)
confusionMatrix(test$Outcome2, RF_predict2)
## Confusion Matrix and Statistics
## 
##           Reference
## Prediction   0   1
##          0 150  25
##          1  32  61
##                                           
##                Accuracy : 0.7873          
##                  95% CI : (0.7334, 0.8347)
##     No Information Rate : 0.6791          
##     P-Value [Acc > NIR] : 5.73e-05        
##                                           
##                   Kappa : 0.5223          
##                                           
##  Mcnemar's Test P-Value : 0.4268          
##                                           
##             Sensitivity : 0.8242          
##             Specificity : 0.7093          
##          Pos Pred Value : 0.8571          
##          Neg Pred Value : 0.6559          
##              Prevalence : 0.6791          
##          Detection Rate : 0.5597          
##    Detection Prevalence : 0.6530          
##       Balanced Accuracy : 0.7667          
##                                           
##        'Positive' Class : 0               
## 
Logistic Regression:
LR_model2 <- train(Outcome2 ~ Pregnancies + Glucose + BMI + DiabetesPedigreeFunction, method = "glm", family = "binomial", data = train)
LR_predict2 <- predict(LR_model2, test)
confusionMatrix(test$Outcome2, LR_predict2)
## Confusion Matrix and Statistics
## 
##           Reference
## Prediction   0   1
##          0 154  21
##          1  37  56
##                                           
##                Accuracy : 0.7836          
##                  95% CI : (0.7294, 0.8314)
##     No Information Rate : 0.7127          
##     P-Value [Acc > NIR] : 0.005295        
##                                           
##                   Kappa : 0.5024          
##                                           
##  Mcnemar's Test P-Value : 0.048885        
##                                           
##             Sensitivity : 0.8063          
##             Specificity : 0.7273          
##          Pos Pred Value : 0.8800          
##          Neg Pred Value : 0.6022          
##              Prevalence : 0.7127          
##          Detection Rate : 0.5746          
##    Detection Prevalence : 0.6530          
##       Balanced Accuracy : 0.7668          
##                                           
##        'Positive' Class : 0               
## 

Model 3 ( Features : Pregnancies + Glucose + BMI + DiabetesPedigreeFunction + SkinThickness)

Random Forest :
RF_model3 <- train(Outcome2 ~ Pregnancies + Glucose + BMI + DiabetesPedigreeFunction + SkinThickness, method = "rf", data = train, ntree = 100)
RF_predict3 <- predict(RF_model3, test)
confusionMatrix(test$Outcome2, RF_predict3)
## Confusion Matrix and Statistics
## 
##           Reference
## Prediction   0   1
##          0 149  26
##          1  36  57
##                                           
##                Accuracy : 0.7687          
##                  95% CI : (0.7135, 0.8178)
##     No Information Rate : 0.6903          
##     P-Value [Acc > NIR] : 0.002792        
##                                           
##                   Kappa : 0.4763          
##                                           
##  Mcnemar's Test P-Value : 0.253038        
##                                           
##             Sensitivity : 0.8054          
##             Specificity : 0.6867          
##          Pos Pred Value : 0.8514          
##          Neg Pred Value : 0.6129          
##              Prevalence : 0.6903          
##          Detection Rate : 0.5560          
##    Detection Prevalence : 0.6530          
##       Balanced Accuracy : 0.7461          
##                                           
##        'Positive' Class : 0               
## 
Logistic Regression:
LR_model3 <- train(Outcome2 ~ Pregnancies + Glucose + BMI + DiabetesPedigreeFunction + SkinThickness, method = "glm", family = "binomial", data = train)
LR_predict3 <- predict(LR_model3, test)
confusionMatrix(test$Outcome2, LR_predict3)
## Confusion Matrix and Statistics
## 
##           Reference
## Prediction   0   1
##          0 155  20
##          1  37  56
##                                           
##                Accuracy : 0.7873          
##                  95% CI : (0.7334, 0.8347)
##     No Information Rate : 0.7164          
##     P-Value [Acc > NIR] : 0.00512         
##                                           
##                   Kappa : 0.5097          
##                                           
##  Mcnemar's Test P-Value : 0.03407         
##                                           
##             Sensitivity : 0.8073          
##             Specificity : 0.7368          
##          Pos Pred Value : 0.8857          
##          Neg Pred Value : 0.6022          
##              Prevalence : 0.7164          
##          Detection Rate : 0.5784          
##    Detection Prevalence : 0.6530          
##       Balanced Accuracy : 0.7721          
##                                           
##        'Positive' Class : 0               
## 

Printing the model Accuracies :

1 -> Pregnancies 2 -> Glucose 3 -> BMI 4 -> DiabetesPedigreeFunction 5 -> SkinThickness

Random Forest

print("RF Model 1 + 2")
## [1] "RF Model 1 + 2"
confusionMatrix(test$Outcome2, RF_predict)$overall[1]
##  Accuracy 
## 0.7313433
print("RF Model 1 + 2 + 3 + 4")
## [1] "RF Model 1 + 2 + 3 + 4"
confusionMatrix(test$Outcome2, RF_predict2)$overall[1]
##  Accuracy 
## 0.7873134
print("RF Model 1 + 2 + 3 + 4 + 5")
## [1] "RF Model 1 + 2 + 3 + 4 + 5"
confusionMatrix(test$Outcome2, RF_predict3)$overall[1]
##  Accuracy 
## 0.7686567
Logistic Regression
print("LR Model 1 + 2")
## [1] "LR Model 1 + 2"
confusionMatrix(test$Outcome2, LR_predict)$overall[1]
##  Accuracy 
## 0.7761194
print("LR Model 1 + 2 + 3 + 4")
## [1] "LR Model 1 + 2 + 3 + 4"
confusionMatrix(test$Outcome2, LR_predict2)$overall[1]
##  Accuracy 
## 0.7835821
print("LR Model 1 + 2 + 3 + 4 + 5")
## [1] "LR Model 1 + 2 + 3 + 4 + 5"
confusionMatrix(test$Outcome2, LR_predict3)$overall[1]
##  Accuracy 
## 0.7873134

Final Conclusion

We can see the accuracies and they are a bit close, but there are some outliers in BMI and PedigreeFunction Variables thus the fit can be faulty in some manner but over all we can see all this 5 variables have some significant effect over diabetes outcome.

Further we can try more combinations and other models to see what fits the best, we cannot just rely on accuracies

The Machine Learning Approach

After some exploratory analysis with the selected attributes from the dataset, I applied some common ML algoritms (widely used for classification problems), though the models perform well and can be used for inference and base level predictions but a deep learning approach is neccesary as it can help us infere more about the data and give much better predicitions (more real world predictions)

The Deep Learning Approach

The dataset which is used in this problem contains only 768 instances (ie, rows or records), so it is sufficient for a machine learning problem but is very small for a deep learning problems as the data will be splitted further into test and train sets and for correct evalutation of the model we would be left with a very small test set which wont give us good or highly significant metrics for our model. So eradicate this problem to a acceptable level we can use K Folds Cross validation method to train and evalusate our model so that even with such a small data we can get pretty good and most important "significant" metrics for our model.

The choice of model is also very simple due to data limitation and so we will use a model with only 2 layers, the model schema is summarized below:

INPUT LAYER : 8 UNITS SHAPE(5,), ACTIVATION = "RELU"
#1 HIDDEN LAYER 1 : 8 UNITS, ACTIVATION = "RELU"
#2 OUTPUT LAYER 2 : 1 UNIT , ACTIVATION = "SIGMOID" (FOR BINARY CLASSIFICATION PROBLEM)

LOSS FUNCTION : BINARY_CROSSENTROPY
OPTIMIZER : STOCHASTIC GRADIENT DESCENT (SGD)
METRICS : ACCURACY (HOW MANY LABELS DID WE CORRECTLY PREDICT FROM TEST SET)

VALIDATION METHOD : K FOLDS CROSS VALIDATION WITH 3 FOLDS
NOTE : Also tweaked and tried 10 folds as well as 5 folds but the final metrics did not have any significant or drastics changes so chose 3 Folds

Results:

About

Analysis and Prediction of Diabetes Outcome with machine learning and deep learning approach

Topics

Resources

Stars

Watchers

Forks