capstone-presentation for diabetes prediction leveraging machine learning.

This project helps us to predict diabetes leveraging different machine learning algorithms that can predict whether an individual is likely to have diabetes or not based on their lifestyle related attributes and historical data.

Problem Statement :

The dataset used in this project is originally from NIDDK(National Institute of Diabetes and Digestive and Kidney Diseases).Predicting diabetes in patients based on diagnostic measurements and thereby, evaluating the predictive performance of different machine learning algorithms in predicting the outcomes and identifying trends for diabetes management and early personalized treatment , based on the following records of the patients : Pregnancies (Number of times a female has conceived) Glucose (Blood Sugar Level)-mMol/L Blood Pressure (Diastolic Blood Pressure)-mm Hg Skin Thickness (Fold Thickness Of Triceps Skin)-mm Insulin (2 hours Serum Insulin Intake)-mL Body Mass Index (Body Mass / Body Weight)- kg Diabetes Pedigree Function (Hereditary Carrier Cases)/(Genetic Influence) Age (Age Of Females)- years So, these are all the independent variables, on the basis of which, I would utilize the different machine learning algorithms to predict the target variable , thereby, to solve this disease prediction problem under health care domain.

WHY AM I SOLVING THE PROBLEM?

For 3 foremost reasons I am performing this problem of diabetes prediction. EARLY IDENTIFICATION OF THE DISEASE- acts as a boon to save a large count of lives before they actually develop this disease. PLANNING THE PERSONALIZED TREATMENTS- Our predictive models actually helps the healthcare professionals to plan a proper personalized treatments for the patients who are suffering or at the likelihood of developing the disease. ALLOCATION OF RESOURCES- Looking at the likelihood of developing this disease in a person, the healthcare professionals tend to take some measures of allocation like , educational campaign, preventive measures, in a way that a patient gets benefitted by that.

CHALLENGES FACED BY THE HEALTH CARE DOMAIN TO PREDICT THE DISEASE OF DIABETES:

Data Availability Data Quality And Standardization Data Privacy And Security Feature Selection And Dimensionality Imbalanced Data Model Interpretability

HOW MY MODEL IS GOING TO HELP:

Accurate Predictions Early Prevention Improved Patient Outcomes Research And Insights Continuous Learning And Improvement Personalized approach

1. Total count and their percentages of diabetic and non-diabetic cases.

=>> OVERVIEW :

Here, getting the total count and percentages of who had diabetes and who did not have diabetes............so, as we can see

that, count of non-diabetic women is the highest.

2. distribution of blood pressure by outcome.

=>> OVERVIEW :

From the above visualisation I can say, that the distribution of Blood Pressure of a person both with diabetes

is slightly left skewed and whereas, without diabetes is almost normally distributed.

3. distribution of pregnancies by outcome.

=>> OVERVIEW :

From the above plot , I can easily interpret that the distribution of Pregnancies in female is right

skewed , both with the diabetic and non-diabetic conditions, that says that , the mean is greater

than the median.

4. frequency distribution of each features

=>> OVERVIEW :

From the above plots I came to know regarding the distribution of various features in the data, that is, the

distribution of some of the features are right skewed, which tells us that the mean is greater than the

median,such as ,'Pregnancies','Skin Thickness','Insulin','Diabetes Pedigree Function' ,'BMI' and 'Age',these

features are right skewed. On the otherhand, the feature, 'Blood Pressure' is slightly left skewed,

that is, median is greater than mean.

5. correlation plot

=>> OVERVIEW :

From this viasualisation, we can see that :

** If it is nearer to 1, then it says that the two variables are perfectly correlated to each other.

** If, it is nearer to -1, then it means the two variables are poorly correlated to eachother.

As, we can see from the table and heatmap, the pairs which are correlated are:

** Age and Pregnancies

** Skinthickness and BMI

** Glucose and Insulin

6. plot of insulin vs outcome

=>> OVERVIEW :

Hereby, it is pretty understandable that the maximum people with non-diabetic condition is having a Insulin

level in the range of (0-200)mL, as its showing that its wider as compared to that of those who are diabetic

(less people). However, we can also see that the median value of Insulin level is more in diabetic females as

compared to that in non-diabetic females, also along with the maximum number of outliers in this category.

7. boxplot of the dataset drawing various statistical inferences.

=>> OVERVIEW :

From the above visualization, I would interpret that all the features(independent variables) are containing

outliers , we need to remove it in order to get a good outcome from this dataset.

lastly, we can conclude that, after training the models and evaluating their performances we come with the conclusion that the random forest is the best performing model

inorder to solve the diabetes prediction with the highest recall value 0.60 among all the four models employed by us, thereby fetching its confusion matrix.

Hence, displaying the feature importances plot

Top 5 features, playing the most significant role in influencing the diabetes prediction

. Glucose

. Body mass index

. Age

. Diabetes pedigree function

. Pregnancies

HOW THIS MODEL CAN BE LEVERAGED IN OTHER DOMAINS OR CAN BE SCALED WITH SOME MORE TWEAKS ?

This model can be leveraged in other domains such as : FINANCE : Random Forests can be applied in finance for tasks like fraud detection, investment prediction, by training the model on the basis of the historical financial data. So, that it can assess credit worthiness, detect anomalous transactions. RETAIL AND E-COMMERCE : Random Forests can be utilized here to predict customer behavior , recommendation of the products on the basis of customer attributes, their purchase history and hence, makes easier for the business to take the data driven decisions. Now, it can be scaled with some more tweaks such as: Feature engineering and domain specific knowledge Additional features makes the task more easier for the model to predict the target variable. Model ensembling can help in improving the overall prediction by leveraging the strength of different algorithms.

In this way I have solved the diabetes prediction problem.