/ensemble-dementia-predictor

An application that allows users to predict the risk of a patient having dementia based from their MRI Scan and other medical data using Three Ensemble machine learning methods

Primary LanguageJupyter Notebook

Prediction of Dementia using Three Ensemble Machine Learning Methods: Project Overview

An application that allows users to predict the risk of a patient having dementia based from their MRI Scan and other medical data

  • Three types of Ensemble methods were performed to model the data .i.e Bagging, Stacking and Boosting.

    • Bagging using Random Forest
    • Stacking using Decision Trees, Naive Bayes and K-Nearest Neighbors
    • Boosting using XGBoost
  • Performed various Data Preprocessing techniques such as missing data imputation and removal of multicolinear features to clean and make the data ready for model building

  • Tuned hyperparameters of the model to achieve best performance.

  • Boosting had an accuracy of 86.76%, f1-score of 83.64% and recall of 85.19%

  • Model was deployed on a web application built using Django available at Dementia Predictor


Model Performamce

Accuracy, F1-Score and Recall were the metrics used to evaluate the performance of the model

Method Accuracy (%) F1-Score (%) Recall (%)
Bagging 85.29 82.14 85.19
Stacking 85.29 80.77 77.78
Boosting 86.76 83.64 85.19

Confusion Matrix

0 1
0 TN FP
1 FN TP

Web application of the model


Features

Variable Data Object Data type
MR Delay The number of days between visits by a patient. Integer
Gender Gender of a patient (M or F) Object
Hand Patient’s dominant hand Object
Age Patient's age at the time of data collection Integer
EDUC Years of Education Integer
SES Socioeconomic status is classified into categories from 1 (highest status) to 5 (lowest status) Integer
MMSE Mini-mental State Examination score (range is from 0-worst to 30-best) Integer
eTIV Estimated total intracranial volume (cm3) Integer
nWBV Normalized whole brain volume Float

Model Deployment

The final model with the best score was deployed on a web application built with Django with the frontend built with HTML & CSS with Boostrap 4 as the CSS Framework.

Web application of the model