Web-Based Early Dementia Detection Using Deep Learning, Ensemble Machine Learning, and Model Explainability Through LIME and SHAP

Abstract

Dementia is a gradual and incapacitating illness that impairs cognitive abilities and causes memory loss, disorientation, and challenges with daily tasks. Treatment of the disease and better patient outcomes depend on early identification of dementia. In this paper, the study uses a publicly available dataset to develop a comprehensive ensemble model of machine learning (ML) and deep learning (DL) framework for classifying the dementia stages. Before using SMOTE to balance the data, the procedure starts with data preprocessing which includes handling missing values, normalization and encoding. F-value and p-value help to select the best seven features, and the dataset is divided into training (70%) and testing (30%) portions. In addition, four DL models like long short-term memory (LSTM), convolutional neural networks (CNNs), multilayer perceptron (MLP), artificial neural networks (ANNs), and 12 ML models are trained such as logistic regression (LR), random forest (RF) and support vector machine (SVM). Hyperparameter tuning was utilized to further enhance each model’s performance and an ensemble voting technique was applied to aggregate predictions from several ML and DL algorithms, providing more reliable and accurate outcomes. For ensuring model transparency, interpretability strategies like as shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME) are applied in ANN and LR. The suggested model’s ANN shows a promising accuracy of 97.32% demonstrating its efficacy in the early diagnosis and categorization of dementia which can support clinical decisions. Furthermore, the proposed work, created a web-based solution for diagnosing dementia in real-time.