Design of an Improved Model for Diabetes Detection Combining Deep Dyna-Q Learning with Ensemble Classification

Abstract

The burgeoning prevalence of diabetes worldwide necessitates advancements in early detection techniques to curb its impacts on public health. Traditional methods, while effective to a degree, often suffer from limitations such as low detection speed, moderate accuracy, and high false positive rates, which could delay timely intervention and management of the disease. This research introduces an innovative approach combining deep reinforcement learning and ensemble classification to enhance the efficiency of diabetes detection. In addressing the limitations of existing detection systems, our study integrates a Deep Dyna-Q Learning Network with an ensemble of classical classifiers, namely Naive Bayes, Multilayer Perceptron (MLP), Support Vector Machine (SVM), Logistic Regression (LR), and k-Nearest Neighbors (kNN). The Deep Dyna-Q Network is meticulously designed with multiple layers that include dense units with ReLU activation, batch normalization, and dropout for regularization, specifically configured to optimize learning from complex, non-linear medical data samples. The ensemble approach leverages the distinct statistical strengths of each classifier, allowing for a more robust and comprehensive analysis of clinical data samples. This ensemble is tailored to improve the generalizability and reliability of predictions by mitigating individual biases inherent in single-model predictions. Our model processes a wide array of clinical parameters such as Glucose levels, Blood Pressure, Body Mass Index (BMI), and lifestyle factors like Smoking and Physical Activity, which are critical for predicting diabetes onset. The use of such diverse data helps in capturing a holistic view of risk factors associated with diabetes, thereby enhancing the model's diagnostic precision. The impact of this hybrid model is significant, demonstrating an enhancement in precision by 8.5%, accuracy by 5.9%, recall by 8.3%, and a reduction in diagnosis delay by 2.9%. Furthermore, the Area Under the Curve (AUC), a key metric for classification performance, improved by 9.5%. These improvements underscore the potential of integrating deep learning techniques with traditional statistical methods to create a more accurate, efficient, and responsive diabetes detection system. This work not only advances the technological framework for disease detection but also offers a scalable model that can be adapted to various healthcare settings, potentially leading to better patient outcomes and reduced healthcare costs through timely and accurate diabetes detection.