Multi-label risk diabetes complication prediction model using deep neural network with multi-channel weighted dropout

• Main Author: Dzakiyullah, Nur Rachman
• Format: Thesis
• Language: English; English
• Published: 2025
• Subjects: Q Science; QA Mathematics
• Online Access: http://eprints.utem.edu.my/id/eprint/29015/

The early diagnosis of diabetes complications using risk factors remains underexplored, particularly with the application of Multi-Label Classification (MLC). This study addresses this gap by leveraging data from the Behavioral Risk Factor Surveillance System (BRFSS)from 2016 to 2021 to categorize seven diabetes complications simultaneously. By employing Artificial Intelligence (AI), this study examines the interconnected nature of these complications. A total of 33 variables from each year of the BRFSS dataset were analyzed, incorporating statistical techniques to understand the data and preprocessing methods to prepare it for machine learning. Seven machine learning models—Artificial Neural Network (ANN), Random Forest (RF), Decision Tree (DTT), k-Nearest Neighbors (k-NN), Naïve Bayes (NB), Support Vector Machine (SVM), and Deep Neural Network (DNN)—were used for multi-label classification of the complications. The study employed two MLC frameworks: Problem Transformation methods (Binary Relevance, Classifier Chains, Label Power Set, and Calibrated Label Ranking) and Algorithm Adaptation. Performance was evaluated using 20 metrics, including AUROC and other advanced indicators. The first experiment revealed that the Algorithm Adaptation framework outperformed Problem Transformation methods across most metrics. Among the models, the DNN achieved superior performance in key metrics such as Subset Accuracy (0.4156), Hamming Loss (0.1272), F1-Score (macro) (0.9113), and AUROC (macro) (0.7935). Feature importance analysis identified the top 10 variables influencing different complications. The second experiment introduced a novel dropout regularization technique called multi-channel weighted dropout, designed to enhance model generalization. Comparative evaluations with existing dropout methods demonstrated the superior performance of the proposed technique, particularly when applied within the Algorithm Adaptation framework using DNNs. The proposed method managed data and model complexity effectively while maintaining high computational efficiency. This study contributes to the field by proposing a new regularization technique, demonstrating the effectiveness of the Algorithm Adaptation framework, and providing insights into the associations between diabetes complications. These findings highlight the potential of AI-driven MLC approaches in advancing diabetes complication.