Analyzing and predicting diabetes with deep learning and visual insights

Diabetes is a prevalent chronic condition often accompanied by comorbidities such as obesity, hypertension, and arthritis, influenced by both individual behaviors and broader socioeconomic factors. This study aims to analyze diabetes trends and associated chronic conditions across U.S. states from 2011 to 2021, and to develop a predictive model to estimate diabetes prevalence using selected health, demographic, and socioeconomic indicators. We compiled a comprehensive dataset by integrating multiple U.S. public health sources, yielding 90 features representing chronic disease rates, demographics, and socioeconomic conditions. Statistical analysis and visualizations—including trend lines and geographic comparisons—were used to identify regional disparities and comorbidity patterns. For predictive modeling, we implemented a hybrid deep learning framework combining Principal Component Analysis (PCA) for feature selection with a Multi-Layer Perceptron (MLP) classifier. The model was evaluated using 5-fold cross-validation and standard performance metrics. Our visual analytics revealed significant state-level disparities in diabetes prevalence, closely associated with factors like obesity, tobacco use, and poverty. States such as Mississippi consistently exhibited higher rates of diabetes and related risk factors. The MLP-PCA model, using only the top 21 features, achieved strong predictive performance: 93.07% accuracy, 92.81% precision, 90.65% recall, and a 91.61% F1-score, with a rapid prediction time of 0.12 s. The top features included population demographics, obesity, COPD, arthritis, hypertension management, and poverty rate. By integrating data visualization and machine learning, this study offers a scalable and interpretable framework for monitoring diabetes at the population level. The implementation strategy enables early detection and supports targeted public health strategies by identifying high-risk regions and influential risk factors. Future work will incorporate time-series forecasting and additional environmental and behavioral variables to enhance prediction and inform long-term health planning. • Created a novel U.S. diabetes dataset integrating multiple public sources. • Analysed 10-year chronic disease trends using statistical and geospatial methods. • Identified regional disparities in diabetes and comorbid conditions across states. • Explored and discussed demographic, socioeconomic, and health factors influencing diabetes rates. • Developed efficient ML models using tailored feature selection strategies. • Reduced 89 features to top 10, 15, and 21 for accurate state-level predictions. • Evaluated models with 5-fold cross-validation and diverse performance metrics.