پیوندهای مفید
آمار
| تعداد نشریات | 21 |
| تعداد شمارهها | 1,294 |
| تعداد مقالات | 11,832 |
| تعداد مشاهده مقاله | 83,226,477 |
| تعداد دریافت فایل اصل مقاله | 121,984,085 |
Enhancing Clinical Reliability in Diabetes Mellitus Diagnosis: A Multi-Stage Deep Neural Network Framework | ||
| Health Management & Information Science | ||
| مقالات آماده انتشار، پذیرفته شده، انتشار آنلاین از تاریخ 07 شهریور 1405 | ||
| نوع مقاله: Original Article | ||
| شناسه دیجیتال (DOI): 10.30476/jhmi.2026.112718.1391 | ||
| نویسنده | ||
| Parisa Eslami* | ||
| چکیده | ||
| Background and Objective: Diabetes Mellitus is a critical global health challenge requiring early and reliable diagnostic frameworks. Although deep learning (DL) models show promise in medical decision support, clinical datasets often suffer from physiological noise and anomalies that degrade predictive performance. This study proposes a multi-stage pipeline combining unsupervised data refinement with deep neural networks to enhance diagnostic reliability on the PIMA Indian Diabetes Dataset. Methods: The framework comprises four phases: data acquisition, preprocessing, architectural modeling, and evaluation. To address data integrity, an unsupervised K-means clustering algorithm (k=3), evaluated via the Davies-Bouldin Index, was applied to standardized physiological attributes to identify and prune severe structural outliers (n=26). Subsequently, a feedforward Artificial Neural Network (ANN) and a Multi-Layer DL architecture incorporating ReLU activation and momentum optimization were evaluated. Validation was conducted using 10-fold Stratified Cross-Validation on both raw baseline data (N=768) and the refined cohort (N=742). Results: Unsupervised outlier mitigation significantly enhanced model stability and convergence. Before preprocessing, baseline ANN and DL models yielded accuracies of 65.11% and 76.17%, respectively. Post-outlier pruning, the proposed DL model achieved superior performance with an accuracy of 98.65%, sensitivity of 97.83%, specificity of 99.02%, and an AUC of 0.990, drastically outperforming the baseline architectures. Conclusion: Combining unsupervised outlier purification with multi-layer deep neural networks effectively mitigates clinical data noise and enhances model performance. This analytical pipeline provides a robust, scalable foundation for integration into Clinical Decision Support Systems (CDSS) to assist early diabetes screening and reduce the health and economic burdens of unmanaged diabetes risk. | ||
| کلیدواژهها | ||
| Diabetes Mellitus؛ Deep Learning؛ PIMA Indian Dataset؛ Outlier Detection؛ Artificial Neural Networks | ||
|
آمار تعداد مشاهده مقاله: 1 |
||