Comparison of Synthetic Performance with Conventional Methods on Guideline in Predictive Diagnosis: A Systematic Review

Abstract

Background: Diabetes is a serious global health problem, and effective methods for its prediction and management are essential. Conventional diagnostic approaches typically rely on tests such as oral glucose tolerance test (OGTT), fasting plasma glucose (FPG) and glycated hemoglobin (HbA1c). Machine learning has the potential to enhance diagnostic accuracy; however, its performance and alignment with clinical guidelines require thorough evaluation.

Methods: This narrative review examines the effectiveness of machine learning in the early diagnosis of diabetes. Articles were selected based on predefined criteria and analyzed in terms of algorithm classification, output measures, involvement of clinical experts, and interpretability. Evaluation metrics such as accuracy, area under the curve (AUC), specificity and sensitivity were used to assess algorithmic performance. Relevant studies comparing prediabetes diagnosis using artificial intelligence and conventional methods were reviewed, and clinical guidelines from both domains were extracted and compared.

Results: Analysis of 41 articles showed that ANN, LR, and DNN were the most frequently used algorithms. Only 2% of the studies incorporated clinical rules and physician involvement, and 12% demonstrated model interpretability. While conventional methods rely on HbA1c and FPG tests, no clinical guidelines currently exist for AI-based diagnosis. Machine learning algorithms outperformed traditional methods, showing 29% higher sensitivity and 23% higher specificity.

Conclusion: Although artificial intelligence demonstrates superior performance in prediabetes diagnosis, limitations such as lack of interpretability and the absence of standardized clinical guidelines hinder its current clinical application. Addressing these challenges could enable AI to become a more efficient and reliable diagnostic tool