Early Stage Diabetes Risk Prediction

Diabetes is a chronic condition that occurs when the body is unable to effectively use insulinor when the pancreas is unable to produce sufficient amounts of the hormone required to regulate bloodsugar levels. Conventional diagnostic methods, such as blood and urine glucose tests and family historyassessments, are commonly used by endocrinologists to detect diabetes. However, delayed detection remainsa significant concern as it increases the risk of severe complications over time. Therefore, early diabetesprediction is essential to improve patient outcomes and mitigate potential health risks.This study evaluates the performance of clustering techniques K-Means, Complete-Linkage, Expectation-Maximization, and Hierarchical K-Means for early-stage diabetes prediction. The impact of feature selec-tion using the Ant Colony Optimization (ACO) algorithm is analyzed to enhance clustering performance.The experiments are carried out using a data set comprising key diabetes risk indicators. The data setis preprocessed and the models are assessed using precision, recall, Rand index, and Fowlkes-Mallowsscore. The results indicate that ACO feature selection significantly improves the accuracy of the clustering,with Expectation-Maximization achieving the highest recall increase of 81.45% and Hierarchical K-Meansimproving the recall by 64. 93%. Precision scores for K-Means, Expectation-Maximization, and Hierar-chical K-Means reached 97% with ACO, while Complete-Linkage exhibited reduced effectiveness, withrecall dropping by 38.24%. False negative rates decreased significantly for Expectation-Maximization (-101instances) and Hierarchical K-Means (-87 instances) after feature selection. These findings demonstrate thatintegrating ACO with clustering methods enhances early diabetes prediction, providing a reliable approachfor clinical decision-making.

糖尿病是一种慢性疾病，当人体无法有效利用胰岛素，或胰腺无法分泌足够的调节血糖水平所需激素时，便会发病。内分泌科医生通常采用血液与尿液葡萄糖检测、家族史评估等常规诊断方法筛查糖尿病，但诊断延误仍是突出问题，随时间推移会大幅提升严重并发症的发生风险。因此，开展糖尿病早期预测对改善患者预后、降低潜在健康风险至关重要。
本研究评估了K均值（K-Means）、全链接聚类（Complete-Linkage）、期望最大化（Expectation-Maximization）以及分层K均值（Hierarchical K-Means）等聚类技术在糖尿病早期预测中的表现；同时分析了基于蚁群优化算法（Ant Colony Optimization, ACO）的特征选择对提升聚类性能的作用。实验采用包含关键糖尿病风险指标的数据集开展，先对数据集进行预处理，随后以精确率、召回率、兰德指数（Rand index）及福克思-马洛指数（Fowlkes-Mallows score）作为模型评估指标。
实验结果表明，蚁群优化算法的特征选择可显著提升聚类精度：期望最大化算法的召回率提升幅度最大，达81.45%；分层K均值的召回率提升64.93%。采用蚁群优化算法后，K均值、期望最大化以及分层K均值的精确率均达到97%；而全链接聚类的效果出现下滑，召回率下降38.24%。特征选择后，期望最大化与分层K均值的假阴性率分别大幅降低101例与87例。综上，将蚁群优化算法与聚类方法相结合可有效提升糖尿病早期预测效果，为临床决策提供可靠的解决方案。