MAA：数据集无关的衰老时钟评估研究

MAA: Dataset-Independent Aging Clock Evaluation Research

摘要: 近年来，基于不同衰老生物标志物的生物年龄预测得到了广泛的研究，虽然先前的工作可以从不同的衰老生物标志物中构建对应的生物年龄预测模型，但是由于所使用的数据集和生物标志物的差异，作为主要评估指标的绝对平均误差在不同数据集之间的表现也存在一定的差异。为了减少数据集差异对模型评估产生的不必要的影响，本文提出了一种新的评估指标MAA (Mean Absolute Accuracy)，通过以特定情况下的绝对平均误差为基准，建立近似分类问题中的准确度度量标准。本文使用当前主流的模型在不同的子数据集上进行了实验，并与传统评估指标进行比较，显示本文提出的MAA具有良好的数据集无关性，为跨数据集和跨衰老生物标志物评估生物年龄预测模型的性能提供了新的工具。

Abstract: In recent years, biological age prediction based on different aging biomarkers has been extensively studied. Although previous work has constructed corresponding biological age prediction models from various aging biomarkers, the mean absolute error, which is the primary evaluation metric, shows some variation across different datasets due to differences in the datasets and biomarkers used. To reduce the unnecessary impact of dataset differences on model evaluation, this paper proposes MAA (Mean Absolute Accuracy) a new evaluation metric. By using the mean absolute error in specific situations as a baseline, it establishes an accuracy measure in the context of approximate classification problems. This paper conducts experiments using current mainstream models on different sub-datasets and compares them with traditional evaluation metrics. The results show that the proposed MAA metric has good dataset independence, providing a new tool for evaluating the performance of biological age prediction models across datasets and aging biomarkers.