Department of Software Convergence, Kyung Hee University, Yongin 17104, Korea.
Sensors (Basel). 2022 Oct 21;22(20):8077. doi: 10.3390/s22208077.
Brain structural morphology varies over the aging trajectory, and the prediction of a person's age using brain morphological features can help the detection of an abnormal aging process. Neuroimaging-based brain age is widely used to quantify an individual's brain health as deviation from a normative brain aging trajectory. Machine learning approaches are expanding the potential for accurate brain age prediction but are challenging due to the great variety of machine learning algorithms. Here, we aimed to compare the performance of the machine learning models used to estimate brain age using brain morphological measures derived from structural magnetic resonance imaging scans. We evaluated 27 machine learning models, applied to three independent datasets from the Human Connectome Project (HCP, = 1113, age range 22-37), the Cambridge Centre for Ageing and Neuroscience (Cam-CAN, = 601, age range 18-88), and the Information eXtraction from Images (IXI, = 567, age range 19-86). Performance was assessed within each sample using cross-validation and an unseen test set. The models achieved mean absolute errors of 2.75-3.12, 7.08-10.50, and 8.04-9.86 years, as well as Pearson's correlation coefficients of 0.11-0.42, 0.64-0.85, and 0.63-0.79 between predicted brain age and chronological age for the HCP, Cam-CAN, and IXI samples, respectively. We found a substantial difference in performance between models trained on the same data type, indicating that the choice of model yields considerable variation in brain-predicted age. Furthermore, in three datasets, regularized linear regression algorithms achieved similar performance to nonlinear and ensemble algorithms. Our results suggest that regularized linear algorithms are as effective as nonlinear and ensemble algorithms for brain age prediction, while significantly reducing computational costs. Our findings can serve as a starting point and quantitative reference for future efforts at improving brain age prediction using machine learning models applied to brain morphometric data.
大脑结构形态随衰老轨迹而变化，利用大脑形态特征预测个体年龄有助于发现异常衰老过程。基于神经影像学的大脑年龄被广泛用于量化个体的大脑健康状况，即与正常大脑衰老轨迹的偏差。机器学习方法正在扩大准确预测大脑年龄的潜力，但由于机器学习算法种类繁多，因此具有挑战性。在这里，我们旨在比较使用从结构磁共振成像扫描中提取的大脑形态学测量值来估计大脑年龄的机器学习模型的性能。我们评估了 27 种机器学习模型，这些模型应用于来自人类连接组计划（HCP，n = 1113，年龄范围 22-37）、剑桥老龄化和神经科学中心（Cam-CAN，n = 601，年龄范围 18-88）和信息提取图像（IXI，n = 567，年龄范围 19-86）的三个独立数据集。在每个样本中，我们通过交叉验证和一个看不见的测试集来评估性能。模型在 HCP、Cam-CAN 和 IXI 样本中的平均绝对误差分别为 2.75-3.12、7.08-10.50 和 8.04-9.86 年，以及预测脑龄与实际年龄之间的皮尔逊相关系数分别为 0.11-0.42、0.64-0.85 和 0.63-0.79。我们发现，在同一数据类型上训练的模型之间的性能存在显著差异，这表明模型的选择会导致预测脑龄的差异较大。此外，在三个数据集，正则化线性回归算法的性能与非线性和集成算法相似。我们的研究结果表明，对于大脑年龄预测，正则化线性算法与非线性和集成算法一样有效，同时显著降低了计算成本。我们的发现可以为使用机器学习模型应用于大脑形态计量数据来提高大脑年龄预测的未来努力提供起点和定量参考。
