Machine learning (ML) has been extensively applied in brain imaging studies to aid the diagnosis of psychiatric disorders and the selection of potential biomarkers. Due to the high dimensionality of imaging data and heterogeneous subtypes of psychiatric disorders, the reproducibility of ML results in brain imaging studies has drawn increasing attention. The reproducibility in brain imaging has been primarily examined in terms of prediction accuracy. However, achieving high prediction accuracy and discovering relevant features are two separate but related goals. An important yet under-investigated problem is the reproducibility of feature selection in brain imaging studies. We propose a new metric to quantify the reproducibility of neuroimaging feature selection via bootstrapping. We estimate the reproducibility index (R-index) for each feature as the reciprocal coefficient of variation of absolute mean difference across a larger number of bootstrap samples. We then integrate the R-index in regularized classification models as penalty weight. Reproducible features with a larger R-index are assigned smaller penalty weights and thus are more likely to be selected by our proposed models. Both simulated and multimodal neuroimaging data are used to examine the performance of our proposed models. Results show that our proposed R-index models are effective in separating informative features from noise features. Additionally, the proposed models yield similar or higher prediction accuracy than the standard regularized classification models while further reducing coefficient estimation error. Improvements achieved by the proposed models are essential to advance our understanding of the selected brain imaging features as well as their associations with psychiatric disorders.

中文翻译：

---

带加权正则化的分类模型提高神经影像特征选择的可重复性

机器学习 (ML) 已广泛应用于脑成像研究，以帮助诊断精神疾病和选择潜在的生物标志物。由于成像数据的高维度和精神疾病的异质亚型，ML 结果在脑成像研究中的可重复性引起了越来越多的关注。脑成像的可重复性主要在预测准确性方面进行了检查。然而，实现高预测精度和发现相关特征是两个独立但相关的目标。一个重要但研究不足的问题是脑成像研究中特征选择的可重复性。我们提出了一个新的指标来量化通过自举进行的神经影像学特征选择的可重复性。我们将每个特征的再现性指数 (R-index) 估计为大量自举样本中绝对平均差的变异倒数系数。然后，我们将 R 指数作为惩罚权重整合到正则化分类模型中。具有较大 R 指数的可重现特征被分配较小的惩罚权重，因此更有可能被我们提出的模型选择。模拟和多模式神经成像数据都用于检查我们提出的模型的性能。结果表明，我们提出的 R 指数模型可以有效地将信息特征与噪声特征分离。此外，所提出的模型产生与标准正则化分类模型相似或更高的预测精度，同时进一步减少系数估计误差。