Khuntia Adyasha, Buciuman Madalina-Octavia, Fanning John, Stolicyn Aleks, Vetter Clara, Armio Reetta-Liina, From Tiina, Goffi Federica, Hahn Lisa, Kaufmann Tobias, Laurikainen Heikki, Maggioni Eleonora, Martinez-Zalacain Ignacio, Ruef Anne, Dong Mark Sen, Schwarz Emanuel, Squarcina Letizia, Andreassen Ole, Bellani Marcella, Brambilla Paolo, Haren Neeltje van, Hietala Jarmo, Lawrie Stephen M, Soriano-Mas Carles, Whalley Heather, Taquet Maxime, Meisenzahl Eva, Falkai Peter, Wiegand Ariane, Koutsouleris Nikolaos
Department of Psychiatry and Psychotherapy, University Hospital, Ludwig-Maximilian University, Munich, Germany.
International Max-Planck School for Translational Psychiatry, Germany Max-Planck Institute of Psychiatry Munich, Munich, Germany.
Neurosci Appl. 2024 Dec 9;4:105407. doi: 10.1016/j.nsa.2024.105407. eCollection 2025.
The current biologically uninformed psychiatric taxonomy complicates optimal diagnosis and treatment. Neuroimaging-based machine learning methods hold promise for tackling these issues, but large-scale, representative cohorts are required for building robust and generalizable models. The European College of Neuropsychopharmacology Neuroimaging Network Accessible Data Repository (ECNP-NNADR) addresses this need by collating multi-site, multi-modal, multi-diagnosis datasets that enable collaborative research. The newly established ECNP-NNADR includes 4829 participants across 21 cohorts and 11 distinct psychiatric diagnoses, available via the Virtual Pooling and Analysis of Research data (ViPAR) software. The repository includes demographic and clinical information, including diagnosis and questionnaires evaluating psychiatric symptomatology, as well as multi-atlas grey matter volume regions of interest (ROI). To illustrate the opportunities offered by the repository, two proof-of-concept analyses were performed: (1) multivariate classification of 498 patients with schizophrenia (SZ) and 498 matched healthy control (HC) individuals, and (2) normative age prediction using 1170 HC individuals with subsequent application of this model to study abnormal brain maturational processes in patients with SZ. In the SZ classification task, we observed varying balanced accuracies, reaching a maximum of 71.13% across sites and atlases. The normative-age model demonstrated a mean absolute error (MAE) of 6.95 years [coefficient of determination ( ) = 0.77, < .001] across sites and atlases. The model demonstrated robust generalization on a separate HC left-out sample achieving a MAE of 7.16 years [ = 0.74, < .001]. When applied to the SZ group, the model exhibited a MAE of 7.79 years [ = 0.79, < .001], with patients displaying accelerated brain-aging with a brain age gap (BrainAGE) of 4.49 (8.90) years. Conclusively, this novel multi-site, multi-modal, transdiagnostic data repository offers unique opportunities for systematically tackling existing challenges around the generalizability and validity of imaging-based machine learning applications for psychiatry.
当前缺乏生物学依据的精神科分类法使最佳诊断和治疗变得复杂。基于神经影像学的机器学习方法有望解决这些问题，但构建强大且可推广的模型需要大规模、具有代表性的队列。欧洲神经精神药理学学院神经影像学网络可访问数据存储库（ECNP-NNADR）通过整理多站点、多模态、多诊断数据集来满足这一需求，从而实现协作研究。新建立的ECNP-NNADR包含来自21个队列的4829名参与者以及11种不同的精神科诊断，可通过研究数据虚拟合并与分析（ViPAR）软件获取。该存储库包括人口统计学和临床信息，如诊断结果以及评估精神症状的问卷，还有多图谱灰质体积感兴趣区域（ROI）。为了说明该存储库提供的机会，进行了两项概念验证分析：（1）对498例精神分裂症（SZ）患者和498例匹配的健康对照（HC）个体进行多变量分类，以及（2）使用1170名HC个体进行正常年龄预测，随后将此模型应用于研究SZ患者异常的脑成熟过程。在SZ分类任务中，我们观察到不同的平衡准确率，各站点和图谱的最高准确率达到71.13%。正常年龄模型在各站点和图谱中的平均绝对误差（MAE）为6.95岁[决定系数（ ） = 0.77， <.001]。该模型在一个单独留出的HC样本上表现出强大的泛化能力，MAE为7.16岁[ = 0.74， <.001]。当应用于SZ组时，该模型的MAE为7.79岁[ = 0.79， <.001]，患者表现出脑老化加速，脑年龄差距（BrainAGE）为4.49（8.90）岁。总之，这个新颖的多站点、多模态、跨诊断数据存储库为系统解决围绕基于成像的机器学习应用于精神病学的可推广性和有效性的现有挑战提供了独特机会。
