RESUMO
Generative Adversarial Networks (GANs) are promising analytical tools in machine learning applications. Characterizing atypical neurodevelopmental processes might be useful in establishing diagnostic and prognostic biomarkers of psychiatric disorders. In this article, we investigate the potential of GANs models combined with functional connectivity (FC) measures to build a predictive neurotypicality score 3-years after scanning. We used a ROI-to-ROI analysis of resting-state functional magnetic resonance imaging (fMRI) data from a community-based cohort of children and adolescents (377 neurotypical and 126 atypical participants). Models were trained on data from neurotypical participants, capturing their sample variability of FC. The discriminator subnetwork of each GAN model discriminated between the learned neurotypical functional connectivity pattern and atypical or unrelated patterns. Discriminator models were combined in ensembles, improving discrimination performance. Explanations for the model's predictions are provided using the LIME (Local Interpretable Model-Agnostic) algorithm and local hubs are identified in light of these explanations. Our findings suggest this approach is a promising strategy to build potential biomarkers based on functional connectivity.
RESUMO
Objective.The purpose of this article is to introduce readers to the concept and structure of the CAAos (CerebralAutoregulationAssessmentOpenSource) platform, and provide evidence of its functionality.Approach.The CAAos platform is a new open-source software research tool, developed in Python 3 language, that combines existing and novel methods for interactive visual inspection, batch processing and analysis of multichannel records. The platform is scalable, allowing for the customization and inclusion of new tools.Main results.Currently, the CAAos platform is composed of two main modules, preprocessing (containing artefact removal, filtering and signal beat to beat extraction tools) and cerebral autoregulation (CA) analysis modules. Two methods for assessing CA have been implemented into the CAAos platform: transfer function analysis (TFA) and the autoregulation index (ARI). In order to provide validation of the TFA and ARI estimates derived from the CAAos platform, the results were compared with those derived from two other algorithms. Validation was performed using data from 28 participants, corresponding to 13 acute ischemic stroke patients and 13 age- and sex-matched control subjects. Agreement between estimates was assessed by intraclass correlation coefficient and Bland-Altman analysis. No significant statistical difference between the algorithms was found. Moreover, there was an excellent correspondence between the curves of all parameters analysed, with the intraclass correlation coefficient ranging from 0.98 (95%CI 0.976-0.999) to 1.00 (95%CI 1 -1). The mean differences revealed a very small magnitude bias indicating an excellent agreement between the estimates.Significance.As open-source software, the source code for the software is freely available for noncommercial use, reducing barriers to performing CA analysis, allowing inspection of the inner-workings of the algorithms, and facilitating networked activities with common standards. The CAAos platform is a tailored software solution for the scientific community in the cerebral hemodynamic field and contributes to the increasing use and reproducibility of CA assessment.