Resumo:
Autism Spectrum Disorder (ASD) is a neurodevelopmental condition whose diagnosis
remains centered on behavioral criteria, which can introduce delays and inter-rater variability.
Resting-state functional magnetic resonance imaging (rs-fMRI) enables investigation
of brain functional connectivity and, combined with machine learning models, has been
explored as an objective aid to diagnosis. However, choices of configuration parameters
and preprocessing strategies strongly influence feature quality and, consequently, classifier
performance and reproducibility. This work assesses the impact of different rs-fMRI preprocessing
configurations on the performance of models for ASD classification. Modular
software was developed to automate image preprocessing, feature extraction and selection,
and the training and evaluation of multiple classifiers. Data from the ABIDE I and II
datasets were used. Initially, a sex-balanced set with 72 participants (half with ASD)
was assembled, to which 108 different preprocessing configurations were applied, varying
spatial smoothing, mask threshold, noise removal, and dimensionality reduction parameters.
Connectivity features were derived using full correlation (FC), partial correlation (PC), and
their combination (FC+PC), followed by statistical tests and incremental feature selection
with features added sequentially in ascending order of p-value. Thirteen classification
algorithms were evaluated with leave-one-out validation, incrementally varying the number
of features used, totaling 113,698 experiments. In this initial set, the best accuracy reached
95.8% and specificity 100%, with multilayer perceptron (MLP)-based classifiers and the
combined use of FC+PC features standing out. Next, to simulate performance in a scenario
closer to clinical practice, the preprocessing pipelines with the best initial results were
applied to an independent set of 216 additional participants. Models were trained on
the first-stage participants and tested on this new set, where accuracy dropped to 61.4%
and specificity to 58.4%. These findings highlight the sensitivity of model performance to
preprocessing choices and reinforce the need for strategies that prioritize generalization.
