Counterfactual Causal Inference of Biomedical Signals: Unveiling Causal EEG Patterns for ASD Evaluation

The examination of complex diseases has greatly benefited from machine-learning techniques for interpreting biomedical signals. However, for diseases with multifactorial etiologies, such as Autism Spectrum Disorders (ASD), the inability of associative methods to distinguish between causal relationships and mere correlations often leads to unreliable outcomes. In contrast, causal models require well-defined causal structures, demanding a thorough understanding of their contributing factors. This study proposes a Generative Adversarial Network framework for Counterfactual Causal Inference (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">C²I-GAN</i>) to uncover causal patterns from biomedical signals without a predefined causal structure. The framework leverages graph attention network to identify key features for directing counterfactual generation, applies generative adversarial learning to produce task-oriented counterfactuals, and supports inference by evaluating how modifications to specific features affect diagnostic outcomes. A case study of ASD evaluation is conducted on a resting-state EEG dataset (74 ASD and 143 TD children) using <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">C²I-GAN</i> against state-of-the-art methods (Associative: GAT; Counterfactual: CounteRGAN, Omnixai, and CXGAN). The findings show that <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">C²I-GAN</i> identified T3, T4, O1, and O2 channels as causal patterns while recognizing C3 and C4 as merely associative, aligning with the latest neuroscience evidence where counterpart methods failed. In terms of performance, the model improved actionability by 30E and accuracy by 10E compared to other counterfactual methods, and increased accuracy by 5E while reducing training loss by 20E against associative method, demonstrating enhanced precision and efficiency.