Machine Learning to Classify Simulated Psychotherapy States in Preschool Children from fNIRS

Artificial Intelligence (AI) is increasingly used in mental health research, yet there remain few applications focusing on early childhood populations. By analyzing neural signals from naturalistic paradigms that more closely resemble real-world therapeutic interactions, AI could offer early insight into therapy engagement or treatment response. The goal of the study was to test whether machine learning algorithms could classify functional near-infrared spectroscopy (fNIRS) brain activity of young children engaging in a therapy-like activity above random chance. 78 children were randomly assigned to a dyadic coloring task that prompted emotion-related thoughts and speech, designed to mimic a therapy-like interaction, or an identically structured interpersonal control condition without emotion-related prompts. Across 200 iterations of 5-fold cross-validation, logistic regression models accurately classified whether a child was engaged in the therapy-like task (mean AUC = 0.60 - 0.76). Performance in a-priori models was modest and variably significant against a permutation null (AUC = 0.60 - 0.63, p_perm = 0.044 - 0.13), whereas post-hoc exploratory models yielded improved and consistently significant performance (AUC = 0.72 - 0.76, p_perm = 0.005). Feature analysis revealed that lower mean activity and variability in the left ventrolateral prefrontal cortex while engaging in calm-emotion reflection increased the likelihood of being classified into the therapy-like condition. We show that naturalistic, brain-based measures captured via fNIRS, analyzed using simple and interpretable AI models, can detect psychotherapy-like states in young children, although modestly. While larger replications with external datasets are required, this approach holds promise for personalized, neuroscience-informed treatment routing and response monitoring.