Random forest and Shapley Additive exPlanations predict oxytocin targeted effects on brain functional networks involved in salience and sensorimotor processing, in a randomized clinical trial in autism.

Intranasal oxytocin (IN-OXT) has shown some promises in rescuing social deficits in autism spectrum disorder (ASD) as well as some inconsistencies in long-term trials. We conducted a target engagement study to study the precise effects of different doses of IN-OXT on brain resting-state functional connectivity (rsFC) in ASD. We examined the effects of varying doses of IN-OXT (0 IU, 8 IU, 24 IU, 48 IU) on rsFC in a double-blind, placebo-controlled, within-subject design in 30 male adults with ASD and 17 neurotypical controls (NT) receiving placebo. Random forest analysis was used to classify individuals as ASD or NT. Shapely Additive explanations values were calculated to rank brain functional networks by level of contribution to ASD deficits and to evaluate IN-OXT dose effects. The model predicted ASD diagnosis with an AUC of 94%. Hypoconnectivity between salience/empathy and visual networks, and hyperconnectivity between reward and sensorimotor networks and theory of mind networks were among the strongest predictors of ASD deficits. IN-OXT had a dose-dependent effect on rescuing both deficits described above. Overall, 48 IU dose was more effective, and 24 IU dose was more effective in those who have lower DNA OXT receptor methylation and lower severity of clinical symptoms. Higher doses of OXT might be necessary to enhance empathic responses, and ASD individuals with less support needs and with a preserved OXT system might benefit most from OXT treatment. Applying machine learning approaches in OXT research can provide data-driven unbiased results that can inform future clinical trials.