Naturalistic fMRI Language Mapping With Rotation Forest

Naturalistic fMRI offers a patient-friendly alternative to task-based language mapping but presents analytical challenges due to complex brain responses. This study presents an automated framework for language mapping using naturalistic fMRI analysed with a Rotation Forest (RotF) classifier. Participants were presented a televised quiz show during fMRI acquisition and completed seven task-based language paradigms, which were used as reference maps. RotF models were trained on voxel-wise time series labelled using task-based activation maps and evaluated on independent participants. Performance was assessed using the structural similarity index measure (SSIM) and the language lateralisation indices (LI). RotF generated activation maps demonstrated strong spatial and lateralisation agreement with task-based activation, capturing both receptive and expressive language regions. Data quality, which was assessed through head motion and post-scan questionnaire score, influenced model performance and provided a practical quality control measure. Feature selection identified a short, informative stimulus segment, enabling reliable language mapping with a scan duration of approximately 5 min. These findings demonstrate the feasibility of accurate, efficient and patient-friendly language mapping using naturalistic fMRI combined with machine learning.