Development and Cross-Cultural Evaluation of a Scoring Algorithm for the Biometric Attachment Test: Overcoming the Challenges of Multimodal Fusion with ``Small Data''

The Biometric Attachment Test (BAT) is a recently developed psychometric assessment that exposes adults to standardized picture and music stimuli-sets while simultaneously capturing their linguistic, behavioral and physiological responses, with the goal of objectively measuring their psychological attachment characteristics. Within this work, (I) we describe a new version of the BAT (v2) that implements a remote photoplethysmography method to obtain physiological measures from video alone. (II) We discuss the specific challenges we found when trying to develop an automatic scoring algorithm for the BAT v2 using machine learning: practicing multimodal fusion over a high-dimensional feature space with a small learning sample. We propose and evaluate an original combination of methods, including a three-step hybrid multimodal fusion procedure, that overcomes these challenges. (III) Using the proposed methodology, we train a scoring algorithm for the BAT v2 on a francophone sample, using the Adult Attachment Questionnaire as ground-truth. (IV) We then validate the scoring algorithm cross-culturally, testing its performance on an independent anglophone sample, showing low error and high correlation and serving as the BAT v2's first convergent validity evidence. We believe this work constitutes a breakthrough in the development of the first objective and automatic measure for adult attachment, and we hope that our ``small data'' learning methodology could be useful for other machine learning projects involving small samples coming from psychological research.