4 - (SYM 13) Ai-based Automated Clutter Ratings to Enhance Hoarding Disorder Assessment

Unbiased and precise assessment of hoarding disorder (HD) is crucial given its prevalence rate, adverse effects (e.g., health and safety concerns), limited treatment engagement, and challenges with clutter reduction. Hoarding is typically assessed using self and/or assessor report as well as clinical interview instruments. To address the visual aspect of HD, the paper-based Clutter Image Rating (CIR: Frost et al., 2008) scale includes sets of 9 “clutter-equidistant” photos for each of 3 rooms (i.e., bedroom, living room, kitchen) that are used to rate clutter severity. Ratings may be provided by clients, family members, clinicians, and human service personnel. While the CIR has been validated, it can be subjective (requiring ratings by a human), potentially biased by the assessor’s perception, and costly and time-consuming, if a clinician performs a home-based assessment. Our prior research utilized two data-driven approaches based on machine learning to develop an automatic, image-based rating of room clutter (Tooke et al., 2016; Tezcan et al., 2018). While these approaches showed promising results, the current study proposes a new automatic clutter assessment method to enhance accuracy.

We developed a new approach based on the visual transformer (ViT), commonly used in computer vision, to assess CIR values from images. We also applied extensive data augmentations during algorithm training. We tested the algorithm on a set of ~1,233 on-line images of hoarded rooms that we collected and were rated by a trained assessor using the CIR.   In four-fold cross-validation on this dataset, the new algorithm resulted in 86% correct classification rate within ±1 of the CIR rating, an improvement  compared to the best-performing prior method (Tezcan et al., 2018).

These results are very promising, improving an AI-based algorithm to automate CIR values, and are consistent with ratings by trained professionals. Next steps include further improvements to the algorithm and finalizing a newly developed mobile app to deploy the algorithm for validation in the community.  Automated real-time CIR assessment would remove human bias from CIR scoring, thus leading to a better precision and consistency across venues and time. Such an objective HD measure would also enhance the scalability of HD assessment, enable real-time monitoring and feedback, and facilitate the coordination of response among clinical and community providers.