Automated recovery of announced clip-level bolus categories from routine videofluoroscopic swallowing study audio: a single-centre development study

Purpose In the videofluoroscopic swallowing study (VFSS) workflow studied here, announced bolus categories were linked manually to fluoroscopic clips for review and reporting. We evaluated whether routine VFSS audio could support automated recovery of announced clip-level bolus categories for supervised fluoroscopic workflow support. Methods This single-centre development study used prospectively acquired recordings. It included 74 clips from 25 patients with an interpretable food-related announcement after exclusion of 3 technically unsuitable recordings and 14 clips (13 without a food-related cue and 1 that was unintelligible). A Whisper-based transcription pipeline combined a curated bolus-term dictionary with fixed heuristic rules. Dictionary and rule development and evaluation used the same 88 clips; surrogate target labels came from the same audio and were verified non-blind. No held-out or external validation dataset was available. Results In a post-hoc restricted development-set analysis, the pipeline made 66 calls among 74 eligible clips; 64 matched the surrogate target (86.5% of eligible clips; 95% CI 78.6–94.3; 97.0% of calls), 8 were no-call, and 2 were wrong calls. The legacy direct dictionary detector showed exact single-category agreement in 47/74 clips, while its candidate set contained the target in 63/74 and included multiple categories in 17 clips. Comparisons were descriptive because development and evaluation used the same recordings. Retaining the post-hoc term yielded 65/74 exact matches. Conclusions Under the studied cue-positive conditions, the pipeline showed apparent development-set agreement for recovery of announced bolus categories. It did not test correctness of the bolus actually administered or independent clinical classification. The single-centre dataset, same-audio surrogate target, non-blind verification, and post-hoc development preclude deployment or workload-reduction claims. A locked pipeline requires external validation with independently assigned labels.