Computer-Vision Procedural Telemetry for Airway Guidance: A Public 30-Run Manikin Evidence-Package Audit
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Computer vision-enabled airway workflows can turn airway video into timestamped model-observation fields, but later blinded review and training studies require source-video linkage, run identifiers, quality-control status, and app/model provenance.
Objective
To audit whether a public post-reconciliation 30-run manikin evidence package from a computer vision-enabled video laryngoscopy workflow preserved prespecified, video-linked procedural telemetry in structured JSON, while keeping detection accuracy, report quality, and reviewer agreement outside the current claim.
Methods
Thirty manikin runs were captured on a HEALTHIBLE Intubation Simulator using an IntuBlade device connected to an iPhone 15 Pro Max. Six predefined conditions were tested with five runs each in planned round-robin order by a board-certified emergency physician operator. The author-affiliated team analyzed corrected Study Metrics JSON exports, the video manifest, app/model metadata, QC fields, and the frozen package checker after reconciliation against the assigned run guide. Blinded video review, independent analysis, and report-quality adjudication were not performed.
Results
After reconciliation, all 30 rows contained parseable Study Metrics JSON, a companion videoFilename, run-named Drive video status, QC pass status, and corrected identifiers matching assigned row labels (30/30 for each completeness field; descriptive exact binomial 95% CI, 88.4% to 100.0%). App/model metadata were complete: appVersion 3.3.0 (75), source revision b94cd63, Navigation model, model version 31, and detection threshold 0.1. The exported JSON target-state flag was true in 25 of 25 target-condition rows (95% CI, 86.3% to 100.0%) and false in 5 of 5 no-target controls (95% CI, 47.8% to 100.0%), with zero glottic-detected frames and zero acceptable-view time in no-target controls. Among target-condition rows, median time to first model-detected glottic target was 2 seconds (IQR 1 to 3), median acceptable-view duration was 2.2 seconds (IQR 1.0 to 3.8), and median glottic visibility was 35.8% (IQR 25.8 to 45.6).
Interpretation
The corrected package supports a bounded formative claim: simulated airway video can be represented as specified, video-linked computer-vision procedural telemetry after documented reconciliation. It supports package completeness, traceability, and assigned-condition consistency only; it does not establish native uncorrected export reliability, computer-vision detection accuracy, report quality, reviewer agreement, training effectiveness, autonomous guidance, tube-placement confirmation, clinical efficacy, or patient outcomes.
