Abstract
Many intensive care unit bedside devices lack interoperable digital outputs, so clinically important readings are transferred into records by intermittent manual charting. Camera-based optical character recognition (OCR) offers an interface-independent alternative, but its value depends on identifying acquisition conditions under which structured extraction remains reliable and conditions under which output should be withheld. We evaluated a frame-sampled pipeline comprising display localisation, segmentation, pre-processing, OCR, clinical dictionary-based normalisation, and layered quality-control gating. A total of 151,763 usable frames were included in the primary evaluation across seven bedside instruments, three camera platforms, distances of 36 to 390 inches, viewing angles of 0 to 80 degrees, five lighting conditions, and two layouts. Per-stratum performance intervals are binomial Wilson and Clopper–Pearson intervals; the sampling-density transition estimates include bootstrap uncertainty. In an independently double-annotated stratified subset of 750 frames, Cohen’s kappa ranged from 0.89 to 0.97 across categorical annotations. Exact agreement was 96.4% for numeric values (95% CI 94.8–97.6%) and 94.8% for strings and labels (95% CI 93.0–96.3%); pooled exact agreement was 95.1% (95% CI 93.4–96.5%) and pooled kappa was 0.93 (95% CI 0.90–0.95). At 36 inches and near-frontal orientation, device-level sensitivity and precision were 0.990 and 0.997, 0.931 and 0.937, and 0.782 and 0.787 for the mirrorless, Raspberry Pi, and consumer webcam platforms, respectively. By 74 inches at near-frontal orientation, device-level sensitivity had fallen to 0.617, 0.515, and 0.309, respectively, showing that physical distance alone does not define a stable bedside operating range. Effective display sampling density described extraction performance more consistently across cameras than physical distance alone; the 0.90 complete device-level extraction transition was approximately 9.8 px/cm (bootstrap 95% CI 8.7–11.4 px/cm), a workflow-specific provisional empirical estimate rather than a validated operational threshold. Precision remained comparatively high as sensitivity declined; this is reported as observed system-level behaviour rather than causal evidence for any individual pipeline stage.