AETHER: Video-Based Smartphone Free Flap Monitoring Using Spatial Phase Coherence and Reference-Corrected Augmented Imaging.

Background

Postoperative free flap surveillance relies on clinical inspection or hardware-dependent adjuncts. Remote photoplethysmography offers a contactless alternative, but motion, lighting variation, specular reflection, and camera autoexposure can mimic perfusion signals. AI-Enhanced Transcutaneous Hemodynamic Evaluation for Reconstructive surgery (AETHER) combines per-patch spatial phase coherence, in-image reference correction, and augmented overlay visualization.

Methods

We retrospectively analyzed bedside smartphone videos of 25 consecutive anterolateral thigh (ALT) free flaps treated between November 2025 and April 2026. No patient-level demographic, comorbidity, clinical-suspicion, or bedside-examination data entered the algorithm; AETHER was evaluated as a covariate-free, video-only signal-processing proof of concept. Flap and nonperfused reference regions of interest, each divided into 100 subregions, underwent plane-orthogonal-to-skin-based pulse extraction, band-pass filtering, and sliding-window correlation with a global mean reference. The Global Coherence Score (GCS) was defined as the percentage of synchronous subregions; corrected GCS was flap GCS minus reference GCS. Compromise was defined as clinically indicated surgical take-back with intraoperative confirmation of arterial compromise.

Results

Six flaps had arterial compromise, and 19 were vital. Raw flap GCS and reference GCS did not separate groups (area under the receiver operating characteristic curve 0.62 and 0.36, respectively). Corrected GCS showed strong observed in-sample case-level separation (area under the receiver operating characteristic curve 0.96, 95% confidence interval 0.87 to 1.00). At an exploratory 17% cutoff, sensitivity was 0.83 and specificity 0.89; predictive values were prevalence-dependent.

Conclusions

AETHER identified a video-derived perfusion-associated signal from routine smartphone video alone. These findings support covariate-free signal extraction but not recording-level diagnostic performance, lead time, or generalizability. Prospective validation is required before clinical use.