Operational integrity screening for telemedicine workflows: an explainable motion and audiovisual coherence framework.

Teleconsultations are exposed to digital impersonation and synthetic media attacks that can alter identity, articulation, or consent evidence, introducing emerging AI-enabled biosecurity risks to digitally mediated healthcare workflows. We evaluate an explainable integrity control based on motion dynamics and audiovisual temporal coherence, designed for conservative operation at extremely low false alarm rates with auditable evidence reporting to support governance and risk-based escalation. Our contribution is a reproducibility-oriented evaluation protocol and an evidence atlas of temporally grounded cues, together with a staged fusion that supports scalable prescreening and targeted verification under platform-style degradations characteristic of real-world dissemination chains. Building on prior biomarker-oriented analyses of physiological and structural cues for synthetic media detection, this work advances toward a calibrated, deployment-oriented integrity control architecture optimized for operational screening in telemedicine workflows. We use the DeepFake RealWorld (DFRW) dataset with 46,371 clips (229.28 h), combining 4,186 Open-Source Intelligence (OSINT) samples and 42,185 controlled, systematically degraded variants emulating recompression, resizing, filtering, and recapture. On the held-out binary-labeled test split, descriptor fusion reaches an AUC of 0.91 and achieves a true positive rate (TPR) of 18.5% (95% CI 16.2-20.8) at a false positive rate (FPR) of 0.1%, compared with 6.2% (95% CI 4.8-7.6) for an Xception baseline fine-tuned on the DFRW dataset. Microbenchmarked latencies motivate a two-stage deployment with explicit abstention and a structured integrity report aligned with healthcare governance and post-incident auditing requirements. This study evaluates telemedicine-oriented integrity controls using a benchmark dataset and does not claim demographic or clinical subgroup generalizability. Before deployment in clinical workflows, the proposed approach requires broader validation across age groups, ethnic backgrounds, speech characteristics, capture conditions, and medical conditions that may affect facial motion, articulation, or voice production. Prospective subgroup validation and governance-oriented assessment should therefore be treated as necessary directions for future work.