Privacy Preserving Diagnosing converts patient videos into standardized avatar representations that retain clinically relevant movement, posture, and facial cues while removing identifiable visual details. These neutral avatars serve as input for machine learning-based diagnosis models–with initial work in autism spectrum disorder (ASD)–demonstrating high classification performance. The system separates visual identity from the behavioral signal needed for diagnosis, creating a de-identified dataset that can be stored, shared, and analyzed with substantially lower privacy risk than raw video. The approach does not require specialized hardware, and it can be deployed in clinics, schools, or home environments. Clinicians and health systems can use the avatar outputs to support earlier screening, triage, and longitudinal monitoring, while reducing administrative burden. The same workflow can be extended other conditions in which facial expression or body movement contribute to diagnosis, enabling scalable, remote, and privacy-aware diagnostic support tools. Background: Demand for developmental and behavioral diagnostics has outpaced the supply of trained specialists, leading to long wait lists, limited follow-up, and inequitable access. Traditional telehealth and video-based assessments rely on identifiable recordings that raise privacy, consent, and data-sharing concerns, especially for children. Existing automated tools often require specialized hardware, depending on raw video, or still rely heavily on manual expert review. Privacy Preserving Diagnosing addresses these issues by providing a privacy-focused video representation that remains informative, reduces reliance on expert review, and enables a lower-risk dataset for model development and large-scale deployment. Applications:
Advantages: