Evaluating AI-Generated Virtual Patient Videos for Psychotherapy Education Through Non-Verbal Behavioural Similarity to Standardized Patient

Sammanfattning

Introduction: Generative AI video models are increasingly able to create realistic human-like videos from text prompt. These Text-to Video (T2V) models can be used to generate Virtual Patient (VP) videos. But their suitability for psychotherapy education remains unclear. In psychotherapy training, patient realism depends not only on spoken dialogue but also on non-verbal behavior such as facial expression, eye gaze, and head movement. These cues are important because small inconsistencies in human-like patient videos may affect how natural or convincing the interaction appears. Research Question: This thesis investigates the question: How closely can generative AI video models produce psychotherapy patient interview videos with non-verbal behavior comparable to actor-acted standardized patient videos? Method: A quantitative comparative evaluation was conducted using 20 actor-acted standardized patient videos from two psychotherapy cases. Each video was replicated using two generative AI video models: Veo 3.1 Fast and Kling 3.0 Omni, resulting in 60 videos in total. An open-source facial behaviour analysis tool, OpenFace 2.2.0 was used to extract facial Action Units, eye-gaze, and head-pose features. The final analysis used 50 selected video-level features. Each AI-generated video was compared with its matched standardized patient reference using robust normalized absolute distance scores. Veo and Kling were then compared using Wilcoxon signed-rank tests. Results: Kling showed lower overall distance from the standardized patient references than Veo in all 20 matched video groups. Kling had a median overall distance of 0.163, compared with 0.391 for Veo. The overall difference was statistically significant. Feature-group analysis showed that Kling was significantly closer in six out of seven non-verbal behavior groups after FDR correction. Discussion: The findings suggest that, within this dataset and measurement approach, Kling produced psychotherapy patient interview videos with OpenFace-measured non-verbal behavior more comparable to actor-acted standardized patient videos than Veo. However, the results should be interpreted as computational similarity in selected non-verbal features, not as direct evidence of emotional realism, perceived authenticity, or educational effectiveness.