📄 health informatics

Evaluating Generative Video AI for Standardized Psychiatric Patient Simulation With Graded Hygiene Deterioration.

This pilot study demonstrates the technical feasibility of using generative video AI to create standardized psychiatric patient simulations with graded hygiene deterioration, while highlighting that although appearance modulation is achievable, fine-motor artifacts necessitate expert human oversight before clinical deployment.

Original authors: Mwangi, B., Jabbar Abdl Sattar Hamoudi, H., Wu, M.-J., Martin, A., Soares, J. C., Soutullo, C. A.

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mwangi, B., Jabbar Abdl Sattar Hamoudi, H., Wu, M.-J., Martin, A., Soares, J. C., Soutullo, C. A.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Idea: "Digital Makeup" for Medical Training

Imagine you are training to be a doctor. To learn how to spot serious mental health issues, you need to see patients who look different. For example, a patient with severe depression might look very unkempt, while someone with schizophrenia might look disheveled.

Usually, to get these "looks" for training videos, you have to hire actors, put them in makeup chairs for hours, and film them. But there's a problem: You can't easily make an actor look slightly messy, then very messy, then extremely messy in the exact same video without re-filming everything. Plus, you can't ethically ask an actor to gain 30 pounds overnight or look like they haven't slept in weeks.

This paper asks: Can we use Artificial Intelligence (AI) to take one video of a calm, well-groomed actor and digitally "paint" different levels of messiness onto them, creating a whole library of training videos without ever re-filming?

The Experiment: The "Magic Re-Animator"

The researchers used a powerful new AI tool called Wan2.2-Animate-14B. Think of this AI as a "digital puppeteer."

The Setup: They took three existing videos of professional actors (playing patients with OCD, Schizophrenia, and Bipolar Disorder) sitting in a chair talking.
The "Makeup" Step: They used a text-to-image AI to take a single photo of each actor and generate five versions:
- Original: Clean and tidy.
- Mild: A little messy.
- Moderate: Clearly unkempt.
- Marked: Very disheveled.
- Severe: Extremely neglected.
The Animation Step: They fed these "messy" photos back into the video AI. The AI was told: "Take this messy face and body, and make it move exactly like the original actor in the video."

They ran this experiment 180 times, changing different settings (like how strictly the AI followed the instructions or how it handled the background) to see which settings worked best.

The Results: What Worked and What Didn't

The researchers measured the quality of the videos in two very different ways, like checking a car in two different ways:

1. The "Statistical Look" (Distributional Fidelity)

The Analogy: Imagine you are judging a painting by looking at the overall colors and lighting. Does the whole picture look like a real room?
The Finding: The AI did a great job here, but only if they used a specific setting called "Replacement Mode."
- Replacement Mode is like a green-screen effect where the AI keeps the real background (the wall, the chair) and only swaps the person. This looked very realistic.
- Animation Mode (where the AI tries to invent the background too) looked much worse and "off."
The Trend: As the AI made the patient look messier (from "Mild" to "Severe"), the video quality statistically dropped. The messier the prompt, the more the video drifted away from looking like the original real footage.

2. The "Physics Check" (Physical Plausibility)

The Analogy: Imagine watching a puppet show. Even if the puppet looks like a real person, if it has six fingers, or if its arm bends backward like a broken stick, you know it's fake. This is about the laws of physics and anatomy.
The Finding: The AI failed at this, and it didn't matter how they set the controls.
- The AI kept making "glitches" like extra fingers, weird joint movements, or shadows that didn't match the light.
- Crucially: These glitches happened just as often when the patient looked "Clean" as when they looked "Severe."
- The Takeaway: The messiness of the patient didn't cause the glitches. The glitches are a fundamental flaw in the AI's brain. The AI is good at copying the "vibe" of the video, but it doesn't truly understand how human bodies move physically.

The Main Conclusion

The paper concludes that:

Yes, we can do it: We can technically generate videos of patients with different levels of hygiene deterioration using AI.
But, be careful: The AI creates "fine-motor artifacts" (weird hands, impossible movements) that look like real medical symptoms (like tremors or movement disorders).
The Warning: Because the AI can accidentally create a "fake tremor" that looks real, these videos cannot be used for training doctors yet without a human expert checking them first. If a student sees a fake glitchy hand and thinks it's a real medical symptom, they could learn the wrong thing.

Summary in One Sentence

The researchers proved they can use AI to digitally "mess up" a patient's appearance for training videos, but the AI still makes weird anatomical mistakes (like extra fingers) that happen regardless of how messy the patient looks, meaning human doctors still need to double-check the work before using it in a classroom.

Technical Summary: Evaluating Generative Video AI for Standardized Psychiatric Patient Simulation With Graded Hygiene Deterioration

Problem Statement

The mental status examination (MSE) relies heavily on a patient's general appearance, grooming, and hygiene as immediate clinical indicators for diagnosis and monitoring, particularly in conditions involving self-neglect such as schizophrenia, severe depression, and neurocognitive disorders. However, the development of clinical AI tools and training curricula is hindered by the logistical impossibility of creating standardized patient (SP) video libraries that systematically manipulate these appearance characteristics.

Traditional simulation methods face strict physical and ethical limitations: actors cannot ethically undergo rapid, clinically meaningful weight changes, nor can conventional filming easily reproduce visible somatic or dermatologic changes associated with substance use or neurodegenerative illness. Consequently, the field lacks controlled, graded stimuli required for diagnostic calibration, inter-rater reliability training, and the validation of automated phenotyping systems. While generative AI offers a potential solution, the fidelity of re-animating a patient's video with deliberately altered appearance remains unproven, particularly regarding the risk of "hallucinated" fine-motor artifacts that could be misinterpreted as clinical pathology (e.g., tremors or dyskinesia).

Methodology

This pilot study investigated the technical feasibility of using the Wan2.2-Animate-14B latent video diffusion model to re-animate standardized psychiatric patients with systematically modified appearances.

Experimental Design:
The study employed a full factorial design ( $2 \times 3 \times 2$ ) crossing three generative variables:

Compositing Architecture: Animation mode (generating subject and background from scratch) vs. Replacement mode (segmenting the subject and compositing them back into the preserved original background).
Classifier-Free Guidance (CFG) Scale: Tested at levels 1, 4, and 6 to control prompt adherence versus diversity.
Pose Retargeting: Enabled vs. disabled to adjust skeleton keypoints to match reference anatomy.

Data Generation:

Source Material: Three SP identities ("Ben" with OCD, "Karthik" with Schizophrenia, "Robbin" with Bipolar Disorder) were drawn from a validated public repository (Martin et al., 2020).
Appearance Perturbation: Static reference portraits were extracted and modified using the Google Gemini text-to-image model to create five conditions: unmodified baseline, and four escalating levels of hygiene deterioration (mild, moderate, marked, severe). Prompts specifically targeted hair, facial hair, and attire while prohibiting changes to identity or somatic geometry.
Preprocessing: Generated images were upscaled using ESRGAN to prevent temporal collapse during diffusion.
Output: The model synthesized 180 unique 10-second videos where the reference subject performed the kinematics of the original driving footage.

Evaluation Metrics:

Distributional Fidelity: Measured via Fréchet Video Distance (FVD) using a ResNet3D-18 feature extractor to quantify the divergence between generated and ground-truth videos.
Physical Plausibility: Evaluated using a multimodal large language model (Qwen3-Omni-30B) acting as an automated physics-aware rater. The model scored videos on a 5-point ordinal scale across five dimensions:
- Finger count consistency (S1)
- Joint movement physics (S2)
- Arm count stability (S4)
- Kinetic momentum conservation (S5)
- Shadow consistency (S6)

Statistical Analysis:
Generalized Estimating Equations (GEE) were used for FVD analysis, while non-parametric tests (Kruskal-Wallis, Mann-Whitney U) analyzed ordinal physics scores. Dose-response relationships between hygiene severity and quality were assessed via Spearman rank correlation and covariate-adjusted GEE.

Key Results

1. Generation Mode Dominates Fidelity

Replacement mode significantly outperformed Animation mode in distributional fidelity (Mean FVD: 8.6 vs. 19.4; Cohen's $d = 1.84$ ).
Variance decomposition showed generation mode accounted for 46.1% of FVD variance, whereas hygiene level accounted for 25.6%.
Animation mode required synthesizing the entire scene, leading to greater distributional drift, while replacement mode constrained the generative space to the subject region, preserving environmental statistics.

2. Hygiene Severity Degrades Fidelity Monotonically

There was a significant, monotonic positive association between the severity of hygiene perturbation and FVD ( $\rho = 0.479, p < 0.001$ ).
As appearance modifications became more severe (from mild to severe), the generated videos drifted further from the ground-truth distribution. All perturbed levels differed significantly from the baseline.

3. The FVD-Physics Dissociation

Critical Finding: While distributional fidelity (FVD) degraded with increased hygiene perturbation, physical plausibility scores did not follow a similar trend.
Fine-motor artifacts (e.g., finger hallucinations, implausible joint trajectories) appeared at comparable rates regardless of whether the reference image was unmodified or severely perturbed.
Only one physics dimension (Arm count stability, S4) showed a significant difference between generation modes, with replacement mode scoring higher.
Correlation between FVD and physics dimensions was weak (strongest was Arm count, $\rho = -0.33$ ), indicating that a video can be distributionally realistic while still containing anatomically impossible artifacts.

4. Optimal Configuration

Combinatorial ranking identified Replacement mode with CFG=1 and retargeting disabled as the globally optimal configuration for minimizing FVD while maintaining physics scores.

Significance and Claims

The paper claims to provide the first systematic empirical evidence that latent video diffusion models can re-animate standardized psychiatric patients with deliberately altered appearances across graded levels of hygiene deterioration.

The study establishes three primary conclusions:

Technical Feasibility: It is technically achievable to generate appearance-modulated clinical video libraries without repeated filming, converting stimulus production from a laborious recording process into a scalable video generation task.
Architectural Necessity: For clinical training applications, replacement mode is superior to animation mode, as it preserves the spatial environment and minimizes distributional drift.
Limitation of Current Models: The persistence of fine-motor artifacts, which are invariant to the severity of appearance modification, underscores that these are intrinsic limitations of the generative architecture rather than byproducts of the perturbation task.

Implications for Clinical Deployment:
The authors assert that because distributional fidelity (FVD) and physical plausibility are dissociated, FVD alone is insufficient as a quality gate for clinical AI. A video can be globally realistic yet contain hallucinated tremors or limb distortions that mimic psychomotor pathology. Therefore, before these materials can be safely deployed in educational or translational settings, expert human oversight and physics-aware evaluation are necessary to screen for anatomical artifacts.

The study does not claim that these videos are currently ready for clinical use but rather that they provide a benchmark framework and optimal configuration parameters for future model development and validation. It highlights the need for separate evaluation of global realism and anatomical correctness in high-stakes clinical applications.