Research

Generative motion, and AI that earns clinical trust

I work on two threads that keep converging. The first is generative modelling of human motion — learning discrete representations of movement and conditioning them on language and audio. The second is medical AI, where the question is less about raw accuracy and more about whether a model changes what a clinician actually does.

Areas

What I am working on

01 — Motion synthesis

Language- and audio-conditioned human motion

Extending masked motion models to richer pose representations. Moving from a 263-dimensional to a 272-dimensional representation lets the model carry direct SMPL joint rotations in 6D form, which removes the inverse-kinematics artefacts that normally appear during BVH conversion and preserves subtle detail like joint twist.

The current system reaches an FID of 0.093 with roughly two-to-three second inference, and exports animation-ready output for Unity, Unreal Engine, and Blender.

Masked motion modelsSMPL 6D HumanML3DBVH
02 — Representation learning

Discrete motion embeddings

A VQ-VAE encoder/decoder that compresses motion into a discrete codebook so a transformer can model movement the way a language model handles tokens. Training used a 512-entry codebook with 2× temporal downsampling.

Codebook resets and EMA stabilisation brought reconstruction error down to an L2 of 0.0159 and an L1 of 0.0640 — low enough that the decoded motion is visually indistinguishable from ground truth.

VQ-VAECodebook resetsEMA
03 — Clinical AI

Decision support that clinicians will use

With the Duke Heart Center and the UNC School of Medicine I have worked on algorithms for organ donation after circulatory death, timing tools for ECMO in SVC syndrome, and a meta-analysis of whether deployed AI actually improves patient outcomes.

The donation algorithm projects a 94% improvement in success rate while reducing ICU workload — the kind of result that only matters if the workflow around it is realistic.

DCDECMOBiomedCLIPSmolVLM
04 — Systems

Making research runnable

Research code that only runs on one machine is research nobody can build on. I spend a lot of time on the deployment path: CUDA optimisation that took text-to-motion inference from 90 seconds to 7, audio preprocessing from four minutes to 30 seconds, and FastAPI services with pre-loaded models behind them.

Generated motion is retargeted onto TADA avatars and rendered in-browser with Three.js, so a result can be inspected the moment it exists.

CUDAFastAPINext.jsThree.js
Publications

Papers

Author order as submitted. Status reflects the most recent editorial decision.

Is Artificial Intelligence Saving Lives? A Meta-Analysis of Real-World Clinical Impact

Chouffani El Fassi, S., Ngan, Z., Alzouby, I., et al.

Meta-analysis examining the real-world clinical impact of AI interventions across healthcare settings.

In revision after peer review · Nature Communications

Gaps and Opportunities in the Medical AI Market

Alzouby, I., Chouffani El Fassi, S., Shahrour, L., et al.

Comprehensive analysis of current gaps and future opportunities in the medical AI industry.

Under review · Nature Medicine

Computer Algorithms to Optimize Organ Donation After Circulatory Death (DCD)

Chouffani El Fassi, S., Alzouby, I., Yüksel, I., Khan, T., et al.

Novel algorithms designed to improve organ donation outcomes after circulatory death.

Under review · Nature Medicine

AI Tool to Time ECMO for SVC Syndrome Relief: A Use Case in Nonseminomatous Germ Cell Tumor Resection

Chouffani El Fassi, S., Khoury, L., Alzouby, I., Haithcock, B.

AI-based tool for optimizing ECMO timing in nonseminomatous germ cell tumor resection cases.

Under review · Nature Medicine
Dissemination

Conference presentations

Oral Presentation, Cardiac Surgery
63rd Eastern Cardiothoracic Surgical Society Annual Meeting
2025
Oral Presentation, Cardiac Surgery — Best Oral Presentation award
62nd Eastern Cardiothoracic Surgical Society Annual Meeting
2024
ESS-AI Tool — poster presentation
UNC Chapel Hill School of Medicine
2024
AI @ Work — FALD-sponsored professional development session
UNC Chapel Hill School of Medicine · invited panellist
2025
Lab seminar

Reading-group presentations

Papers and tutorials I have presented to the group, mostly tracing the line from autoencoders through discrete latents to modern motion and video generation.

From Autoencoders to VQ-VAEs: A Mathematical Timeline
Tutorial
Introduction to Human Motion Generation Models
Tutorial
BAMM: Human Motion Modeling & Architecture
Paper
Training BAMM — Bidirectional Autoregressive Motion Model
Tutorial
MoMask: Generative Masked Modeling of 3D Human Motions
Paper
Diffusion Transformers with Representation Autoencoders
Paper
Index Backpropagation Quantization
Paper
Taming Teacher Forcing for Masked Autoregressive Video Generation
Paper
Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow
Paper
Soft-Masked Diffusion Language Models
Paper
Live

Deployed research systems

Research that left the lab and can be used in a browser today.

Text-to-motion

BAMM Motion Studio

Try our text-to-motion platform live — research commercialized with my advisor, built to take academic motion generation off the page and into creators' hands.

Clinical tooling

ECMO Timing Tool

A web front end for the ECMO timing work developed with the UNC School of Medicine, used to walk clinicians through the model's reasoning.

Centre

UNCC AI4Health

The AI4Health Center site, where much of the motion-modelling and clinical AI work at UNC Charlotte is coordinated.

Interested in collaborating?

I'm open to research collaborations across motion generation, medical AI, and embodied systems.