CGM: Chaos-Guided Graph Attention and Expert Routing for Autism Stereotypy Detection

Hafsa Aslam, Yue Li, Saba Aslam, Abdur Rasool, Qiao Pan, Muhammad Rizwan, Ovidiu BagdasararXiv 2026 · arXiv

TL;DR

This paper introduces CGM, a novel deep learning framework that detects autistic motor stereotypy (like hand flapping or rocking) by focusing on the unpredictability of movement rather than just its magnitude. By integrating a "chaos" signal directly into the architecture of graph attention networks and mixture-of-experts routing, the model achieves a 0.930 balanced accuracy and provides a reliable continuous severity index for clinical monitoring.

What Problem Does This Paper Address?

Motor stereotypies—repetitive, rhythmic movements—are core diagnostic criteria for Autism Spectrum Disorder (ASD). While tracking these movements is vital for monitoring therapy progress, it is traditionally done by hand, which is time-consuming and subjective.

Existing automated methods using Graph Neural Networks (GNNs) often suffer from "blind attention." They calculate joint importance based on spatial content alone, treating an erratic, flapping wrist the same as a still one if their static poses look similar. Furthermore, most models focus on movement magnitude (how much a child moves), but the authors argue that the true hallmark of stereotypy is its temporal irregularity or "chaos"—how unpredictably the movement unfolds over time.

Fig. 1. Temporal irregularity has no pathway in conventional skeleton pipelines: attention follows spatial similarity alone and routing treats calm

Methods

The authors propose the Chaos-Guided Model (CGM), which uses a measured irregularity signal to drive its internal computations.

  1. Chaos Formation: The model extracts 2D skeletons from video and calculates the Sample Entropy of angular velocities for each joint. This creates a "chaos tensor" that quantifies moment-to-moment unpredictability.
  2. Chaos-Modulated Graph Attention (CM-GAT): Instead of standard attention, CGM adds a learned pairwise term to the attention logits. This forces the network to concentrate its "focus" on irregular joints and pairs of joints that are moving chaotically together.
  3. Chaos-Guided Mixture-of-Experts (CG-MoE): The model uses a sparse gate to route each frame to specific "experts" (sub-networks). This gating is conditioned on the chaos signal, ensuring that chaotic stereotypies are processed by dedicated experts while calmer movements are handled by others.
  4. Stereotypy Severity Index (SSI): Parallel to classification, the chaos tensor is used to derive a continuous score (0–100) that reflects how far, how often, and how widely irregularity is elevated across the body.
The CGM framework extracts a chaos tensor to modulate graph attention and guide expert routing.

Datasets

The study utilizes several public datasets to ensure robustness:

  • SSBD & SSBD+: 86 ASD stereotypy clips from rehabilitation and home settings.
  • CMU Panoptic Studio: 30 clips of typically developing (TD) children at play, used as a comparison cohort.
  • MMASD: A public dataset of 453 clips from 22 children with ASD, used to replicate the component ordering and test generalization.
  • Held-out Panoptic sessions: Used to test the model's ability to transfer to entirely new recording environments without retraining.

Key Insights and Ideas

  • Chaos > Magnitude: The authors found that while traditional kinematic descriptors (like jerk or speed) often decrease during stereotyped clips in their cohort, Sample Entropy consistently rises. This confirms that unpredictability is a more reliable marker than raw movement intensity.
  • Structural vs. Feature Injection: Simply adding chaos as an input feature provides little benefit. The "aha moment" is that chaos must be injected at the structural level—modulating the attention weights and routing logic—to truly overcome the limitations of standard GNNs.
  • Mechanistic Validation: By "zeroing out" the chaos tensor during testing, the authors proved that the model actually uses the signal as intended: attention shifts away from irregular joints, and the routing pattern collapses, leading to a significant drop in performance.
Fig. 2. Chaos Formation. A seven-node anatomical graph is converted to timestamp-aware angular velocity ωt,j , isolating movement dynamics from static posture. Sample Entropy is evaluated on a stride-one window to quantify moment-to-moment unpredictability independent of magnitude; insets order three windows from regular to irregular. Stacking these estimates yields the dynamic chaos tensor that drives attention and routing in CGM.

Key Results

  • High Accuracy: CGM achieved a 0.930 balanced accuracy in subject-grouped cross-validation, outperforming the standard graph attention baseline by nearly 19%.
  • Robust Generalization: On the MMASD dataset, the model's performance gains were replicated. On held-out Panoptic Studio sessions, CGM maintained 0.86 specificity, whereas chaos-free baselines dropped as low as 0.29.
  • Reliable Severity Scoring: The Stereotypy Severity Index (SSI) showed high split-half reliability (0.87) and stayed consistent within individual children across different therapy sessions, making it a viable tool for longitudinal monitoring.
  • Ablation Gains: Chaos modulation alone (without the MoE) contributed a +0.132 gain in balanced accuracy, highlighting its importance as a structural prior.

Limitations and Future Work

While promising, the study has some limitations:

  • Cohort Size: The primary training cohort is relatively small (116 clips), and the comparison group (TD) initially relied on a single participant, though this was mitigated by testing on broader datasets.
  • 2D vs 3D: The model currently uses 2D pose estimation from standard video. While this makes it accessible, it may miss subtle movements in the Z-axis.
  • Future Directions: The authors suggest moving toward phone-based monitoring for home use and conducting larger, multi-site clinical validation studies to further prove the SSI's utility as a session-to-session outcome measure.

Summary updated

Abstract

Skeleton graph attention for autism video analysis weights joints by content alone, giving an erratic wrist the same standing as a still one. Stereotypy, however, is defined by how unpredictably movement unfolds rather than how much of it occurs: in our cohort, every amplitude-based kinematic descriptor falls in stereotyped clips, while windowed Sample Entropy of angular velocity alone rises (Hedges' g = +1.05). We compute this irregularity as a per-joint, per-frame tensor that structurally drives computation. The proposed chaos-guided graph attention and mixture-of-experts model (CGM) couples two mechanisms. First, a learned pairwise term added to the graph-attention logit concentrates attention on irregular joints and re-ranks the remaining edge weight toward jointly irregular pairs; its unconstrained coefficient neither vanishes nor inverts, and zeroing the tensor realigns attention in 94% of clips. Second, a gate conditioned on the same signal routes each frame to two of four experts, whose weights track frame-level irregularity. Under subject-grouped five-fold cross-validation on 116 clips, CGM reaches 0.930 balanced accuracy, exceeding the graph-attention baseline by +0.187; the component ordering replicates on the public MMASD dataset, and on held-out Panoptic Studio sessions, CGM keeps 0.86 specificity without retraining, against at most 0.35 for chaos-free baselines. In ablation, routing without irregularity yields no reliable gain, whereas chaos modulation alone contributes +0.132. A severity index from the same tensor reaches 0.87 split-half reliability, separates the cohorts with Hedges' g = +1.08, and stays consistent within children across therapy recordings, supporting video-only monitoring of stereotypy during intervention.
BibTeX
@misc{aslam2026cgm,
  author = {Hafsa Aslam and Yue Li and Saba Aslam and Abdur Rasool and Qiao Pan and Muhammad Rizwan and Ovidiu Bagdasar},
  title = {{CGM: Chaos-Guided Graph Attention and Expert Routing for Autism Stereotypy Detection}},
  year = {2026},
  eprint = {2611.25730},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.25730},
}
Figure