EARS-UDE: Evaluating Auditory Response in Sensory Overload with Universal Differential Equations

TL;DR

The paper addresses autism-related auditory hypersensitivity by targeting heterogeneity in sensory processing. It introduces Universal Differential Equations to hybridize a mechanistic sensory-adaptation ODE with a learnable neural component, achieving interpretable personalization while capturing nonlinear dynamics. Results show a 90.8% improvement over pure Neural ODEs and a 73.5% reduction in parameters, with accurate recovery of $\alpha$, $\gamma$, and $\delta$ and a quantified overload risk of $17.2\%$ for pulse stimuli. This framework supports personalized interventions and wearable/clinical applications, while noting the need for clinical validation, time-varying parameters, and multi-modal extensions.

Abstract

Auditory sensory overload affects 50-70% of individuals with Autism Spectrum Disorder (ASD), yet existing approaches, such as mechanistic models (Hodgkin Huxley type, Wilson Cowan, excitation inhibition balance), clinical tools (EEG/MEG, Sensory Profile scales), and ML methods (Neural ODEs, predictive coding), either assume fixed parameters or lack interpretability, missing autism heterogeneity. We present a Scientific Machine Learning approach using Universal Differential Equations (UDEs) to model sensory adaptation dynamics in autism. Our framework combines ordinary differential equations grounded in biophysics with neural networks to capture both mechanistic understanding and individual variability. We demonstrate that UDEs achieve a 90.8% improvement over pure Neural ODEs while using 73.5% fewer parameters. The model successfully recovers physiological parameters within the 2% error and provides a quantitative risk assessment for sensory overload, predicting 17.2% risk for pulse stimuli with specific temporal patterns. This framework establishes foundations for personalized, evidence-based interventions in autism, with direct applications to wearable technology and clinical practice.

Figures (4)

• Figure 1: Classical ODE model of sensory adaptation. (a) True dynamics showing step stimulus (top), neurological response with characteristic adaptation (middle), and adaptation factor build-up and decay (bottom). (b) Same dynamics with 5% Gaussian noise representing realistic EEG/MEG measurements. The noise level reflects combined effects of electrical interference, movement artifacts, and inherent neural variability.
• Figure 2: Performance comparison on noisy sensory data. Top: Step stimulus active from 10-40s. Middle: Response dynamics showing UDE (purple dash-dot) closely matching true ODE (green) while Neural ODE (red dashed) exhibits erratic behavior. Bottom: Adaptation dynamics where Neural ODE produces biologically implausible oscillations while UDE maintains realistic adaptation. Black dots show noisy observations the models must learn from.
• Figure 3: (a) Training loss evolution showing UDE converging to 10× lower loss (0.002606) than Neural ODE (0.028115). Note logarithmic scale and learning rate reduction at epoch 500. (b) Sensory overload risk assessment for pulse stimuli. Red shading indicates periods where response exceeds threshold (R > 0.8). Risk score: 17.2% of time above threshold.
• Figure 4: Future response prediction demonstrating generalization to novel stimulus patterns. The trained UDE accurately predicts responses to pulse stimuli (not seen during training with step stimulus). Vertical line at t=60s marks end of training data, showing successful extrapolation. This predictive capability enables assessment of real-world environments before exposure.