Trajectory- and Topology-Aware Multimodal Modeling for Pediatric Sleep Event Prediction

Scott Ye, Harlin Lee

Keywords: pediatric sleep disorders, multimodal learning, time series analysis, PHATE, topological data analysis, electronic health records

TL;DR: We propose a multimodal late-fusion model that augments pediatric PSG embeddings with PHATE trajectory features, topological descriptors, and EHR context to improve rare sleep event prediction.

Abstract: Sleep disorders in children are common yet often underdiagnosed, and manual scoring of overnight polysomnography (PSG) is slow while labels for key events are sparse. We study 30-second pediatric PSG epochs represented by fixed embeddings from a multimodal masked-autoencoder. We investigate and augment these embeddings with (i) PHATE-derived per-epoch coordinates and whole-night movement descriptors, (ii) persistent-homology summaries computed on the high-dimensional embedding cloud, and (iii) routine EHR context. An AHI-stratified screen shows clinically coherent shifts in movement/topology. In predictive benchmarks, a late-fusion MLP that integrates all branches improves rare-event detection over a linear probe, leading in 3/4 binary tasks (Desaturation AUPRC = 0.370, EEG arousal = 0.484, Hypopnea = 0.290), while Apnea favors the EHR-only late-fusion variant (AUPRC = 0.147). Results suggest that clinical context and latent geometry/topology provide complementary signals beyond the generative embeddings, yielding interpretable links to disease burden and better performance under extreme imbalance.

Submission Number: 96