BiTimeCrossNet: Time-Aware Self-Supervised Learning for Pediatric Sleep

TL;DR

BTCNet tackles pediatric sleep analysis by introducing time-aware, multimodal self-supervised learning over overnight PSG. It combines unimodal MAE+contrastive pretraining with a random modality-pair cross-attention stage that is conditioned on global night-time context and refined via LoRA adapters. The approach yields superior representations across six downstream tasks on the NCH dataset and generalizes well to an independent CHAT cohort, especially for respiration-related events. By leveraging random modality sampling, cross-modal attention, and global temporal conditioning, BTCNet advances robust, transferable sleep representations with practical implications for scalable pediatric sleep analysis.

Abstract

We present BiTimeCrossNet (BTCNet), a multimodal self-supervised learning framework for long physiological recordings such as overnight sleep studies. While many existing approaches train on short segments treated as independent samples, BTCNet incorporates information about when each segment occurs within its parent recording, for example within a sleep session. BTCNet further learns pairwise interactions between physiological signals via cross-attention, without requiring task labels or sequence-level supervision. We evaluate BTCNet on pediatric sleep data across six downstream tasks, including sleep staging, arousal detection, and respiratory event detection. Under frozen-backbone linear probing, BTCNet consistently outperforms an otherwise identical non-time-aware variant, with gains that generalize to an independent pediatric dataset. Compared to existing multimodal self-supervised sleep models, BTCNet achieves strong performance, particularly on respiration-related tasks.

Figures (4)

• Figure 1: PHATE visualization moon2019visualizing of frozen embeddings from BTCNet (time-aware) and BCNet (non-time-aware) for a representative patient using SPO2 and CAPNO signals. Each point is 30 seconds of sleep colored by time of night. The time-aware model exhibits a smoother and more coherent global structure compared to the non-time-aware model.
• Figure 2: Overview of BTCNet. Stage 1: Unimodal self-supervised pretraining using masked reconstruction and contrastive learning to obtain modality-specific encoders. Stage 2: Cross-modal self-supervised learning with randomly sampled modality pairs, time-aware conditioning, and cross-attention.
• Figure 3: PHATE and UMAP embeddings for a representative patient with apnea using CAPNO and SPO2 channels. Positive events are indicated by $\times$ markers, while color denotes temporal order within the patient record.
• Figure 4: PHATE and UMAP embeddings for a representative patient with hypopnea using EEG O1-M2 and SPO2 channels. Positive events are indicated by $\times$ markers, while color denotes temporal order within the patient record.