Does random-pair cross-attention improve multimodal self-supervised pediatric sleep representations?

Random signal pairs instead of fixed modality sets

The paper asks whether a self-supervised model can learn useful relationships between physiological sleep signals by training on randomly chosen pairs of channels rather than a fixed modality combination. SleepPairNet treats each polysomnography channel as its own modality. At every training iteration it samples two of the 16 NCH channels, masks half of each one's temporal tokens, and lets each modality's tokens attend to the other modality through cross-attention. Because the pair changes each step, one model is exposed to many signal combinations instead of one.

This design matters for a practical reason the paper states directly: fixed-combination models such as SynthSleepNet need new multimodal pretraining whenever the channel combination changes. A single SleepPairNet model, by contrast, handles all 120 pairs formed from its 16 channel-specific encoders without further cross-modal pretraining, which the authors say reduces compute and storage costs.

Two-stage pretraining separates per-channel learning from pairing

Pretraining happens in two stages. Stage 1 independently pretrains a NeuroNet encoder for each of the 16 channels, using masked reconstruction of CNN backbone features plus an NT-Xent contrastive loss on two independently masked views, with 75% of tokens masked per view. Stage 2 freezes those per-channel backbones and encoders, then trains the cross-attention encoder and decoder on random pairs, masking each modality at a rate of 0.5 and computing reconstruction loss only at masked positions.

The paper's ablation isolates Stage 2 by comparing SleepPairNet against a "Concat" model that mean-pools and concatenates the same Stage-1 representations without cross-modal pretraining. That comparison is the cleanest evidence for what random-pair cross-attention itself adds, separate from the strength of the per-channel encoders.

Respiratory tasks gain most on the NCH test set

On the NCH test set, SleepPairNet outperformed all evaluated baselines on the respiratory tasks in both F1 and AUPRC, the metrics the authors emphasize because event detection is extremely imbalanced. The largest gain was for hypopnea, at 5.8 percentage points in both F1 and AUPRC. The paper reports task prevalence of 0.8% for apnea, 1.9% for hypopnea, 8.7% for oxygen desaturation, 4.6% for EEG arousal, and 4.6% for arousal, which is why accuracy alone would be misleading and why AUPRC is reported alongside F1.

The picture is not uniform. For sleep scoring and EEG arousal, SynthSleepNet and SleepFM performed best respectively, and SleepPairNet's F1 for sleep scoring was 3.6 points below its own Concat ablation. The authors interpret the remaining EEG-task gap partly as a consequence of pretraining exposure: SleepFM was pretrained jointly on brain, respiratory, and EMG signals, and the evaluated SynthSleepNet models on EEG–EOG/EMG combinations, whereas SleepPairNet samples pairs from all 16 channels and so trades repeated EEG-centric exposure for broader coverage.

Cross-attention adds respiratory gains over concatenation

The Concat ablation shows where Stage-2 cross-modal pretraining helps. SleepPairNet improved F1 over Concat by 3.6 to 4.8 percentage points on NCH respiratory tasks, while the two stayed within 2.2 points on sleep scoring and arousal. Concat itself was often competitive with existing multimodal self-supervised models, which the authors read as evidence that the Stage-1 encoders provide a strong foundation for Stage 2.

That pattern supports a narrow reading: the random-pair cross-attention stage is associated with respiratory-task improvements rather than across-the-board gains. The paper does not report a fixed-pair cross-attention variant, and it lists comparing random- and fixed-pair pretraining as future work, so the specific contribution of randomness versus cross-attention is not separated in the current results.

Transfer to CHAT widens respiratory gains

For cross-dataset evaluation, the NCH-pretrained models were frozen and only linear classifiers were trained on 422 baseline CHAT PSGs, restricted to the 12 channels overlapping with NCH. SleepPairNet again outperformed all baselines on every respiratory task, with gains larger than on NCH, reaching up to 14.4 percentage points in AUPRC for desaturation. Some NCH channels are unavailable in CHAT: for apnea, EOG ROC–M1 was replaced by EEG F3–M2, and for hypopnea, EEG CZ–O1 by EEG F4–M1. Even with those substituted pairs, SleepPairNet gained 5.4 and 13.1 percentage points in F1 for apnea and hypopnea respectively.

The authors present this as evidence that a single model transfers across both cohorts and channel combinations. On CHAT, the sleep-scoring gap to SynthSleepNet narrowed to 2.6 points in F1 and 2.4 points in AUPRC, which the authors suggest may reflect more robust transfer from exposure to diverse pairs. That explanation is the authors' interpretation, not a separately tested mechanism.

What the comparison does and does not establish

All representation models were pretrained on the same NCH train split, and SleepFM and PedSleepMAE were evaluated on the same task-specific modality pairs as SleepPairNet so that multimodal methods received identical inputs. SynthSleepNet, however, was restricted to modality configurations studied in its original work rather than extended to respiratory modalities, so its respiratory numbers reflect a narrower setup than SleepPairNet's. The reported baselines are those the authors chose to evaluate, not an exhaustive field comparison.

The evaluation also depends on pair selection: for each task, pairs were chosen from individual-channel rankings plus domain knowledge, using the pair with the highest SleepPairNet validation F1. That selection procedure favors the proposed model's configuration, and the paper's own future-work list includes evaluating on held-out channel pairs and extending beyond pairs to larger signal sets.

Where random-pair cross-attention earns its cost

The decision-relevant split in this paper is by task family, not by overall average. If the target is respiratory event detection, SleepPairNet's combination of random-pair pretraining and cross-attention produced the best F1 and AUPRC among the evaluated models on both NCH and CHAT, with the CHAT desaturation AUPRC gain of 14.4 points the largest reported. If the target is sleep scoring or EEG arousal, the evidence points the other way: SynthSleepNet and SleepFM led those tasks, and SleepPairNet's advantage over its own Concat ablation was within 2.2 points.

The second consideration is channel flexibility. A team facing changing or partially unavailable channel sets gets one pretrained model covering all 120 pairs of 16 encoders, and the CHAT substitutions show that performance held when two NCH channels were replaced. The paper's own framing of the trade-off is exposure breadth versus repeated EEG-centric training, which is the most useful lens for deciding whether this pretraining scheme fits a given downstream task.

Sources used in this answer