Triaxial Accelerometry Sensor for Respiratory-Effort Phase Landmark Detection During Sleep: A Cross-Modal Weak Supervision Approach

Abstract

Reliable respiratory phase sensing is important for respiratory-synchronized hypoglossal nerve stimulation (HGNS), but conventional sensing approaches may impose additional burden. This pilot study investigated whether a surface triaxial accelerometry sensor (TAS) could detect thoracic respiratory inductance plethysmography (RIP) troughs, which were used as operational phase landmarks representing the transition from end expiration to subsequent inspiration, rather than as directly established physiological respiratory onsets. Ten participants underwent synchronized overnight polysomnography and surface TAS recording, providing 5377.3 min of analyzable data and 75,499 automatically generated RIP landmarks. Two complementary Conformer models were trained using Gaussian weak labels and combined using multi-offset dual-conformer fusion. Generalization was evaluated using outer leave-one-subject-out validation, with fusion weights and event thresholds selected from training-side validation data. At a ±1.0 s matching tolerance, the fusion achieved participant-mean precision, recall, and F1 scores of 0.78, 0.82, and 0.80, respectively, with a 95% confidence interval of 0.75–0.85 for the mean F1 score. It outperformed each component model in 9 of 10 participants; however, its F1 gain over otherwise identical single-grid fusion was 0.0036, below the prespecified exploratory advancement margin. These findings provide preliminary evidence that surface TASs can recover an RIP-trough-derived operational phase landmark in unseen same-center participants. Independent multimodal adjudication, external replication, and full-night causal testing are required before clinical or respiratory-synchronized HGNS applications.