Lost in Time? A Meta-Learning Framework for Time-Shift-Tolerant Physiological Signal Transformation

TL;DR

This work tackles the problem of transforming multimodal physiological signals under time-shift misalignment, a barrier to accurate ABP estimation from noninvasive sensors. It introduces ShiftSyncNet, a bi-level meta-learning framework with TransNet for waveform transformation and SyncNet for time-shift correction, leveraging Fourier-domain phase shifts to produce differentiable, aligned supervision. Through a k-step lookahead meta-gradient and a sample-selection training strategy, the approach effectively corrects misaligned labels and exploits shifted data, achieving state-of-the-art performance across real-world and public datasets and improving downstream SBP/DBP predictions to clinically relevant levels. The method promises robust, time-shift-tolerant physiological waveform transformation suitable for continuous, low-cost health monitoring, with potential extensions to other misalignment and artifact scenarios.

Abstract

Translating non-invasive signals such as photoplethysmography (PPG) and ballistocardiography (BCG) into clinically meaningful signals like arterial blood pressure (ABP) is vital for continuous, low-cost healthcare monitoring. However, temporal misalignment in multimodal signal transformation impairs transformation accuracy, especially in capturing critical features like ABP peaks. Conventional synchronization methods often rely on strong similarity assumptions or manual tuning, while existing Learning with Noisy Labels (LNL) approaches are ineffective under time-shifted supervision, either discarding excessive data or failing to correct label shifts. To address this challenge, we propose ShiftSyncNet, a meta-learning-based bi-level optimization framework that automatically mitigates performance degradation due to time misalignment. It comprises a transformation network (TransNet) and a time-shift correction network (SyncNet), where SyncNet learns time offsets between training pairs and applies Fourier phase shifts to align supervision signals. Experiments on one real-world industrial dataset and two public datasets show that ShiftSyncNet outperforms strong baselines by 9.4%, 6.0%, and 12.8%, respectively. The results highlight its effectiveness in correcting time shifts, improving label quality, and enhancing transformation accuracy across diverse misalignment scenarios, pointing toward a unified direction for addressing temporal inconsistencies in multimodal physiological transformation.

Figures (6)

• Figure 1: Time shift between multimodal signals degrades physiological signal transformation performance.
• Figure 2: Overview of ShiftSyncNet. It follows a bi-level optimization structure: the lower level updates TransNet$f_\theta$ to minimize the training loss on the misaligned dataset $D^{\prime}$ using labels corrected by SyncNet$h_\alpha$, while the upper level updates SyncNet to minimize TransNet’s loss on the clean metaset $D$.
• Figure 3: Signal transformation visualization under varying time-shift conditions. "Shift" indicates supervision with time-shift ${s}$; "ReconsY" and "PseudoY" represent predictive signals and pseudo-labels, respectively. For clarity, we also show the true original position of "ShiftY" as "ShiftYOrigin" and the input-aligned target as "AlignY", both unknown during training.
• Figure 4: Effectiveness of ShiftSyncNet.
• Figure 5: Training loss distribution of aligned and shifted samples for Co-teaching and ours at epochs 5, 30, and 60.