physiological time-series synthesis

Biomedical time-series data are severely scarce due to stringent privacy regulations and high annotation costs, hindering AI model development. To address this, we propose the first synthetic data generation framework for physiological signals based on deep temporal forecasting models, integrating statistical feature alignment and spectral preservation mechanisms to faithfully reconstruct multimodal signals—including EEG and EMG. Our method pioneers the adaptation of state-of-the-art forecasting architectures to physiological signal synthesis, jointly preserving temporal dynamics and frequency-domain structural characteristics while guaranteeing strict privacy compliance and enabling scalable, open-source data generation. Experiments demonstrate that the synthetic data significantly improve downstream AI model performance in cross-subject tasks, achieving generalization and stability comparable to real data—effectively bridging the critical gap between data demand and availability.

In mental health research, acquiring authentic electroencephalography (EEG) data is costly, privacy-sensitive, and often yields low-fidelity emotional/psychological signal representations. To address this, we propose a synthetic EEG generation method based on explicit cross-band correlation modeling. Our approach is the first to explicitly model and constrain inter-band dependencies among the δ, θ, α, β, and γ frequency bands, thereby guiding structured random sampling to produce high-fidelity, privacy-preserving synthetic EEG data. Experimental results demonstrate statistical equivalence between synthetic and real EEG data across distributional properties, band-wise correlations (mean absolute error < 0.02), and classification indistinguishability (random forest discrimination accuracy ≈ 50%), confirmed via PERMANOVA (p > 0.05)—outperforming existing methods. Furthermore, synthetic data enhances downstream task performance while eliminating risks of raw data leakage.

Existing ECG generation methods struggle to model long-range temporal dependencies, limiting synthetic signals to single beats or a few seconds—insufficient for clinical applications requiring multi-minute continuous recordings. To address this, we propose a three-tier diffusion-based synthesis framework: (1) a diffusion model generating high-fidelity single-beat waveforms; (2) a cross-beat temporal feature modeling module to capture rhythm dynamics; and (3) a feature-guided stitching mechanism that integrates multi-scale physiological constraints for coherent long-sequence synthesis. This work achieves the first end-to-end generation of diagnostic-grade, multi-minute (≥5 min) ECGs. Experiments demonstrate that the synthesized signals significantly outperform baselines in morphological fidelity and rhythm consistency. In downstream arrhythmia classification, models trained on our synthetic data achieve a 4.2% accuracy improvement, validating both clinical utility and methodological advancement.

This study addresses the challenge of missing critical physiological signals in clinical monitoring—often due to invasiveness, high cost, or unavailability—and the limitations of existing methods in handling multimodal, irregularly missing time-series data alongside static covariates. The authors propose ReCoGen, a two-stage framework: first, a masked autoencoder extracts compact, missingness-robust token sequences for each modality; second, a flow-matching generator synthesizes target signals by fusing these tokens with static conditions. The approach innovatively decouples conditional representation from generation and introduces a learnable cross-attention mechanism along with a dual-path fusion strategy for static conditions (via tokens and AdaLN). Evaluated across 16 tasks on AI-READI, MIMIC-III, and MIMIC-IV, ReCoGen achieves state-of-the-art downstream utility, surpassing even real-signal baselines in 13 cases.

This survey addresses the fragmented research landscape, inconsistent evaluation practices, and lack of standardized benchmarks for deep generative modeling of physiological signals (ECG, EEG, PPG, EMG). Following the PRISMA guidelines, we conduct the first systematic review in this domain, integrating bibliometric analysis with cross-modal technical categorization to comprehensively synthesize works published between 2018 and 2023. Our method yields a structured knowledge graph covering model architectures, publicly available datasets, evaluation metrics, and clinical application scenarios. We clarify the technical evolution trajectory, identify core challenges—including data scarcity, inter-modal heterogeneity, and non-comparable evaluations—and propose a standardized benchmarking framework. The findings provide theoretical foundations and practical guidance for algorithm development, reproducible experimentation, and clinical translation of generative models in physiological signal analysis.

This work addresses the challenge of limited performance in epilepsy EEG analysis due to data scarcity and high annotation costs by proposing GP-EEG, a novel framework that integrates Gaussian process regression with a domain-adaptive variational autoencoder to generate high-quality, interpretable epileptic EEG signals. The method effectively captures the long-range dependencies, high dimensionality, and non-stationarity inherent in EEG data. Experimental results on two public datasets demonstrate that the synthesized signals closely resemble real EEG recordings in both qualitative and quantitative evaluations, and significantly enhance the performance of downstream classification tasks.

This work addresses the challenge that existing electrocardiogram (ECG) models struggle to capture the evolving electrophysiological dynamics of the heart under external interventions such as pharmacological treatments. To overcome this limitation, the authors propose the first physiology-informed, action-conditioned ECG world model, which structurally embeds ordinary differential equation (ODE)-based physiological mechanisms into latent-space diffusion dynamics through energy-based regularization, enabling credible prediction of post-intervention ECG trajectories. The study further introduces a novel uncertainty-aware evaluation strategy grounded in sampling stochasticity to quantify clinical risk and its variability. Experiments on real-world clinical and drug-response datasets demonstrate that the proposed method significantly improves waveform fidelity and risk calibration, with predictions closely aligned with expert therapeutic preferences.

Existing digital twin models for physiological signals rely on point-wise evaluation metrics, which fail to capture critical dynamic characteristics such as oscillatory behavior, frequency, phase, and state transitions, often leading to misinterpretation. This work proposes TimeSynth, the first systematic evaluation framework designed specifically for assessing dynamic fidelity in synthetic physiological signals. TimeSynth integrates a parameterized generator fitted to real signals with multidimensional diagnostic tools to enable quantitative analysis of frequency spectra, phase coherence, and state-transition dynamics. Comparative evaluations across multiple neural network architectures reveal that models emphasizing local temporal structure better preserve these dynamic properties. Notably, even when conventional metrics appear comparable, phase errors can differ by up to 53° (approximately 123 ms), underscoring the necessity of aligning model architecture with the specific requirements of the target application.

This study addresses the performance limitations of deep neural networks (DNNs) in electrocardiogram (ECG) abnormality classification due to the scarcity of real-world ECG data. The authors propose a Gaussian mixture-based synthesis method that integrates medical prior knowledge to generate single-lead (Lead II) ECG signals encompassing four clinically significant abnormalities: atrial fibrillation, atrial flutter, premature ventricular contractions, and Wolff-Parkinson-White (WPW) syndrome. These synthetic data are used for pretraining multiple DNN architectures. The work provides the first systematic validation of the efficacy of synthetic-data pretraining for few-shot real ECG classification tasks, demonstrating substantial performance gains across three of the four abnormalities. Notably, atrial flutter classification accuracy improves by 33.2% on average, with greater benefits observed as the amount of available real training data decreases.

This study addresses the limitations in cardiac amyloidosis (CA) research imposed by small-scale, class-imbalanced, and heterogeneous electrocardiogram (ECG) datasets. To overcome these challenges, the authors propose an interactive generative adversarial network (GAN)-based tool that enables clinical researchers to train class-specific generators on demand, efficiently synthesizing high-quality, label-preserving ECG beats. The approach faithfully reconstructs the underlying data distribution, substantially alleviating data scarcity and class imbalance in CA. Furthermore, a graphical command-line interface enhances usability, offering a scalable data augmentation solution to support early diagnosis and patient stratification in cardiac amyloidosis.