xECG: Cross-Context ECG Analysis

• xECG is a cross-context electrocardiogram analysis framework that integrates deep learning, multi-scale signal processing, and adaptive filtering to address diverse ECG challenges.
• Foundation modeling with xLSTM layers and self-supervised learning enables robust ECG representation across varied clinical conditions, achieving high accuracy on multiple benchmarks.
• Robust biometrics and fetal ECG extraction methods in xECG utilize multi-scale convolution and adaptive filtering to ensure reliable patient authentication and precise fetal heart rate detection.

xECG refers to advanced methods for cross-context Electrocardiogram (ECG) analysis, encompassing multiple domains: (1) foundation modeling across diverse ECG datasets and tasks, (2) robust biometrics invariant to physiological state (rest, exercise), and (3) fetal ECG extraction from maternal abdominal recordings. These tasks share the common theme of extracting, representing, and comparing ECG signals under challenging real-world variability, hence the "x" denoting "cross-state," "cross-domain," or "extraction." The following sections synthesize the primary research frontiers and technical methodologies in xECG, with direct reference to recent benchmarking, model design, and application developments.

1. Foundation Modeling: xECG as a Baseline for Universal ECG Representation

xECG is established as a state-of-the-art foundation model for comprehensive ECG representation learning (Lunelli et al., 12 Sep 2025). It is constructed as a deep bidirectional stack of nine xLSTM layers (combining sLSTM and mLSTM blocks) with an embedding dimension of 256 per layer. The architecture introduces exponential gating and multi-head matrix memories at each layer to enhance temporal pattern retention over long sequences.

Pretraining leverages SimDINOv2 self-supervised learning, which combines patch-level masked reconstruction, view-based contrastive losses, and a coding-rate regularizer to maximize the diversity and informativeness of the learned ECG embeddings. The model is trained on an aggregated corpus of ∼2.4M ECGs, spanning diverse clinical populations and durations, with input preprocessing reduced to minimal resampling and cleaning (all signals are resampled to 100 Hz, no further normalization).

Key features of xECG's learning objective:

• Patch reconstruction loss: enforces local correspondence between masked patches of teacher and student embeddings.
• Multi-view consistency loss: promotes alignment across global and local views sampled from patient ECGs.
• Coding-rate regularization: encourages spread in the embedding space for improved downstream separability.

2. BenchECG: Standardized Evaluation and Tasks

BenchECG (Lunelli et al., 12 Sep 2025) is a standardized benchmark that enables rigorous comparison of foundation models across seven held-out datasets (421,000 patients, 1.67M recordings) and ten tasks. These cover classification (PTB-XL, CPSC2018, MIMIC-IV-ECG), segmentation (Sleep Apnea), detection (R-peak on ambulatory and exercise ECG), regression (age estimation), and survival analysis (5-year mortality).

xECG's performance establishes new baselines on this benchmark:

Task/Dataset	xECG FT / Lin-Probe	Key Metric
PTB-XL (classification)	0.928 / 0.915	AUROC
CPSC2018 (classification)	0.981 / 0.968	AUROC
MIT-BIH arrhythmia (F1)	0.677 / 0.674	F1-score
R-peak MIT-BIH	0.995 / 0.945	F1 @20ms
R-peak Exercise-ECG	0.996 / 0.948	F1 @20ms
Sleep Apnea	0.932 / 0.856	AUROC
DeepBeat PPG AF	0.887 / 0.751	AUROC
Blood-test (MIMIC-IV-ECG)	0.747 / 0.733	AUROC
Age estimation	0.072 / 0.081	SMAPE
5y Mortality (CODE→MIMIC-IV)	0.710 / 0.703	C-index

All values reflect strong cross-dataset and cross-task generalization, with xECG outperforming previous transformer (ViT) or large-scale supervised CNN baselines especially on long-context and frozen-encoder evaluations.

3. Cross-State ECG Biometrics: Robust Authentication Across Physiological Variability

“xECG” also denotes cross-state biometric identification, where authentication must be invariant to significant physiological changes such as from rest to post-exercise (Zheng et al., 20 Oct 2025). CrossStateECG establishes a state-of-the-art method for this scenario:

• Multi-scale convolutional architecture: Four parallel 1D-conv branches with kernel sizes {3,5,7,11} extract local and global cardiac morphology.
• Deep convolutional block: Stacked Conv1D + BatchNorm + ReLU layers deepen representation.
• Self-attention enhancement: Projects features into queries, keys, and values via 1x1 convolutions, computes temporal attention, and outputs a scaled residual.
• Embedding head: Global pooling, linear projection to 128 dimensions, L2 normalization.

Loss is a combination of focal loss (to mitigate class imbalance) and an improved Multi-Similarity metric learning loss that ensures intra-class compactness and inter-class separability even across state shifts.

Quantitative results on the exercise-ECGID dataset:

Scenario	ACC (%)	FAR (%)	FRR (%)	AUC-ROC (%)
Rest2Rest	99.94	0.00	0.12	99.90
Exercise2Exercise	99.86	0.00	0.28	99.80
Mix2Mix	97.85	0.27	4.56	98.85
Rest2Exercise	92.50	1.61	13.33	98.72
Exercise2Rest	94.72	8.95	1.61	99.46

Compared to earlier methods (Saleh et al. 86.45%, Wang et al. 61.40% on Rest2Exercise), CrossStateECG achieves substantially higher cross-state identification accuracy while maintaining near-perfect performance in same-state scenarios (Zheng et al., 20 Oct 2025).

4. Fetal ECG Extraction: Source Separation from Maternal Recordings

In a different context, xECG refers to the extraction of fetal ECG from abdominal maternal recordings where the observed signal is a superposition of maternal ECG, fetal ECG, and noise (Wang, 2014). The canonical methodology involves:

Signal model: $x(t) = m(t) + b(t) + n(t)$

• Chest channels ($c_i$) predominantly sense $m(t)$ (maternal ECG).
• Abdominal channels ($a_j$) contain both $m(t)$ (attenuated) and $b(t)$ (fetal ECG).

Estimation uses adaptive linear filtering:

• Wiener filter: Batch approach assuming stationary statistics, solving for $w$ to minimize the MSE between filtered chest signals and the abdominal signal.
• LMS filter (adaptive): For non-stationary heart rate/morphology, real-time adaptation of $w$ via $w(t+1) = w(t) + \mu e(t) x(t)$, where $e(t)$ is the instantaneous prediction error, yielding fetal ECG as $\hat{b}(t) = AS(t) - y_m(t)$.

Implementation specifics:

• Filter length $L=1$ or $2$ taps—greater $L$ only if more complex delays/morphologies need to be modeled.
• Normalized/phase-scheduled learning rates for improved convergence.
• Optionally, maternal template subtraction applied prior to adaptive filtering.
• Cascaded filtering and multi-reference averaging can further improve SNR.
• Standard detection techniques (e.g., Pan-Tompkins) are then used on $\hat{b}(t)$ to obtain fetal heart rate.

Performance metrics:

• RMSE against reference fetal ECG, SNR improvement ($\Delta$SNR), R-peak sensitivity/PPV.
• On representative experimental data: Two-stage LMS (phase I $\mu=1/[3 \mathrm{trace}(R)]$, phase II $M=10^{-5}$, $L=1$) achieves RMSE ≈ 1.26, $\Delta$SNR gain 8–12 dB, fetal-R sensitivity $>95\%$—substantially surpassing baseline filtering or naive subtraction (Wang, 2014).

5. Methodological Insights and Design Principles

• xLSTM-based architectures (foundation modeling) maintain linear computational scaling in sequence length and superior transferability to ECG tasks with long context compared to transformer-based models (Lunelli et al., 12 Sep 2025).
• Multi-scale convolution (biometric modeling) captures physiological variation across time-scales, improving discrimination in rapidly changing conditions such as exercise.
• Self-supervised training objectives that encourage both local and global agreement across diverse ECG "views" enable robust performance even when fine-tuning is minimal.
• Adaptive filtering in source separation leverages the statistical redundancy between chest and abdominal signals, exploiting their joint structure rather than independent channel analysis.

6. Limitations and Open Research Directions

• Pretraining corpus coverage remains a limiting factor for foundation models; comparisons across models are confounded by different training set scales and patient cohorts (Lunelli et al., 12 Sep 2025).
• Benchmark breadth: While BenchECG spans a wide range of tasks, it excludes some real-world scenarios, e.g., >24 h Holter monitoring or neonatal ECGs.
• Cross-state biometrics currently depend on accurate R-peak detection and fixed-length segments; future architectures may eliminate this constraint (Zheng et al., 20 Oct 2025).
• In fetal ECG extraction, extremely low fetal SNR or abrupt heart rate/motion changes still degrade performance; further adaptive techniques (e.g., RLS, ICA, multimodal fusion, adaptive windowing) are suggested for increased robustness (Wang, 2014).
• Prospective clinical trials and regulatory validation are required for clinical adoption—technical superiority in benchmarking does not guarantee safety or effectiveness in operational medical settings (Lunelli et al., 12 Sep 2025).

7. Summary and Significance

xECG now denotes not a single technique, but the intersection of advanced ECG analysis across foundational medical AI, robust physiological biometrics, and physiological signal source separation. Key technical innovations—multi-scale recurrent and convolutional architectures, self-supervised and metric learning objectives, and adaptive filtering—define the methodological state-of-the-art. BenchECG provides a multi-task, cross-dataset yardstick for progress. Ongoing challenges include model scaling, coverage of rare or long-duration cases, and translation to clinical utility across settings. Collectively, these research strands establish xECG as central to next-generation, real-world-ready ECG analysis (Lunelli et al., 12 Sep 2025, Zheng et al., 20 Oct 2025, Wang, 2014).