Although the problem of predicting atrial fibrillation (AF) has been extensively studied, the electrophysiological changes that precede AF onset remain insufficiently understood. From a clinical perspective, better understanding of such changes should facilitate future monitoring, prediction strategies, and therapeutic decisions. Objective: To investigate the temporal characteristics of P wave morphological variability (PMV) in a 60-min segment preceding the onset of atrial fibrillation (AF). Methods: Ambulatory ECG recordings from 32 patients comprising 176 AF episodes are analyzed. The P wave is emphasized by a linear, spatial transformation of the ECG. Using a novel, robust parameter, the PMV is computed in successive 2-min subsegments of a 60-min pre-AF segment. A piecewise linear model, accounting for basic PMV characteristics, is fitted to the resulting data points, followed by clustering of the model parameter estimates together with the supraventricular ectopic burden. Results: In 63% of the segments, an increase in PMV was found with a change point ranging from 4 to 17 min before AF onset. Clustering demonstrated heterogeneous characteristics manifested by two clusters associated with an increase in PMV, and two other clusters associated with a low and stable PMV of which one had a much larger ectopic burden. Conclusions: Clustering suggests that an AF episode may be preceded by focal ectopic activity, conduction destabilization, or without any observed manifestation. Significance: The proposed framework provides a physiologically interpretable stratification of the pre-AF dynamics of the P wave morphology.
Since fetal ECG (FECG) produces much lower signal amplitudes than maternal ECG (MECG), it is highly susceptible to volume-conduction effects, respiration, and fetal movements. Therefore, the observed abdominal signals exhibit pronounced nonlinear mixing and time-varying characteristics, which severely limit extraction accuracy. To tackle this complex maternal–fetal mixing, this paper pro poses a tensor decomposition of multi-dimensional sub space representation (TD-MDSC) scheme for improving ex traction accuracy and robustness. First, an adaptive multi subspace decoupling (AMSD) mechanism is developed. It employs multi-subspace mapping and hierarchical decomposition to transform the nonlinear mixtures into an approximately linear and more separable representation, thereby facilitating subsequent FECG extraction. Furthermore, a multi-domain tensorized decomposition (MDTD) strategy is developed based on tensor-decomposition theory in multidimensional signal processing. It constructs a third order covariance tensor from the second-order statistics of the observed data and their time-lagged versions to capture multidimensional statistical properties and global spatiotemporal correlations in the observed mixtures. This representation enables effective modeling and separation of the complex maternal–fetal mixtures. Experiments on four public datasets show that the proposed scheme achieves competitive performance compared with several mainstream methods in terms of fetal R-peak detection, indicating robustness under nonlinear and nonstationary mixing conditions.
Objective: Dual-energy contrast-enhanced digital breast tomosynthesis (DECE-DBT) could provide the quantification of iodine contrast with a pseudo-three-dimensional lesion reconstruction, outperforming CE two-dimensional mammography, and hence be a cost-effective alternative to DCE magnetic resonance for breast cancer imaging. However, limited-angle artifacts in DBT reduce its quantitative accuracy. In this work, we propose a deep learning (DL)-based method for DECE-DBT that performs artifact-robust material decomposition for iodine concentration estimation. Methods: Evaluation was performed over 869 digital compressed breast phantoms containing heterogeneous lesions with varying iodine concentrations. The DL network consists of residual blocks, each containing a three-layer CNN followed by a U-Net, and predicts voxel-wise fractions of adipose-blood, fibroglandular-blood and iodine-blood mixtures. Results: Optimization of the number of residual blocks showed that the 2-block model was optimal for the available dataset, generating hallucinations in under 2% of the tested cases. The network performance strongly depends on lesion size: the average dice similarity coefficient was 0.82 for lesions larger than 0.7 cm in diameter. Therefore, early-stage breast tumors (≤ 2 cm) fall mostly within the size range where the network performs reliably. Conclusion: Aimed to shorten the path in making DECE-DBT a functional modality, these findings demonstrate the early potential of our DL-based approach for accurate quantification of iodine concentrations.
Objective: Grouped multivariate variational mode decomposition (GMVMD) is sensitive to the preset group number $G$ and common mode number $K$. To improve its adaptability to multichannel electroencephalogram (EEG) and electrocardiogram (ECG) signals, we determine both automatically. Methods: We propose dual adaptive grouped multivariate variational mode decomposition (DA-GMVMD) with three stages: adaptive grouping, adaptive $K$ search and grouped decomposition. Adaptive grouping constructs a channel graph and determines the channel groups using the Laplacian eigengap and spectral clustering. Adaptive $K$ search uses probe-mode energy shares to select a mode number for each group. Grouped decomposition applies multivariate variational mode decomposition (MVMD) to each group and combines the results. Results: On synthetic signals, DA-GMVMD recovers the underlying grouping structure and remains stable across noise levels. In a separate synthetic test, it recovers the group-specific mode numbers. On the CHB-MIT, MODMA, and PTB-XL datasets, it achieves lower reconstruction error than GMVMD with fixed parameters and several baselines related to MVMD. The reduction relative to GMVMD is significant on CHB-MIT and MODMA. On PTB-XL, DA-GMVMD has a lower observed mean reconstruction error. The reconstructed signals retain seizure classification information on CHB-MIT, regional band power and bilateral asymmetry on MODMA, and R-peak timing and ECG intervals on PTB-XL. Conclusion: DA-GMVMD adapts the channel groups and their mode numbers to the signal content while retaining the evaluated EEG and ECG information in the reconstructed signals. Significance: DA-GMVMD makes multichannel EEG and ECG decomposition less sensitive to a manually preset group number and common mode number.
Objective: While neurosurgical treatment is the primary intervention for Drug-resistant epilepsy (DRE) patients, its success rate is limited to approximately 60%. Precise localization of epileptogenic zone (EZ) is a critical factor for surgical success. This study aims to achieve cross-subject localization of the EZ in DRE patients using causal brain network features extracted from stereotactic electroencephalography (SEEG). Methods: We collected SEEG recordings from 12 DRE patients who achieved complete seizure freedom (Engel Class I). The causal coupling algorithm, full convergent cross mapping (FCCM), was applied to both ictal and interictal states to extract causal brain network features from SEEG recordings. We fed these features into multiple machine learning models to achieve automated identification of EZ within a leave-one-patient-out cross-validation framework. Under a unified evaluation protocol, FCCM was compared with directed information, frequency-domain convergent cross mapping (FD CCM),high-frequencyoscillation-based methods, source-sink con nectivity, and SEEGformer. Results: At the individual level, the Mann–Whitney U test demonstrated that the raw FCCM values sig nificantly differentiated the EZ from the non-epileptogenic zones (NEZ) in 83.3% (10/12) of patients during the ictal state and in all patients (12/12) during the interictal state (p < 0.001). Per mutational multivariate analysis of variance on the 8 dimensional statistical features revealed significant global network variations in 83.3% (10/12) of ictal and 91.7% (11/12) of interictal states (p < 0.001). The combined model, which integrated both ictal and interictal features using the logistic regression classifier, achieved the best performance among all evaluated methods This approach achieved the highest accuracy (0.83 ± 0.10), recall (0.87 ± 0.09), and F1-score (0.80 ± 0.12), while the second-best baseline (FD CCM) yielded an accuracy of 0.79 ± 0.15, a recall of 0.81 ± 0.20, and an F1-score of 0.75±0.18. These results demonstrate that our method achieves state-of-the-art performance. Conclusion: Causal brain network features extracted by FCCM are robust biomarkers for localizing the EZ. Significance: This study could support the development of more personalized treatment strategies for DRE patients. It is expected to promote the development of precision medicine for DRE.
