The omnipolar mapping technology (OT) was introduced to overcome the sensitivity of bipolar recordings to catheter orientation and relies on electrodes arranged in regular geometries, such as squares or triangles. Recent studies demonstrated that OT can be applied without the need for specialized catheter geometries. However, whether OT can be effectively applied in sequential mapping without specialized catheter designs remains an open question. In this study, we proposed a variant of OT which could be applied in sequential mapping. A key challenge of this approach is that the electrical field has to be reconstructed from multiple wavefronts recorded across different beats rather than from a single wave as in standard OT, with electrodes arbitrarily positioned. Despite OT was found sufficiently tolerant to the spatial variability of the electrodes, the temporal variability may play an important role. Therefore, to test the efficacy of the algorithm, we investigated the impact of physiological inter-beat variability on the proposed algorithm, with a particular focus on changes in conduction velocity (CV). We performed multiple two-dimensional planar wave simulations with CV values spanning physiological ranges to emulate multiple atrial beats. For each simulation, two bipolar signals were randomly sampled within a circular region of radius 5 mm around a fixed reference point and used to apply OT. Three experimental conditions were considered: i) all bipolar signals were generated from a wavefront with fixed maximum CV, ii) all bipolar signals originated from a single wavefront with a randomly selected CV per simulation, and iii) each bipolar signal originated from a wavefront with an independently selected random CV. Results showed that sequential OT consistently outperformed standard bipolar mapping across all experimental conditions for the characterization of the voltage, exhibiting higher median values (e.g., 5.04 vs 4.62 in Experiment 1 and 4.18 vs 3.68 in Experiment 3). Wavefront direction estimation remained accurate in all cases, with a maximum error of 0, (−3.05,, 3.25)° in Experiment 3. CV estimation of both standard and sequential OT showed a systematic positive bias in Experiment 1 (median 0.99 vs reference 0.9 m/s) and increased variability in Experiments 2 and 3. Sequential OT improved voltage estimation compared with conventional bipolar mapping and enabled reliable wavefront direction assessment. Although the promising results, some limitations persisted, such as a positive bias in CV estimation and incomplete recovery of the reference voltage.

As the leading cause of dementia worldwide, Alzheimer’s Disease (AD) has prompted significant interest in developing Deep Learning (DL) approaches for its classification. However, it currently remains unclear whether these models rely on established biological indicators. This work compares a novel DL model using structural connectivity (namely, BC-GCN-SE adapted from functional connectivity tasks) with an established model using structural magnetic resonance imaging (MRI) scans (namely, ResNet18). Unlike most studies primarily focusing on performance, our work places explainability at the forefront. Specifically, we define a novel Explainable Artificial Intelligence (XAI) metric, based on gradient-weighted class activation mapping. Its aim is quantitatively measuring how effectively these models fare against established AD biomarkers in their decision-making. The XAI assessment was conducted across 132 brain parcels. Results were compared to AD-relevant regions to measure adherence to domain knowledge. Then, differences in explainability patterns between the two models were assessed to explore the insights offered by each piece of data (i.e., MRI vs. connectivity). Classification performance was satisfactory in terms of both the median true positive (ResNet18: 0.817, BC-GCN-SE: 0.703) and true negative rates (ResNet18: 0.816; BC-GCN-SE: 0.738). Statistical tests (p < 0.05) and ranking of the 15% most relevant parcels revealed the involvement of target areas: the medial temporal lobe for ResNet18 and the default mode network for BC-GCN-SE. Additionally, our findings suggest that different imaging modalities provide complementary information to DL models. This lays the foundation for bioengineering advancements in developing more comprehensive and trustworthy DL models, potentially enhancing their applicability as diagnostic support tools for neurodegenerative diseases.

Vectorcardiography (VCG) can evaluate the vector loops of electrocardiographic waves, being a time-spatial representation of the heart vector into the three orthonormal leads. During atrial fibrillation (AF), F waves reflect the disorganized depolarization of the atria, replacing the organized P wave. Usually, paroxysmal AF (PAF) spontaneously terminates, differently from chronic AF (CAF), possibly due to the still-preserved main direction of the P-wave vector loop. To investigate this hypothesis, this study aims to evaluate the similarities between the P-wave vector loop and F-wave vector sway in subjects affected by PAF and CAF. Overall, 10-s VCG were acquired from 10 healthy (HEA) subjects showing normal sinus rhythm, 10 subjects affected by PAF (one during normal sinus rhythm and one during AF), and 10 subjects affected by CAF. P waves were extracted using ECGdeli software, while F waves were extracted after QRST cancellation. Ellipse axes and eccentricities were calculated as the root mean square of VCG components and the ratio between axes, respectively. Overall, 84 beats of HEA, 205 beats of PAF (89 beats during normal sinus rhythm and 116 during fibrillation), and 103 beats of CAF were analyzed. Distributions of axes and eccentricities of PAF are not statistically different (P-value>0.05) than normal sinus rhythm but features related to the Z axis of CAF were statistically lower than PAF (P-value〈10-3). F-wave vector sway in PAF resembles the P-wave vector loop, suggesting the maintenance of the atrial depolarization main direction in subjects with self-terminating AF. Moreover, the F-wave vector sway is more manifest in PAF than in CAF.

Style transfer techniques based on Deep Learning have shown significant promise in biomedical signal processing, particularly in generating synthetic physiological signals from real ones. In this study, we explored the use of Invertible Conditional Generative Adversarial Networks (IcGANs) for style transfer, specifically transforming 12-lead ECG heartbeats from normal sinus rhythm to myocardial infarction (inferior and antero-septal myocardial infarction). Unlike CycleGAN, another style transfer technique which requires multiple models for each class transformation, Ic-Gan only requires training a single conditional GAN and an encoder, offering a more efficient and flexible framework. We trained both IcGAN and CycleGAN on ECG heartbeats extracted from the PTB-XL dataset available on Physionet. We assessed the quality of the generated ECG signals using visual inspection, GAN-train and GAN-test scores, and quantitative metrics such as ST-segment amplitude comparisons. The results showed that the IcGAN effectively captured the relevant features affected by myocardial infarction while preserving the original ECG ones, generating clinically meaningful variations. The comparison between IcGAN and CycleGAN with similar model architectures demonstrated the advantages of the former in terms of efficiency and performance. In conclusion, the potential of IcGAN for controlled ECG feature modification may find applications in domain adaptation, synthetic data generation for rare conditions, and enhancing model generalization for personalized treatment.

Scaffold design for bone tissue engineering along with recent advances in 3D printing represents a major technology to promote the healing of critical bone defects. However, designing the scaffold geometry from a given set of desired properties is challenging and typically tackled by means of high-complexity finite element simulations. Recently, the use of artificial neural networks (ANNs) has started emerging in this field and has provided promising results. In this study, we investigated the development of an ANN to predict the design parameters of a specific geometrical structure from the family of Triply-Periodic Minimal Surfaces, i.e., the gyroid. Unlike other studies, we investigated i) the possibility of using only morphological characteristics of the scaffold to predict the design parameters; ii) the prediction of anisotropic scaffolds since anisotropy has been found to significantly improve bone regeneration; and iii) the comparison of the performance of the ANN with a generalized additive model (GAM) previously designed for the same task. We generated a synthetic dataset of 6940 gyroids where the 90% was used to train the models and 10% for performance evaluation. A feature selection procedure was implemented to select the optimal feature set for the prediction. With the same feature set, the ANN outperformed the GAM in the prediction of all design parameters with Pearson’s correlation coefficients ranging from 0.53 to 0.82, while the GAM’s ranged from 0.41 to 0.51. The ANN also displayed a lower mean absolute error than GAM. The results of the study support the use of ANN for scaffold design. Further evaluations are needed to assess the feasibility of this technology in clinical applications.

Background Detecting subtle patterns of atrial fibrillation (AF) and irregularities in Holter recordings is intricate and unscalable if done manually. Artificial intelligence-based techniques can be beneficial. In fact, with the rapid advancement of AI, deep learning (DL) demonstrated the capability to identify AF from ECGs with significant performance. However, further development and validation on larger cohorts is still needed. Purpose The main purpose of this study was to develop a Residual-attention DL model by considering a large cohort of 2‑lead Holter recordings. Methods We developed a residual DL model by collecting a large dataset of 661 Holter recordings, which was labeled manually by an expert cardiologist. The DL model leveraged attention mechanisms, allowing it to capture long-range dependencies and intricate temporal relationships crucial for identifying subtle patterns indicative of AF. Results Experimental results demonstrated that our model achieved a sensitivity (detection of AF) of Se=0.928 and a specificity of Sp=0.915, with an AUC-ROC of AUC=0.967 on our dataset. Additionally, when evaluated with an external test dataset, specifically IRIDIA-AF, our DL model obtained Se=0.942, Sp=0.932, and AUC=0.965. Finally, when compared under similar experimental conditions with other state-of-the-art models, our DL model achieved slightly better performance overall. Conclusion The Residual-attention DL model we proposed offers a promising solution for AF detection. The validation on external datasets contributes to its potential for deployment in clinical settings, providing clinicians with a valuable decision support system.

Synthetic electrocardiograms (ECGs), obtained with Generative Artificial Intelligence (GenAI), are currently used to support the training of other AI algorithms, most often decision support systems, by augmenting the dataset for the minority classes. In this study, we proposed a Text-to-ECG (T2ECG) framework, which could generate synthetic ECG beats from textual data. The framework made use of two components. The first, Bio_ClinicalBERT, produced an embedded vector from the input text (e.g., "left ventricular hypertrophy"). Then, a second component leveraged such representation to generate a 12-lead ECG heartbeat by means of a Wasserstein Generative Adversarial Network with gradient penalty. The training was performed on the PTB-XL dataset, freely available on Physionet. The framework was designed to generate five different diagnostic classes: i) normal sinus rhythm; ii) inferior myocardial infarction (IMI); iii) antero-septal myocardial infarction (ASMI); iv) left anterior fascicular block (LAFB); and v) left ventricular hypertrophy (LVH). The realism of the generated signals was assessed through three different methodologies, involving both visual inspection and quantitative analyses. Our results show that the T2ECG framework was able to generate heartbeats of sufficient quality, except for the the ASMI class. In conclusion, the framework proposed does not only support data augmentation but also facilitates the creation of ECGs by non-technical users, offering a textual interface to the GenAI model.

Phase-Rectified Signal Averaging (PRSA) computes the average of portions of a time series aligned at a given anchor point. In this study we explored the variance of such samples around the PRSA, or “Phase-Rectified Signal Variance” series (PRSV), and derived its analytical formulation for a stationary Gaussian process. The mathematical prediction was compared with estimates obtained from a set of synthetic time series generated using autoregressive models. The formula, which matched the numerical analysis, reveals interesting insights for the creation of new measures to quantify divergence from Gaussianity and stationarity, in addition to the well known acceleration and deceleration capacities.

Triply-Periodic Minimal Surfaces (TPMS) analytical formulation does not provide a direct correlation between the input parameters (analytical) and the mechanical and morphological properties of the structure. In this work, we created a dataset with more than one thousand TPMS scaffolds for the training of Machine Learning (ML) models able to find such correlation. Finite Element Modeling and image analysis have been used to characterize the scaffolds. In particular, we trained three different ML models, exploring both a linear and non-linear approach, to select the features able to predict the input parameters. Furthermore, the features used for the prediction can be selected in three different modes: i) fully automatic, through a greedy algorithm, ii) arbitrarily, by the user and iii) in a combination of the two above methods: i.e. partially automatic and partially through a user-selection. The latter, coupled with the non-linear ML model, exhibits a median error less than 3% and a determination coefficient higher than 0.89 for each of the selected features, and all of them are accessible during the design phase. This approach has been applied to the design of a hydroxyapatite TPMS scaffolds with prescribed properties obtained from a real trabecular-like hydroxyapatite scaffold. The obtained results demonstrate that the ML model can effectively design a TPMS scaffold with prescribed features on the basis of biomechanical, mechanobiology and technological constraints.

In recent years, subject identification through electrocardiograms (ECGs) broaden the possibilities of existing biometric systems. In this study, we proposed a novel ECG-based biometric identification method designed to be computationally inexpensive, while being sufficiently accurate for a broad variety of application contexts. Specifically, we adapted an established deep learning model known as Deep-ECG to process raw ECG data with minimal preprocessing. We examined the robustness of the model by investigating the identification accuracy across three experiments, obtaining results comparable to more complex state-of-the-art methods. For all experiments, we utilized the SHAREE dataset, containing 24h Holter recordings form 139 subjects, collected in uncontrolled conditions and trained the network by randomly selecting ECG segments during daytime. In the first experiment, we quantified the performance by varying the number of subjects to identify and the number of ECG leads concurrently fed in input. In the second experiment, we varied the number of training samples per individual and the duration of the ECG segments. In the third experiment, we reimplemented the original pipeline of the Deep-ECG model to compare the performance with the new approach with minimal preprocessing. We obtained that the new approach achieved similar performance to the original Deep-ECG model. Also, the new approach obtained accuracies > 80% for individual leads and > 90% for multiple leads when using ECG segments of 2 seconds. Using this ECG duration, the minimal number of training samples per individual to achieve an accuracy > 80% was 100. Our study showed that the computational cost of the Deep-ECG model could significantly be improved by changing the pipeline previously proposed with another one with minimal preprocessing. The source code replicating the results of this study is available on GitHub.

COVID-19 is an infectious disease that has greatly affected worldwide healthcare systems, due to the high number of cases and deaths. As COVID-19 patients may develop cardiac comorbidities that can be potentially fatal, electrocardiographic monitoring can be crucial. This work aims to identify electrocardiographic and vectorcardiographic patterns that may be related to mortality in COVID-19, with the application of the Advanced Repeated Structuring and Learning Procedure (AdvRS&LP). The procedure was applied to data from the “automatic computation of cardiovascular arrhythmic risk from electrocardiographic data of COVID-19 patients” (COVIDSQUARED) project to obtain neural networks (NNs) that, through 254 electrocardiographic and vectorcardiographic features, could discriminate between COVID-19 survivors and deaths. The NNs were validated by a five-fold cross-validation procedure and assessed in terms of the area under the curve (AUC) of the receiver operating characteristic. The features’ contribution to the classification was evaluated through the Local-Interpretable Model-Agnostic Explanations (LIME) algorithm. The obtained NNs properly discriminated between COVID-19 survivors and deaths (AUC = 84.31 ± 2.58% on hold-out testing datasets); the classification was mainly affected by the electrocardiographic-interval-related features, thus suggesting that changes in the duration of cardiac electrical activity might be related to mortality in COVID-19 cases.

The rapid advancement of biomedical sensor technology has revolutionized the field of functional mapping in medicine, offering novel and powerful tools for diagnosis, clinical assessment, and rehabilitation [...]