Abstract
Pulmonary hypertension (PH) is a hemodynamic condition describing elevated pulmonary artery pressure. To date, right heart catheterism is the gold standard diagnostic test for PH, but it is an invasive and expensive procedure. Deep learning (DL) techniques applied to heart sounds have previously shown promising performances for PH screening. In this work, we analyze the impact of different input representations for PH detection with convolutional neural networks (CNNs). We found that considering each heartbeat as an independent input yielded systematically lower performance than considering the recordings as a whole: preserving the information about the variability over the heartbeats is key. Time-domain feature maps outperformed handcrafted features and combining the time- and frequency-domain proved consistently most effective. Reducing the number of heartbeats to 30 did not affect the performance, and even reducing to 10 beats preserves the diagnostic value. The proposed analysis moves one step further the applicability of DL for PH detection from heart sounds in the clinical practice. © 2025 IEEE Computer Society. All rights reserved.

Abstract
We introduce a multimodal Signal Quality Indicator (SQI) for assessing fidelity of synchronous electrocardiogram (ECG) and phonocardiogram (PCG) signals recorded in ambulatory, non-standardized settings. The method uses a bidirectional fiducial-matching algorithm to test the temporal alignment of QRS complexes and T waves (ECG) with S1 and S2 sounds (PCG) respectively. Validation employed 564 synchronous ECG–PCG pairs collected with the FDA-cleared Rijuven Cardiosleeve at the aortic, pulmonary, tricuspid, and mitral valves sites. Expert annotations served as ground truth. In a three-class task, the SQI reached an area under the ROC curve greater than 79%, showing strong discriminative power. This physiology-based metric supports batch-online monitoring and reliable quality control of opportunistic cardiac recordings.

Abstract
Automated sound-event detection is crucial for large-scale passive acoustic monitoring of wildlife, but the availability of ready-to-use tools is narrow across taxa. Machine learning is currently the state-of-the-art framework for developing sound-event detection tools tailored to specific wildlife calls. Gray wolves (Canis lupus), a species with intricate management necessities, howl spontaneously for long-distance intra- and inter-pack communication, which makes them a prime target for passive acoustic monitoring. Yet, there is currently no pre-trained, open-access tool that allows reliable automated detection of wolf howls in recorded soundscapes. We collected 50 137 h of soundscape data, where we manually labeled 841 unique howling events. We used this dataset to fine-tune VGGish-a convolutional neural network trained for audio classification-effectively retraining it for wolf howl detection. HOWLish correctly classified 77% of the wolf howling examples present on our test set, with a false positive rate of 1.74%; still, precision was low (0.006) granted extreme class imbalance (7124:1). During field tests, HOWLish retrieved 81.3% of the observed howling events while offering a 15-fold reduction in operator time when compared to fully manual detection. This work establishes the baseline for open-access automated wolf howl detection. HOWLish facilitates remote sensing of wild wolf populations, offering new opportunities in non-invasive large-scale monitoring and communication research of wolves. The knowledge gap we addressed here spans across many soniferous taxa, to which our approach also tallies.

Abstract
Accurate segmentation of heart sound signals (phonocardiograms, PCGs) is a critical step for the early diagnosis of cardiovascular diseases (CVDs). Although deep learning models, particularly convolutional neural networks (CNNs) like the UNet, have achieved strong performance in PCG segmentation, the impact of signal preprocessing remains underexplored. In this study, we evaluate how different preprocessing strategies, namely wavelet-based denoising, Butterworth filtering, and their combination, affect the segmentation performance of a 1 D UNet model. Using the PhysioNet 2016 database, we evaluated segmentation quality based on sample accuracy, positive predictive value, and sensitivity. The results show that minimal preprocessing, specifically Butterworth bandpass filtering alone, yields the best segmentation performance, outperforming more aggressive preprocessing pipelines. These findings highlight that preserving the baseline structure of PCG signals is crucial for optimal learning and that lightweight preprocessing remains an essential consideration, especially when applying modern deep learning architectures. © 2025 IEEE.

Abstract
Gastric cancer (GC) remains a significant global health issue, and convolutional neural networks (CNNs) have shown their high potential for detecting precancerous gastrointestinal (GI) conditions on endoscopic images [1] [2]. Despite the need for high resolution to capture the complexity of GI tissue patterns, the impact of endoscopic image resolution on the performance of these models remains underexplored. This study investigates how different image resolutions affect CNNs classification of intestinal metaplasia (IM) using two datasets with different resolutions and imaging modalities. Our results reveal that the often adopted input resolution of 224×224 pixels does not provide optimal performance for detecting IM, even when using transfer learning from networks pre-trained on images with this resolution. Higher resolutions, such as 512×512, consistently outperform 224 × 224, with notable improvements in F1-scores (e.g., InceptionV3: 94.46% at 512 × 512 vs. 91.49% at 224 × 224). Additionally, our findings indicate that model performance is constrained by the original image quality, underscoring the critical importance of maintaining the higher original image resolutions and quality provided by endoscopes during clinical exams, for the purposes of training and testing CNNs for gastric cancer management.Clinical Relevance- This research highlights the importance of image quality, particularly when endoscopes capture lower-resolution images. Understanding how image resolution impacts diagnostic accuracy can guide clinicians in improving imaging techniques and employing Artificial Intelligence-driven tools effectively for more accurate GC detection and better patient outcomes.

Abstract
In this study, we analyzed federated learning (FL) for ECG and PCG data from the PhysioNet 2016 challenge dataset. We tested multiple approaches of FL and evaluated how these approaches affect the performance metrics of cardiac abnormality detection while preserving data privacy. We compared the performance of the centralized and federated models with two and four clients. The results demonstrated that multimodal federated models using both ECG and PCG data consistently outperformed centralized single-modality ECG or PCG models; in fact the gains provided by multimodal approaches can compensate for the loss in performance induced by distributed learning. These findings highlight the potential of multimodal federated learning to not only provide decentralization advantages but also to achieve comparable performance with the centralized single-modality approaches.Clinical relevance- The clinical relevance of this research lies in its potential to improve cardiovascular disease detection by exploring multimodal models and federated learning. It can also help to optimize machine learning models for real-world clinical deployment while preserving patient privacy and achieving comparable performance metrics.