Accurate embryo selection is a key factor in improving implantation success rates in Assisted Reproductive Technologies. This study presents a deep learning framework, EmbryoVision AI , designed to enhance blastocyst assessment using Time-Lapse Imaging and eXplainable AI techniques. A customized convolutional neural network was developed to capture both morphological and temporal dynamics, enabling a precise classification of the embryo. To ensure transparency, Gradient-weighted Class Activation Mapping was integrated, allowing visualization of decision-critical embryonic structures and ensuring clinical alignment. The model demonstrated strong predictive performance across different embryo grades, achieving an accuracy of 91.5% for Grade AA, 88.4% for Grade AB, and 79.3% for Grade BC. The AUC-ROC values were 0.95, 0.90, and 0.81 for Grade AA, AB, and BC, respectively, indicating strong discriminatory capabilities. The findings suggest that AI-driven embryo selection can enhance objectivity, reduce human variability, and improve ART outcomes. However, the results also underscore the need to refine AI models to better handle morphological variability in lower-quality embryos, highlighting the importance of improving generalization and strengthening clinical integration. • The study presents a novel deep learning framework integrating time-lapse imaging and explainable AI techniques to enhance blastocyst selection in ART procedures. • The integration of Grad-CAM allows for visual interpretation of model decisions, ensuring clinical trust and alignment with embryological expertise. • The proposed model suggests robust performance in embryo classification.

Abstract Interventions utilizing autobiographical memory (AM) frequently depend on verbal remembrance; nevertheless, their effectiveness may be constrained in the absence of sensory stimuli and emotional reinforcement. Socially Assistive Robots (SARs) provide a multimodal option; yet, there is limited understanding of how robotic contact influences users’ emotional experiences during autobiographical memory recall. We established a pipeline wherein the humanoid robot Pepper performs life-span autobiographical interviews and produces synthetic visuals from real-time speech transcripts. Fifteen adults participated in two counterbalanced conditions during a single session: (i) grasping the robot hand while recounting two memories, and (ii) recounting two memories without tactile interaction. The results show that touch markedly improved affective valence and resulted in a more substantial post-session decline in Negative Affect, mostly due to reductions in evaluations of "nervous" and "Hostile." Arousal and dominance exhibited stability. Participants assessed the system as amiable and intelligent, despite its only mild anthropomorphic qualities. These findings suggest that robotic touch can enhance the enjoyment of robot-mediated memory and specifically alleviate anxiety without modifying the recalled content. The use of a tactile channel enhances the emotional effectiveness of SAR-based AM treatments, establishing a foundation for longitudinal studies including older persons and cognitively at-risk groups.

Abstract Most machine learning models for diabetes prediction rely on small, homogeneous datasets and fixed thresholds, producing binary outputs with limited clinical utility. These approaches lack generalizability, probabilistic awareness, and interpretability, which are essential for real-world healthcare adoption. We present Diabetes Risk via Explainable AI Modeling (D.R.E.A.M.), a framework for Type 2 diabetes mellitus (T2DM) risk prediction that delivers continuous, calibrated probabilities with transparent explanations. D.R.E.A.M. integrates two complementary datasets (PIMA and BRFSS 2015) after excluding gestational diabetes cases, applies clinically guided feature engineering and class balancing, and trains ensemble models (Random Forest, XGBoost, LightGBM). Decision thresholds are optimized using precision–recall curve analysis rather than default cutoffs, enabling clinically meaningful stratification. Model interpretability is achieved through SHapley Additive exPlanations (SHAP), providing both global and patient-level insights. All models achieved Area Under the Curves above 0.83 and F1-scores of 0.78, with Random Forest offering the best balance of sensitivity (recall = 0.89 at an optimized threshold of 0.389) and interpretability. SHAP confirmed the contribution of both physiological and behavioral factors, including glucose, BMI, blood pressure, cholesterol, and physical activity. Accessible via a lightweight web interface, D.R.E.A.M. provides real-time, explainable risk scores to support personalized preventive strategies. In summary, D.R.E.A.M. advances beyond conventional post-hoc explainability by integrating calibrated probabilistic predictions, PRC-based thresholding, and direct clinician-facing deployment. This combination transforms it from a research prototype into a transparent and clinically actionable decision support system.

Abstract Biometric identification based on the electrocardiogram (ECG) is gaining attention as a secure and reliable approach to healthcare authentication, employing the unique physiological patterns detected in the ECG signal. Traditional approaches often depend on raw waveform analysis or the extraction of fiducial points, both of which are computationally intensive and challenging to implement in real-time systems. This work presents CardioIdNet, a lightweight convolutional neural network designed to perform biometric identification directly from ECG images, eliminating the need for complex signal preprocessing steps. ECG recordings from 21 subjects in the MIT-BIH arrhythmia database were segmented and converted to grayscale waveform plots, generating a comprehensive well-suited dataset for image-based deep learning classification. The CardioIdNet architecture consists of convolutional and pooling layers for hierarchical feature extraction, followed by fully connected layers for subject classification. Training was carried out using sparse categorical cross-entropy and the Adam optimizer. The dataset was split 80/20 for training and testing, and early stopping was applied to prevent overfitting and improve generalization. The results show that CardioIdNet achieves excellent performance, with accuracy of 99%, precision, recall, and F1-score of 98.18%, an AUC of 99%, and a false negative rate of 1.85%. CardioIdNet suggests to be a promising solution for biometric authentication in healthcare real-time settings, offering a balance of simplicity, interpretability, and efficiency through image-based deep learning.

The accurate evaluation of periodontal risk during orthodontic tooth movement remains a major clinical challenge, as excessive lower incisor displacement beyond anatomical limits can result in bone damage and gingival recession. This study proposes a novel artificial intelligence framework, FusionNet-ARC, designed to predict alveolar bone availability directly from standard two-dimensional cephalometric radiographs, enabling data-driven assessment of periodontal risk without the need for high-radiation imaging modalities. A total of 996 anonymized lateral cephalometric radiographs were processed through a dedicated image analysis pipeline to extract six anatomical landmarks and compute two alveolar parameters: ARC1 (lingual bone availability) and ARC2 (vestibular bone availability). The proposed FusionNet-ARC model combines a convolutional neural network (ResNet50) for visual feature extraction with a Graph Attention Network (GAT) for landmark-based spatial reasoning. The fused embeddings were regressed to predict ARC1 and ARC2 values, and model accuracy was assessed using Mean Absolute Error (MAE), coefficient of determination (R2), and prediction match rate within a ±0.5,mm tolerance. The model achieved high predictive precision for both parameters, with 66.1 % of ARC1 and 61.4 % of ARC2 predictions within clinical tolerance, showing strong agreement with manual cephalometric measurements. When applied to periodontal risk classification, it accurately identified the corresponding Normal, Borderline, and At-Risk categories, confirming its reliability for automated estimation of alveolar bone availability and related periodontal risk from standard cephalometric radiographs. The proposed system provides an objective and quantitative method for assessing alveolar bone support within routine orthodontic workflows.

Abstract The spread of the COVID-19 pandemic is expected to be uncontrollable by 2020. The main precautions to avoid virus spread have been the introduction of surgical masks or FFP2, sanitization of the hands, and maintaining social distancing. Due to their reliability, molecular tampons are the main detection and prevention methods known as the “Gold Standard”. However, these methods can be particularly uncomfortable. In this case, the analysis of electrocardiogram traces appears to be an alternative method for detecting COVID-19. The dataset used is made up of 1937 images from a study conducted in Pakistan that were preprocessed to train six different neural networks, including MobileNetV2, ResNet-18, ResNet-50, AlexNet, SqueezeNet, and an ad hoc defined neural network. The results show high classification performance, with an accuracy close to 98.94%, as reached by the Resnet-18 network. Moreover, significant attention was devoted to analyzing confusion matrices, revealing the capacity of the networks to identify distinctive features indicative of COVID-19 within ECG data. Finally, it is suggested that in nearly all experiments, including those with low performance, COVID-19 patients are correctly classified, further enhancing the diagnostic potential of ECGs data and DL approach.

Cold spray additive manufacturing (CSAM) is an effective technique for applying metallic layers to various surfaces, particularly beneficial for thermosensitive materials, such as polymers and composites. However, optimizing coating outcomes remains challenging due to several complex factors influencing process efficacy. Machine learning (ML) offers a powerful solution to enhance the quality of CSAM by predicting key coating properties, such as particle penetration depth and flattening. This study addresses the problem of accurately predicting key coating characteristics, specifically particle penetration depth and flattening, by integrating finite element model (FEM) with supervised ML techniques. A dataset of 132 FEM simulations was generated, covering multiple metal–polymer combinations and a wide range of impact velocities. The study evaluates and compares several ML algorithms, including support vector regression, decision trees, Gaussian process regression (GPR), and neural networks (NNs), with the goal of minimizing prediction error measured via root-mean-square error (RMSE). Results show that GPR achieves the best performance for particle flattening (RMSE = 3.9), while a bilayered NN provides the most accurate prediction of penetration depth (RMSE = 2.3). The findings highlight the need for distinct models due to the differing physical mechanisms governing each output: penetration depth exhibits a more linear and predictable relationship with impact velocity and material density, whereas flattening is influenced by complex local deformation and interfacial dynamics. This study demonstrates the feasibility and efficiency of using ML to generalize FEM results, reducing computational cost and enabling fast prediction of coating behavior across varying process conditions.

Cardiovascular disease refers to a group of disorders affecting the heart and blood vessels, including conditions like coronary artery disease, stroke, and heart failure. Arrhythmias are irregularities in the rhythm of the heart, where the heart may beat too fast, too slow, or erratically. This study presents a comparison between ChatGPT-4o and a group of cardiologists in the analysis of electrocardiogram images for assisting in the diagnosis of cardiovascular conditions. The purpose of this comparison is to evaluate the potential of using large language models like ChatGPT-4o in clinical environments, specifically for interpreting electrocardiogram traces. To achieve this, we designed an experiment where both the model and a cohort of cardiologists analyzed the same set of ECG images, and their interpretations were compared to assess performance. The evaluation focused on key diagnostic aspects: heart rate determination, rhythm interpretation, and the overall diagnosis of potential cardiovascular abnormalities. Cardiologists were asked to provide their expert insights through a structured survey that captured their diagnostic reasoning. ChatGPT-4o, in turn, was provided with the same set of images and asked to produce diagnostic outputs. Given that large language models are not explicitly trained in medical image analysis, the responses were generated based on the model’s ability to infer from the textual and visual information presented. The model’s outputs were processed and evaluated for accuracy against the responses of the cardiologists and the ground truth labels provided by the dataset. The results revealed notable differences in diagnostic accuracy between the outputs of ChatGPT-4o and the cardiologists’ assessments. ChatGPT-4o achieved an accuracy of 29.20%, sensitivity of 29.20%, and an F1-score of 0.29 when compared to the ground truth labels. In contrast, the cardiologists collectively performed significantly better, achieving an accuracy of 58.70%, sensitivity of 58.70%, and an F1-score of 0.59.

Diabetes is a chronic metabolic disorder characterized by elevated blood glucose levels due to insufficient insulin production or insulin resistance. It primarily manifests in two forms: Type 1 diabetes, an autoimmune condition typically diagnosed in younger individuals, and Type 2 diabetes, which is more prevalent and often linked to lifestyle factors such as obesity and inactivity. This study evaluates the performance of Long Short-Term Memory networks in diagnosing the two types of diabetes from Italian medical text across four progressively refined pre-processing scenarios. Each scenario incrementally builds on the previous one to enhance text cleaning and data preparation, allowing for a more refined and effective data processing pipeline. In parallel, this study introduces DiabeTalk, a chatbot developed on the WizardLM model, designed to provide specialized advice and support for diabetes diagnosis. While the WE-long short term memory models were fine-tuned with clinical data, DiabeTalk was tested without prior training on clinical diaries, allowing us to evaluate its performance in a real-world context. The results indicate that, despite the lack of domain-specific pre-training, DiabeTalk effectively employs natural language understanding and decision-making algorithms to predict diabetes type and respond to user inquiries. However, the testing revealed limitations in accuracy (77.56% versus 97.80%), with the chatbot achieving a lower performance than the WE-long short term memory model, which was applied to minimally pre-processed raw data. The findings underscore the importance of training large language models on relevant clinical datasets to enhance their response capabilities.

BACKGROUND: Melanoma is one of the deadliest tumors in the world. Early detection is critical for first-line therapy in this tumor pathology and it remains challenging due to the need for histological analysis to ensure correctness in diagnosis. Therefore, multiple computer-aided diagnosis (CAD) systems working on melanoma images were proposed to mitigate the need of a biopsy. However, although the high global accuracy is declared in literature results, the CAD systems for the health fields must focus on the lowest false negative rate (FNR) possible to qualify as a diagnosis support system. The final goal must be to avoid classification type 2 errors to prevent life-threatening situations. Another goal could be to create an easy-to-use system for both physicians and patients. RESULTS: To achieve the minimization of type 2 error, we performed a wide exploratory analysis of the principal convolutional neural network (CNN) architectures published for the multiple image classification problem; we adapted these networks to the melanoma clinical image binary classification problem (MCIBCP). We collected and analyzed performance data to identify the best CNN architecture, in terms of FNR, usable for solving the MCIBCP problem. Then, to provide a starting point for an easy-to-use CAD system, we used a clinical image dataset (MED-NODE) because clinical images are easier to access: they can be taken by a smartphone or other hand-size devices. Despite the lower resolution than dermoscopic images, the results in the literature would suggest that it would be possible to achieve high classification performance by using clinical images. In this work, we used MED-NODE, which consists of 170 clinical images (70 images of melanoma and 100 images of naevi). We optimized the following CNNs for the MCIBCP problem: Alexnet, DenseNet, GoogleNet Inception V3, GoogleNet, MobileNet, ShuffleNet, SqueezeNet, and VGG16. CONCLUSIONS: The results suggest that a CNN built on the VGG or AlexNet structure can ensure the lowest FNR (0.07) and (0.13), respectively. In both cases, discrete global performance is ensured: 73% (accuracy), 82% (sensitivity) and 59% (specificity) for VGG; 89% (accuracy), 87% (sensitivity) and 90% (specificity) for AlexNet.

Premature ventricular contractions (PVCs) are abnormal heartbeats that begin in the lower ventricles or pumping chambers and disrupt the normal heart rhythm. The electrocardiogram (ECG) is the most often used tool for detecting abnormalities in the heart's electrical activity. PVCs are very frequent and usually harmless, but they can be extremely harmful in patients with significant heart problems. As a result, appropriate prevention combined with adequate treatment can improve patients' lives. This paper presents preliminary results on the main challenge associated with the detection of PVCs: identifying common patterns. The images used were extrapolated from the MIT-BIH Arrhythmia Database and then pre-processed to remove any signal noise before creating a distance matrix based on the wave distances of each pair of analyzed images. Finally, we clustered the distance into four groups using clustering algorithms such as K-means. We used a graph-based structure to graphically represent and explore cluster elements in this work. Preliminary results suggest the presence of four distinct patterns.

BACKGROUND: This research aims to increase our knowledge of amyloidoses. These disorders cause incorrect protein folding, affecting protein functionality (on structure). Fibrillar deposits are the basis of some wellknown diseases, such as Alzheimer, Creutzfeldt-Jakob diseases and type II diabetes. For many of these amyloid proteins, the relative precursors are known. Discovering new protein precursors involved in forming amyloid fibril deposits would improve understanding the pathological processes of amyloidoses. RESULTS: A new classifier, called ENTAIL, was developed using over than 4000 molecular descriptors. ENTAIL was based on the Naive Bayes Classifier with Unbounded Support and Gaussian Kernel Type, with an accuracy on the test set of 81.80%, SN of 100%, SP of 63.63% and an MCC of 0.683 on a balanced dataset. CONCLUSIONS: The analysis carried out has demonstrated how, despite the various configurations of the tests, performances are superior in terms of performance on a balanced dataset.

In recent years, the integration of researches in Computer Science and medical fields has made available to the scientific community an enormous amount of data, stored in databases. In this paper, we analyze the data available in the Parkinson's Progression Markers Initiative (PPMI), a comprehensive observational, multi-center study designed to identify progression biomarkers important for better treatments for Parkinson's disease. The data of PPMI participants are collected through a comprehensive battery of tests and assessments including Magnetic Resonance Imaging and DATscan imaging, collection of blood, cerebral spinal fluid, and urine samples, as well as cognitive and motor evaluations. To this aim, we propose a technique to identify a correlation between the biomedical data in the PPMI dataset for verifying the consistency of medical reports formulated during the visits and allow to correctly categorize the various patients. To correlate the information of each patient's medical report, Information Retrieval techniques have been adopted, including the Latent Semantic Analysis technique suitable for constructing a concept space on patient information. Then, patients are grouped and classified into affected or not by using clustering algorithms according to the similarity of medical reports projected in the concept space. Results revealed that the proposed technique reached 95% of effectiveness in the classification of patients.