Depressive symptoms in different populations emerge from factors that relate to the lifestyle and background of individuals. For the university student populations, it is important to understand how sociodemographic and lifestyle characteristics influence their mental health and contribute to the development of depressive symptoms in order to come up with effective support and targeted interventions. In this study, we apply unsupervised Machine Learning (ML) to identify depression-related patterns in university students using a set of behavioral and demographic features collected from wearable devices and questionnaires. Using a selected set of features that includes parental education, income, academic activity, and daily behavioral and physiological measures, we clustered student profiles and assessed how these structural factors relate to depressive symptoms measured through the Patient Health Questionnaire-9 (PHQ-9) outcome. Feature importance and ANOVA analyses were used to evaluate cluster formation and interpretability. Our findings suggest that behavioral features, such as walking activity levels and exercise patterns, have a major role in producing depression-related clusters. Moreover, we observed meaningful variability and associations between demographic factors, like parental education and study time, with the depression outcomes. This work contributes to the growing field of fair and explainable ML in mental health and emphasizes the potential of data-driven tools for personalized decision-making for mental health screening in students and younger populations.
Post-cardiac arrest coma, is a consequence of post-cardiac arrest anoxic brain injury (PCABI) that significantly impacts patient survival, with neurological prognosis guiding withdrawal of life-sustaining therapy (WLST). Continuous electroencephalography (EEG) facilitates real-time monitoring of brain activity. However, there is no consensus on which EEG-based patterns may link with recovery or deterioration. In this study, we applied Topological Data Analysis (TDA) to analyze EEG-derived functional connectivity (FC) networks, revealing that patients with poor outcomes exhibit higher network density and efficiency. These findings suggest that a network reorganization may be linked to worse prognoses, highlighting the potential of TDA as a complementary approach for brain network characterization in coma assessment.
This study explores an approach for customizing a large language model (LLM) for use in a medical context through fine-tuning. Structured cardiovascular knowledge and a dataset of cardiac risk cases—specifically, the Cardiovascular Disease Ontology (CVDO) and patient profiles from the Framingham dataset—were used for model adaptation. The objective was to customize a compact natural language model using different types of cardiovascular information to estimate cardiac risk. The results showed that the model’s responses varied depending on the training data, and integration analysis revealed that contextual training reshaped the model’s semantic structure. These findings highlight the model’s adaptability to different domains and the value of combining structured and semi-structured medical data sources to guide its performance.
Endoscopic assessment of the stomach is difficult since visibility is limited, food or gastric pathology may hinder exploration and the passive organ motion is variable. Artificial intelligence (AI) may improve the navigation experience, but it requires depth and pose annotations. Current public datasets do not provide these annotations and they are generated from a small number of synthetic stomach models, thereby restraining the possible scenarios. In this work, a realistic synthetic endoscopy dataset is generated from 3D stomach models reconstructed from real CT-stomachs (30 patients). Gastric wall textures were learned by generative adversarial networks (GANs) trained with mucosa patches extracted from real endoscopic images. A virtual camera and light source were animated to follow a retrograde trajectory within each stomach model, devised to simulate realistic endoscopic navigation. This simulation produced RGB frames, each paired with a depth map and annotated 6-DoF camera pose. An expert evaluation against a reference synthetic dataset [1] showed a preference for the proposed dataset in 80% of cases, with highest scores in the antrum and incisura regions. The dataset was used to fine-tune a depth and pose estimation network, achieving 86.26% depth accuracy and 0.289 cm mean translation error, and released for public access, available at https://doi.org/10.5281/zenodo.17178144.
This study aimed to assess the prognostic value of radiomic features extracted from cardiac magnetic resonance (CMR) perfusion imaging, both alone and in combination with clinical variables, for mortality prediction in patients with coronary artery disease (CAD). Logistic regression and random forest models were trained with class weighting and stratified cross-validation across the three main coronary territories: left anterior descending, right coronary, and left circumflex arteries. Models based solely on radiomic features achieved moderate predictive performance, with AUC values between 0.64 and 0.69. When clinical variables such as leukocyte count, C-reactive protein (CRP), BNP, infarct location, and ejection fraction were included, model performance improved significantly, reaching AUCs between 0.83 and 0.85. These findings demonstrate that integrating radiomic and clinical information enhances predictive accuracy and supports the use of CMR radiomics as a complementary tool for non-invasive prognosis assessment in CAD. Future multicenter studies and deep learning approaches are warranted to validate and generalize these results.
Detecting the QRS complex is fundamental in the analysis of electrocardiographic signals since it allows to evaluate frequency, rhythm, arrhythmias and cardiac variability. This work evaluates three artificial intelligence architectures, LSTM recurrent networks, convolutional networks with residual memory and a hybrid model combining both for detecting and classifying of QRS complexes, in electrocardiograms recorded from Lead II. The paper describes the complete process, from dataset preparation and model training to implementation and interpretation of results.
Craniosynostosis is the early fusion of the cranial sutures and requires corrective surgery. Although computer aided surgical planning can help improve treatment outcomes, planning accuracy is limited due to the lack of methods to predict anatomical changes between diagnostic image acquisition and treatment and by the risk of repeated radiation exposure and/or sedation from longitudinal image acquisition. Recently, we leveraged 3D photogrammetry, a popular radiation-free alternative image modality that provides similar anatomical accuracy to computed tomography, to develop a generative neural network trained on cross-sectional data that enabled personalized prediction of pathological development from CT images or 3D photogrammetry. However, this method suffered from over-conditioning, mode collapse, and reconstruction artifacts. In this work, we present an improved generative network featuring a multi-headed discriminator to enforce conditional independence of age, sex, and pathology, enhanced domain adversarial training to create unbiased latent representations, and utilize bilinear upsampling blocks for higher anatomical fidelity. This model was trained on head surfaces from 744 patients with craniosynostosis and 1,302 normative children (ages 0–5) and evaluated our predictive accuracy on a longitudinal dataset of 76 patients with craniosynostosis and 33 normative subjects. The model outperformed existing methods with head surface growth prediction errors of 2.79 ± 1.01 mm and 2.82 ± 1.22 mm, and volumetric errors 55.9 ± 48.2 mL and 78.7 ± 92.4 mL, in patients with craniosynostosis and normative subjects, respectively. The predictions generated by our model could be used to improve the accuracy of surgical planning.
Identification of early stages of Parkinson’s Disease (PD) is still an important evaluation issue. Assessments and diagnostic scales rely on neurologists’ expertise, observational tests, and associated clinical factors to reach a final diagnosis. This paper introduces a custom spatiotemporal deep learning model to recognize early stages of Parkinson’s disease. This model, a three-block convolutional architecture compresses multi-joint correlated patterns into a single low dimensional embedding which describes lower limb coupling by finding local and global temporal correlations. The method discriminated a series of patients with early PD from a collection of control volunteers. The proposed architecture achieved Accuracy of 81.11%, F1-score of 0.80 and Recall of 0.79, outperforming alternative models, including LSTM, BiGRU, Transformer, and Random Forest.
Skin cancer, affecting over 1.5 million in 2022, poses a significant public health concern. Traditional methods such as visual inspection and dermoscopy can lead to misdiagnosis due to the similarity of the pixels and textures of malignant and benign lesions. Faced with the challenges inherent in visual diagnosis, computer-aided diagnosis (CAD) systems have become increasingly relevant to improving diagnostic accuracy and efficiency. In this context, deep learning models, specifically convolutional neural networks (CNNs), have shown considerable progress in classifying skin lesions through complex feature extraction This study uses the HAM10000 dataset to investigate and compare the performance of different deep learning models, including CNNs, vision transformers and ensemble methods, in classifying skin lesions. The results obtained with an ensemble model of CNNs and vision transformers outperform the results previously obtained in the literature with the same database. An accuracy of 93.61% and an F1-score of 96.04% was obtained.
Obstructive sleep apnea (OSA) and chronic obstructive pulmonary disease (COPD) are common respiratory disorders that often coexist, forming overlap syndrome. OSA involves recurrent upper airway collapse during sleep, while COPD causes persistent airflow limitation. The standard treatment for both OSA and overlap syndrome is continuous positive airway pressure (CPAP). However, its effectiveness is often compromised by poor adherence, largely due to suboptimal pressure regulation in commercial devices, which may result from misclassification of respiratory events. This study developed an offline algorithm for detecting respiratory events, such as apneas and hypopneas, using simulated respiratory signals that included both normal and obstructed breathing patterns. The performance of the algorithm was evaluated by comparing its results with the events embedded in the simulated signals and with those detected by a commercial CPAP device (ResMed AirSense 10). The findings indicate that the proposed algorithm enhances the detection of OSA events and shows potential for integration into automated respiratory monitoring systems, such as CPAP devices or mechanical ventilators. However, clinical implementation requires adaptation for real-time operation and validation with real patient data.
Brain asymmetry, reflecting structural and functional differences between hemispheres, is a fundamental characteristic of brain organization. Aging has been associated with changes in white matter microstructure and asymmetry patterns, but the extent to which these changes occur across the lifespan in healthy populations remains unclear. We analyzed diffusion tensor imaging (DTI) metrics from 412 healthy subjects of the Human Connectome Project - Aging (HCP-Aging) dataset. Asymmetry indices were computed for white matter bundles, and principal component analysis followed by linear modeling, was used to evaluate age-related asymmetry patterns. All four DTI metrics showed age-related changes in brain asymmetry, though with relatively small effect sizes. Distinct patterns of asymmetry were observed across bundles, including both increases and decreases with age. Nonlinear trends in some bundles highlight complex dynamics not fully captured by linear models. Our findings suggest that brain asymmetry is influenced by age, although the effect appears to be small. This could imply that other factors, such as gender and hormonal effects, as reported in the literature may also play a role.Overall, the results presented here contribute to a better understanding of how brain asymmetry evolves across the lifespan and may serve as a reference for future studies exploring the interplay between structural connectivity, aging, or neurological conditions. The establishment of normative patterns of asymmetry across the lifespan could aid in the early identification of structural alterations associated with neurodegenerative or psychiatric disorders.
Gastric intestinal metaplasia (GIM) has been considered as a pre-neoplasic event, yet a very low percentage of patients with these changes actually develops cancer. GIM consists in a series of benign cellular transformations expressing several intestinal phenotypes. Whether or not GIM follows a chronological sequence is still unclear, although visual estimation is the very basis of the Operative Link on Gastric Intestinal Metaplasia (OLGIM) protocol, which reduces GIM quantification to a binary decision. Despite the OLGIM protocol is worldwide applied, it has been reported to underestimate cancer risk. To bridge this gap, this paper presents a continuous approximation to metaplasia in a latent space constructed by an adapted branched variational autoencoder (B-VAE). Both a small encoding architecture and a label-dependent term in the loss function ensure the latent space is as linear and continuous as possible. Distances in this latent space are estimated by a Support Vector Machine with a linear kernel. A total of 2,434 glands from 15 manually annotated cases are used to generate the latent space and train a linear SVM. The linear SVM achieves an F1-Score of 0.85 for the classification task. This representation space demonstrates a scale of quantifiable GIM changes, from healthy glands to the complete and incomplete subtypes of GIM.
This paper presents a new approach for the classification of macular edema types in low-resolution retinal fundus images. The proposed method moves away from traditional spatial-based techniques and instead operates directly on the raw byte-stream representation of the images. An evolutionary framework is employed, utilizing a genetic algorithm to discover and optimize a two-stage binary classifier. Each classifier is composed of an evolved regular expression for byte-level pattern matching, and a subsequent Binary Trainable Decision Function that maps the matched sequence to a final class label. Evaluated on a hand-labeled dataset containing Drusen and Cotton Wool Spot-Type edema, the method achieved a classification accuracy of 0.9535 and an F1 Score of 0.9286. These results significantly outperform a suite of state-of-the-art deep learning models, demonstrating that byte-level feature discovery is a highly effective and competitive strategy for biomedical classification task on low-resolution images.
Deep histological classification holds powerful clinical potential: streamlining clinical evaluation and enhancing care decision-making. However, large and complex images, combined with a relatively small number of patients, make deep learning difficult to leverage effectively in this context. Pre-training and vision transformers have shown promise in overcoming this challenge, but choosing from among these methods is non-trivial. This work evaluates three pre-trained histology models and three vision transformer models toward improved prediction of pediatric Crohn’s disease endoscopic remission from whole-slide images (WSI) of the terminal ileum. Evaluation over these methods reveals the benefits of histology-specific model pre-training, specifically the GigaPath model (median validation F1-score = 0.77, median test F1-score = 0.70). Notably, high and stable model performance was also possible using the largest vision transformer pre-trained on ImageNet-21k, suggesting that a lack of domain-specific pre-training can be overcome by using a larger capacity model. Generally, ensembling tile-level predictions using the mean probability output produced superior predictive performance to global average pooling (GAP).
Accurate breast cancer (BCa) histopathology diagnosis depends on high-resolution imaging, but advanced microscopy is often inaccessible in low-resource settings. Deep learning–based super-resolution (SR) provides a computational alternative, though generic models struggle to preserve critical histological details. This study compares two SR architectures, ESRGAN with a histopathology-trained KimiaNet discriminator and EDSR as a non-adversarial baseline, focusing on diagnostic fidelity and computational efficiency. A dataset of 100 whole-slide images from The Cancer Genome Atlas was processed into 570,000 2048×2048 patches, with low-resolution versions created at ×2, ×4, ×8, and ×16 via bicubic downsampling. ESRGAN+KimiaNet preserved diagnostically relevant features best, reaching PSNR 32.20 dB and SSIM 0.921 at ×2, surpassing EDSR in perceptual quality while keeping inference under 100 ms and VRAM under 200 MB. Recommended use of ×2 for diagnostics (PSNR ¿ 30 dB), 4 for preliminary analysis (PSNR ¿ 25 dB), and ×8–×16 for morphology exploration. Results demonstrate the feasibility of integrating DL-based SR into pathology workflows, offering a scalable, cost-effective solution to enhance diagnostic imaging quality in resource-limited environments, reducing reliance on expensive optical hardware.
Polyp segmentation is a critical task in automated colonoscopy analysis, playing a vital role in the early detection and prevention of colorectal cancer. However, the scarcity of pixel-level annotations in medical imaging limits the scalability of fully supervised methods. We introduce FSTS-Net, a feature-mixing aided semi-supervised teacher-student frame-work designed specifically for polyp segmentation. FSTS-Net employs a U-Net style encoder-decoder architecture and integrates an Exponential Moving Average (EMA) teacher to guide a student model through pseudo-labeling and multi-scale consistency learning. The model performs feature-level semantic mixing in latent space, preserving anatomical structure and contextual integrity crucial for polyp detection. Extensive experiments on benchmark datasets, including CVC-Clinic-DB, Kvasir, and Colon-DB, demonstrate that FSTS-Net consistently surpasses existing semi-supervised methods, particularly in low-label settings (15% labeled data), achieving Dice scores of 0.8636, 0.9308, and 0.8249 respectively. By leveraging unlabeled data effectively and maintaining anatomical fidelity through feature-space mixing, FSTS-Net sets a new benchmark for semi-supervised polyp segmentation, with strong potential for clinical use. The code for our proposed framework is available on https://github.com/Utathyaworks/FSTS-Net.
Preeclampsia (PE) is a significant cause of maternal morbidity and mortality globally. This study aimed to explore the potential of using Electrocardiogram (ECG) and Photoplethys-mogram (PPG) signals combined to assess an automatic detection of post-preeclamptic condition with Detrended Fluctuation Analysis (DFA). We conducted a retrospective case-control pilot study involving 11 post-preeclamptic and 23 healthy pregnant women. Signals were recorded using the BIOPAC MP150 system during induced hypoxia and baseline periods. Several classification models, including Decision Tree (DT), Support Vector Machine (SVM), and Random Forest (RF), were evaluated using repeated stratified k-fold cross-validation. The SVM with Radial Basis Function (RBF) kernel demonstrated the highest performance with an Area Under the Receiver Operating Characteristic (AU-ROC) of 0.730. Key features identified included Hurst exponents from different signal intervals and morphological instability time. This study highlights the potential of using ECG and PPG signals for the classification of post-preeclamptic women.
Lung adenocarcinoma (LUAD) is a highly heterogeneous disease with diverse biological behaviors and clinical outcomes, which complicates diagnosis and treatment strategies. Recent research has demonstrated that LUAD can originate from multiple alveolar cell types, including GRAMD2+ alveolar type I (AT1) cells and SFTPC+ alveolar type II (AT2) cells, each giving rise to tumors with distinct histological features and transcriptomic profiles. Activation of the KRASG12D oncogene in either AT1 or AT2 cells leads to the development of multifocal LUAD, but it is not fully understood how the cell of origin affects tumor characteristics observable through imaging techniques. In this study, we investigated the influence of cell origin on radiomic signatures by comparing LUAD models derived from SFTPC+ AT2 cells (SKTG) and GRAMD2+ AT1 cells (GKTG). Using fast spin echo (FSE) MRI sequences, we identified distinct and cell-of-origin–specific radiomic profiles, with the SKTG tumors showing a greater number of significantly (p < 0.05) different features, indicating unique imaging phenotypes between the two tumor types. Notably, SKTG tumors exhibited increasing radiomic heterogeneity over time, which may correlate with a more aggressive tumor behavior and progression compared to GKTG tumors. These findings underscore the diagnostic and prognostic potential of radiomics to non-invasively differentiate LUAD lesions based on their cellular origin. By capturing tumor heterogeneity and progression longitudinally through imaging, radiomic analysis offers promising avenues for personalized treatment planning and improved patient management tailored to AT1- versus AT2-derived LUAD. This approach may ultimately enhance early detection, therapeutic response monitoring, and clinical outcomes in lung cancer patients.
Chronological brain age prediction using functional connectivity (FC) matrices derived from functional magnetic resonance imaging (fMRI) is an emerging biomarker for assessing neurological health, with deviations from biological age norms suggestive of cognitive decline or other brain disorders. While artificial neural networks (ANNs) outperform traditional age prediction methods by learning hierarchical patterns directly from FC data, most frameworks overlook the biological distinction of within- and between-network connectivity that exhibit divergent trajectories with age. This study addresses this gap by proposing a structured ANN that explicitly models within- and between-network connectivity as separate submodels, aligning their architectural design with modular organization of the brain. Leveraging resting-state fMRI data from 357 healthy adults, FC matrices were partitioned into six brain networks. The dedicated submodels processed within- and between-network connections, with their outputs concatenated prior to age prediction. Model performance was evaluated using mean absolute error (MAE). Grad-Cam was used for interpreting the model findings. Results demonstrated robust predictive performance (MAE was comparable to values in the literature) and revealed that between-network connectivity increased in importance when predicting age in older individuals, while within-network contributions remained stable with age. This finding aligns with prior non-ANN work showing age-related increases in cross-network integration and functional de-differentiation with age. By integrating biologically informed architecture with explainable AI, this work advances personalized brain-age prediction and clarifies network-specific mechanisms. Future directions include validation using longitudinal data and integration with multimodal data to investigate structural-functional coupling in aging.
Identifying alterations in white matter (WM) circuitry in bipolar disorder (BD) may offer crucial insights into the neurobiology of the illness and support biomarker discovery for interventional and genetic studies. Although prior diffusion MRI (dMRI) studies have reported WM abnormalities in BD, few have used fine-scale spatial mapping along tracts (tractometry), and most rely solely on fractional anisotropy (FA), overlooking complementary dMRI metrics related to other microstructural properties. Large-scale multisite tractometry studies remain scarce, limiting power to detect subtle effects and generalizability. Here, we applied BUndle ANalytics (BUAN) to conduct along-tract tractometry of FA, and axial, radial, and mean diffusivity (AD, RD, MD) across 37 major WM tracts in 148 individuals with BD (mean age=36.8 ± 13.5 years; 48% female) and 258 healthy (37.9 ± 13.7 years; 58% female) controls from 6 international sites from the ENIGMA Bipolar Disorder Working Group. Larger samples led to better detection of group-level effects, supporting the benefits of large-scale data harmonization. We identified spatially localized WM microstructural differences in BD across visual, limbic, interhemispheric, language, and sensorimotor pathways including the optic radiation, uncinate and arcuate fasciculi, forceps major of the corpus callosum, extreme capsule, and brainstem tracts. In the largest study to date, our findings demonstrate the value of large-scale, multidimensional tractometry for mapping WM alterations along brain circuits in psychiatric illness.
