Ainhoa Osa-Sanchez

Background: The global shift towards value-based, preventive healthcare requires diagnostic strategies that are accurate, patient-centered, and resource-efficient. Breast cancer remains [1] the most prevalent cancer in women worldwide, making effective screening a public health priority. While mammography is a cornerstone of early detection, current limitations particularly the high rate of false positives in BI-RADS 0 cases (incomplete category, requiring additional imaging or comparison with prior studies), generate patient anxiety, unnecessary biopsies , and a significant economic burden [2, 3]. Approach: This study employed a quantitative, computational approach using mammography images and clinical data from the publicly available Cancer Imaging Archive (CBIS-DDSM) [4]. The research pipeline included data cleaning, AI model design, implementation of explainability algorithms, and discussion of results with a multidisciplinary team. Data cleaning focuses on selecting only patients with BI-RADS 0, a clinically challenging group characterized by indeterminate findings that frequently require further tests and cause uncertainty. The final dataset comprised 68 patients (54 benign and 14 malignant cases). Radiomic features were extracted and combined with clinical variables. For feature selection and classification, machine learning models including Random Forest, LightGBM, and XGBoost were applied. Model interpretability was prioritized through SHAP and LIME. Results: The key findings indicate that out of 45 features, the most relevant were five clinical variables (breast density, mass shape, mass margins, subtlety, and calc type) alongside 40 radiomic features covering shape, first-order statistics, and texture matrices. Using LightGBM, Random Forest, and XGBoost, models achieved high performance with accuracy 88.9–91.6%, recall 85.3–96.7%, F1 90.1–91.0%, and AUC 96.9–98.0%. Among them, Random Forest delivered the best overall results (91.6% accuracy, 85.3% recall, 91.0% F1, 98.0% AUC). Interpretability with LIME and SHAP highlighted clinically relevant features mass margins, mass shape, and texture descriptors as key for distinguishing. High values of GrayLevelNonUniformity and irregular mass margins supported malignancy, while measures like SmallDependenceLowGrayLevelEmphasis and certain calcification patterns supported benignity. Implications: This study highlights the potential of interpretable AI to enhance preventive breast cancer diagnostics. In a cohort of 68 BI-RADS 0 cases, the model accurately identified all 14 malignancies, highlighting its ability to reduce unnecessary procedures, save healthcare resources, and alleviate patient anxiety. By reliably stratifying indeterminate cases, it improves workflow efficiency, supports timely clinical decision-making, and ensures patient-centered care. Explainability enables clinicians to validate AI recommendations, fostering trust, collaboration, and informed multidisciplinary decision-making. Scaled deployment could advance equitable access, optimize resource allocation, and strengthen value-based, patient-focused healthcare systems. References: 1.World Health Organization. Breast cancer. Geneva: WHO; 2023. Available from: https://www.who.int/news-room/fact-sheets/detail/breast-cancer 2.Xu Z et al. A bimodal nomogram as an adjunct tool to reduce unnecessary breast biopsy following discordant ultrasonic and mammographic BI-RADS assessment. Eur Radiol. 2024;34(4):2608-2618. 3.Li N et al. Value of inversion imaging to diagnosis in differentiating malignant from benign breast masses. BMC Med Imaging. 2023;23(1):206. Published 2023 Dec 8. 4.Sawyer-Lee R et al. Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM) [dataset]. The Cancer Imaging Archive; 2016

Lipid transfer proteins (LTPs) are clinically relevant allergens widely present in plant-based foods, and their reliable detection in complex food matrices remains a major challenge. In this study, we developed an integrated framework combining near-infrared spectroscopy (NIRS), deep learning, and explainability methods to enable accurate and interpretable identification of LTPs. A total of 11,688 spectral measurements were collect ed using a FLAME-NIR spectrometer (940-1700 nm) from homogenized food samples, purified Pru p 3 and Ara h 9 proteins, and their mixtures with LTP-free matrices such as yogurt and powdered milk. Spectral preprocessing involved first derivative transformation, Standard Normal Variate correction, and feature scaling, followed by dimensionality reduction through a 1D convolutional autoencoder, which generated 64-dimensional latent embeddings. These representations were used to train two deep learning classifiers Convolutional Neural Networks (CNNs) and TabTransformer optimized via Bayesian optimization. The inclusion of purified protein embeddings substantially improved classification performance. The CNN model achieved the highest performance with 95.8% accuracy, 97.3% precision, 96.9% F1-score, and an AUC-ROC of 0.954, outperforming the TabTransformer, which nonetheless reached 95.19% accuracy and 96.4% F1-score. Model explainability was addressed using SHAP and LIME, which identified key latent features corresponding to specific spectral regions (940-1700 nm) associated with allergenic signatures. Compared to baseline models, the protein-enhanced framework demonstrated marked improvements in specificity and overall robustness. These results highlight the value of incorporating purified protein information into AI-based spectral analysis, offering a portable, non-destructive, and interpretable strategy for allergen detection in food safety applications.

The automated stratification of retinal diseases remains a significant challenge due to data heterogeneity and the closed-box nature of deep learning models. Although foundational models have demonstrated remarkable success in general computer vision, their clinical reliability and interpretability in multimodal ophthalmology remain insufficiently explored. In this work, we introduce an Explainable Multimodal Foundational AI framework trained on a large-scale integrated corpus of 760,243 retinal images collected from over 15 heterogeneous repositories, encompassing both fundus photography and optical coherence tomography (OCT). We systematically evaluate self-supervised learning (SSL) paradigms DINO and iBOT across convolutional (ResNet) and Transformer-based (Vision Transformer, ViT) architectures. Our results show that ResNet-DINO achieves state-of-the-art performance, reaching 93.53% accuracy and a 0.935 F1-score in 6-class multimodal retinal disease classification, while exhibiting superior robustness under data-limited conditions, attributed to its inductive bias. Notably, we observe emergent clinical localization capabilities in Vision Transformer models (ViT-DINOv2 and ViT-iBOT). Using frozen pre-trained weights and without exposure to expert-labeled data or ground truth labels, these models autonomously highlight clinically relevant biomarkers, including subretinal fluid and drusen, demonstrating intrinsic pathological awareness. By bridging the semantic gap between unsupervised representation learning and targeted clinical diagnosis, this study establishes a benchmark for robust, explainable, and label-efficient AI in ophthalmology. Our findings indicate that large-scale foundational pre-training not only enhances diagnostic accuracy but also induces meaningful visual priors aligned with established clinical biomarkers, supporting the deployment of trustworthy AI systems in real-world clinical decision support.

Food allergies have become a significant public health issue, particularly lipid transfer protein (LTP) allergies, which are highly stable allergens and can cause severe allergic reactions. This research aims to develop and validate an AI-driven solution for detecting LTPs in food using near-infrared spectroscopy (NIRS), exploring the feasibility of non-invasive allergen identification using AI-assisted spectroscopy. The methodology involves collecting spectral data from various food samples, building a machine learning model, and optimizing it iteratively to improve detection accuracy. The results show that the AI model achieved an accuracy of 87% and an F1-score of 89.91%, indicating its potential for enhancing food safety. In conclusion, this solution demonstrates the viability of using NIRS and AI for allergen detection, with promising future applications in healthcare.

Sleep apnea, a prevalent disorder affecting millions of people worldwide, has attracted increasing attention in recent years due to its significant impact on public health and quality of life. The integration of wearable devices and artificial intelligence technologies has revolutionized the treatment and diagnosis of sleep apnea. Leveraging the portability and sensors of wearable devices, coupled with AI algorithms, has enabled real-time monitoring and accurate analysis of sleep patterns, facilitating early detection and personalized interventions for people suffering from sleep apnea. This article presents a systematic review of the current state of the art in identifying the latest artificial intelligence techniques, wearable devices, data types, and preprocessing methods employed in the diagnosis of sleep apnea. Four databases were used and the results before screening report 249 studies published between 2020 and 2024. After screening, 28 studies met the inclusion criteria. This review reveals a trend in recent years where methodologies involving patches, clocks and rings have been increasingly integrated with convolutional neural networks, producing promising results, particularly when combined with transfer learning techniques. We observed that the outcomes of various algorithms and their combinations also rely on the quantity and type of data utilized for training. The findings suggest that employing multiple combinations of different neural networks with convolutional layers contributes to the development of a more precise system for early diagnosis of sleep apnea.

A food allergy is a sensitivity to a food or one of its components that triggers the immune system to react. It affects around 8% of children and 10% of adults. This study looks into the possibilities of near-infrared spectroscopy mixed with artificial intelligence approaches for quick and non-destructive detection of LTP, a severe food allergy. NIRS spectra were acquired from a wide variety of foods, both with and without LTP, using a miniature spectrometer. Three deep learning architectures (convolutional neural networks, vision transformers, and TabTransformer) were employed to classify foods according to the presence of LTP, optimizing hyperparameters using Bayesian optimization. The findings indicated that ViT and CNN-based models had significant potential, with accuracies of more than 90% in both accuracy and F1. Key wavelengths in the 1325-1455 nm range were identified as useful for identifying foods with LTP, indicating that the models account for changes in water, fat, and protein content. Additionally, AI explainability methods (SHAP, LIME, and GradCAM) were used to better comprehend model decisions. The study's findings show the potential of NIRS and AI as useful tools for enhancing food safety through rapid and dependable allergy detection.

This project developed a portable air quality station housed in a 3D-printed enclosure, designed to streamline data sampling and minimize material use in laboratory settings. With health concerns related to specific gases and particulates, especially for vulnerable populations such as asthmatics and children, this innovation has significant potential for improving public health. The importance of indoor ventilation has been underscored by COVID-19, which is primarily transmitted through airborne particles, highlighting the need for efficient monitoring and risk reduction strategies. The station utilizes open-source Python software, with a Raspberry Pi as the core data collection and storage unit, interfacing with various sensors via GPIO, serial, and I2C connections. The modular design of the device allows users to customize measurements and focus on specific pollutants. Validation through end-user testing confirmed the system’s effectiveness and usability in practical settings. The portable setup offers a cost-effective solution for building air quality networks that address the needs of vulnerable groups. The module demonstrated a high reliability rate of 95.30% in detecting common pollutants, validated through CO2 monitoring in classrooms (with a 90.47% reliability compared to commercial devices) and outdoor air quality assessments (with an 85.63% reliability rate.