Histopathological image analysis is the gold standard for cancer diagnosis but is time-consuming, requires a high level of expertise, and is subject to inter-observer variability. The advancement of Digital Pathology enables the application of deep learning models for automating and enhancing diagnostic accuracy. In particular, Convolutional Neural Networks (CNNs) have emerged as a powerful tool for identifying morphological features in histopathological images, achieving accuracy comparable to that of medical experts in specific tasks such as tissue classification. Within the framework of this study, an interpretable CNN model is developed and evaluated for classifying gastric histopathological images as benign or malignant using the GasHisSDB dataset. The Monte Carlo Dropout method is applied to estimate the uncertainty of the model's predictions, while the Gradient-weighted Class Activation Mapping (Grad-CAM) method is combined with quantitative image feature analysis towards enabling the spatial localization of image regions that influence the model's predictions. The proposed approach achieved an AUC score of 98.6% while maintaining low computational cost and architectural simplicity. The generated interpretations yielded useful insights into spatially important image regions and their associated nuclear morphological characteristics.

The COVID-19 pandemic intensified the demand for rapid, accessible diagnostic methods. Machine learning models using cough audio recordings have shown potential for remote COVID-19 detection but often exhibit performance disparities across demographic and clinical subgroups. This study investigates fairness-aware machine learning models for COVID-19 diagnosis using crowdsourced data from the COVID-19 Sounds dataset. A Random Forest and a VGGish-based deep neural network classifier were developed. To mitigate bias, four different methods were applied and comparatively evaluated, namely correlation remover as a pre-processing approach, exponentiated gradient and adversarial debiasing as in-processing approaches, and threshold optimizer as a post-processing approach. Evaluation across sensitive attributes including gender, age, recording device's operating system, and the intersection of age and gender revealed that fairness could be substantially improved with minimal loss in predictive accuracy. The best equalized odds ratio values were$0.956,0.998$, and 0.88 with respect to gender, age, and recording device's operating system, respectively. Similarly, for the combination of gender and age the corresponding metric was 0.804. These findings support the feasibility of fair and reliable audio-based diagnostic systems, emphasizing the importance of integrating fairness into clinical machine learning pipelines.

Carotid atherosclerosis represents a major risk factor for ischemic stroke, requiring accurate risk stratification for effective clinical intervention. While deep learning models demonstrate excellent performance in medical image analysis, their lack of uncertainty quantification limits deployment in clinical environments. This study investigates the use of two uncertainty estimation techniques, Monte Carlo Dropout (MCD) and Deep Ensembles (DE), in cardiovascular risk prediction from B-mode carotid ultrasound images. Using the CUBS dataset for training and two datasets (ATTIKON and BUSI) for external evaluation under distributional shift and out-of-distribution (OOD) conditions, the model's performance, calibration, and robustness are assessed. Results demonstrate that MCD provides superior generalization and OOD detection capabilities ($\text{AUC}=0.9589$), while DE excels in graceful degradation through rejecting high uncertainty samples, achieving 16.49 % accuracy improvement on low uncertainty samples. These findings highlight the complementary strengths of both methods and underscore the critical importance of uncertainty aware AI in clinical decision support systems.

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental disorder characterized by significant phenotypic heterogeneity. Emerging evidence is associating ASD with disruptions in the gut microbiome pointing towards the gut-brain axis as a major contributor to ASD pathophysiology and offering microbial biomarkers with potential as early predictors. In this study supervised machine learning (ML) models trained and evaluated in a nested crossvalidation (NCV) framework are leveraged to classify ASD based on gut microbial profiles from nine different cohorts$(\mathrm{n}=929)$. To address the challenges posed by high dimensionality and biological variability, compositional data augmentation techniques - Aitchison Mixup, Compositional CutMix, and Feature Dropout - were integrated into the modeling pipeline. Among the evaluated ML models, XGBoost with Feature Dropout achieved the best performance (Accuracy: 73.3%, AUC:$\mathbf{8 2. 3 \%}$, F1-score:$\mathbf{7 3. 3 \%}$). To interpret the ML model's predictions, SHAP values and information gain were employed, highlighting key microbial species such as Enterobacter kobei, Dialister hominis, Leyella stercorea, and Comamonas kerstersii. These results reinforce the potential of the gut microbiome as a promising source of ASD screening biomarkers and highlight the utility of interpretable ML models to extract biologically meaningful patterns in complex microbiome datasets.

Serious Games (SGs) are nowadays shifting focus to include procedural content generation (PCG) in the development process as a means of offering personalized and enhanced player experience. However, the development of a framework to assess the impact of PCG techniques when integrated into SGs remains particularly challenging. This study proposes a methodology for automated evaluation of PCG integration in SGs, incorporating deep reinforcement learning (DRL) game testing agents. To validate the proposed framework, a previously introduced SG featuring card game mechanics and incorporating three different versions of PCG for non-player character (NPC) creation, has been deployed. Version 1 features random NPC creation while versions 2 and 3 utilize a genetic algorithm approach. These versions are used to test the impact of different dynamic SG environments on the proposed framework's agents. The obtained results highlight the superiority of the DRL game testing agents trained on Versions 2 and 3 over those trained on Version 1 in terms of win rate (i.e. number of wins per played games) and training time. More specifically, within the execution of a test emulating regular gameplay, both Versions 2 and 3 peaked at a 97% win rate and achieved statistically significant higher (p = 0.009) win rates compared to those achieved in Version 1 that peaked at 94%. Overall results advocate towards the proposed framework's capability to produce meaningful data for the evaluation of procedurally generated content in SGs.

Technological advancements in glucose sensing, insulin pumps, and closed-loop glucose control algorithms open new opportunities towards the realization of Artificial Pancreas (AP). However, the effective management of meal disturbances in these systems still remains a challenge. Meal detection algorithms eliminate the need for meal announcements and enable the shift to more automated and reliable AP systems. The aim of the present study is to develop and evaluate a personalized approach for the detection of meal disturbances in patients with Type 1 Diabetes Mellitus (T1DM). Long Short Term Memory Neural Networks (LSTM)'s inherent ability to efficiently handle sequential data is leveraged within an ensemble learning strategy towards the development of different versions of ensemble models. The models receive as input sequences of Continuous Glucose Monitoring (CGM) measurements (glucose profiles) of a 120-min duration and classify them as positive or negative for the onset of an ingested meal. In silico evaluation is performed using the UVA-PADOVA T1DM Simulator. All ensembles achieve acceptable discriminative performance (mean c-statistic: 75.12%-79.52%) and are able to detect meals in a timely manner (mean detection time: 7.08-12.84 min). Statistical analysis demonstrates the superiority of the simple averaging combination scheme over the other schemes in terms of the c-statistic.

Cardiovascular Disease (CVD) is an important cause of disability and death among individuals with Diabetes Mellitus (DM). International clinical guidelines for the management of Type 2 DM (T2DM) are founded on primary and secondary prevention and favor the evaluation of CVD-related risk factors towards appropriate treatment initiation. CVD risk prediction models can provide valuable tools for optimizing the frequency of medical visits and performing timely preventive and therapeutic interventions against CVD events. The integration of explainability modalities in these models can enhance human understanding on the reasoning process, maximize transparency and embellish trust towards the models' adoption in clinical practice. The aim of the present study is to develop and evaluate an explainable personalized risk prediction model for the fatal or non-fatal CVD incidence in T2DM individuals. An explainable approach based on the eXtreme Gradient Boosting (XGBoost) and the Tree SHAP (SHapley Additive exPlanations) method is deployed for the calculation of the 5-year CVD risk and the generation of individual explanations on the model's decisions. Data from the 5-year follow up of 560 patients with T2DM are used for development and evaluation purposes. The obtained results (AUC=71.13%) indicate the potential of the proposed approach to handle the unbalanced nature of the used dataset, while providing clinically meaningful insights about the model's decision process.

The estimation of long-term diabetes complications risk is essential in the process of medical decision making. Guidelines for the management of Type 2 Diabetes Mellitus (T2DM) advocate calculating the Cardiovascular Disease (CVD) risk to initiate appropriate treatment. The objective of this study is to investigate the use of sophisticated machine learning techniques toward the development of personalized models able to predict the risk of fatal or nonfatal CVD incidence in T2DM patients. The important challenge of handling the unbalanced nature of the available dataset is addressed by applying novel ensemble strategies. Hybrid Wavelet Neural Networks (HWNNs) and Self-Organizing Maps (SOMs) constitute the primary models for building ensembles following a subsampling approach. Different methods for combining the decisions of the primary models are applied and comparatively assessed. Data from the 5-year follow up of 560 patients with T2DM are used for development and evaluation purposes. The highest discrimination performance (Area Under the Curve (AUC): 71.48%) is achieved by taking into account both the HWNN- and SOM- based primary models' outputs. The proposed method is superior to the Binomial Linear Regression (BLR) model justifying the need to apply more sophisticated techniques in order to produce reliable CVD risk scores.

This paper describes a probabilistic causal model for the caring procedure to be followed on wheelchair users with spinal injury. Uncertainty in the caring procedure arises mostly from incomplete information about patient findings (i.e. the signs and symptoms) due to loss of sensation and movement caused by the spinal cord injury. As a result, it may not be easy to assess the extent of a condition -- and, thus, make an accurate diagnosis. Bayesian networks are used for diagnostic reasoning because they offer a way of conducting probabilistic inference about the conditions associated with the caring procedure in the face of uncertainty. The network structure and numerical parameters are based on data elicited from the qualified staff nurses and literature of the National Spinal Injury Centre, Stoke Mandeville Hospital, Aylesbury, UK. We also present the model and report the results of the diagnostic performance tests using the AgenaRisk Bayesian network package.