The creation of the Catalan Association for Artificial Intelligence (ACIA) was driven not only by practical and scientific objectives, but also by a powerful symbolic vision. It showcases the potential of Catalan society to produce innovative ideas and add value on a global scale. This paper illustrates the pivotal role of the ACIA in the development and consolidation of artificial intelligence (AI) research in Catalonia and across Europe. Founded in 1994, ACIA emerged from the need to create a cohesive AI research community in Catalan-speaking territories and to promote AI literacy. Over the past three decades, ACIA has made significant contributions to AI research through initiatives detailed in this paper, such as the International Conference on AI (CCIA), the magazine NODES, the Marc Esteva Vivanco Award for the Best PhD thesis in AI, and the donesIAcat working group. By presenting these initiatives, analyzing the evolution of AI research topics in Catalonia, and detailing ACIA's involvement in Europe-including its role within EurAI and its industrial and international impact-, this article highlights how local AI societies contribute to the advancement of AI in Europe while preserving their unique cultural and academic identities. Finally, the future evolution of AI in Europe is discussed.

Emotion recognition from electroencephalography (EEG) signals is a pivotal component of affective computing with broad applications in human-computer interaction, mental health, and emotion-aware systems. However, the intrinsic complexity and non-stationarity of EEG data pose significant challenges for accurate and robust emotion decoding. In this paper, we propose HEART (Hybrid Embedding for Affective and Riemannian Trajectories), a novel framework that jointly leverages Riemannian geometry-based spatial features and concept drift-inspired temporal segmentation to model the evolving neural dynamics underlying emotional experiences. HEART extracts covariance matrices from EEG segments, projects them into tangent spaces on the manifold of symmetric positive definite matrices, and employs geodesic distance measures to capture temporal transitions driven by data-adaptive segmentation. Furthermore, it integrates continuous Valence-Arousal-Dominance (VAD) regression with categorical emotion classification via a meta-classifier to exploit complementary affective representations. Evaluated on the FACED, SEED-VII and SEED datasets, HEART demonstrates high performance compared to state-of-the-art methods, showing robustness to channel reduction and subject variability. The proposed framework offers a compact, interpretable, and effective solution for real-time EEG-based emotion recognition, advancing the integration of neurophysiological insights and machine learning for affective computing.

Background: Autistic people are more likely to experience mental health difficulties than non-autistic people, yet they are less likely to access formal support. This study explored autistic adults' experiences of poor mental health at three key stages: at onset; at the stage of seeking and obtaining support; and after recovery. Methods: Twelve autistic adults who had experienced, but were no longer experiencing, poor mental health, participated in a semi-structured interview. Their responses were analysed using thematic analysis clustered around the three stages of this journey. Results: In terms of participants' first experience of poor mental health, findings confirm previous research highlighting social isolation as the main trigger for poor mental health, and also as a consequence. Participants also reported having difficulty recognising the symptoms and using self-help strategies. When seeking support participants reported not knowing how to seek support, seeking support from family, having difficulty communicating their need for support or not knowing what support to ask for, and having to ask for support more than once. After recovery, participants felt more confident and experienced improved psychological awareness. However, some participants still felt uncertain of what support they would seek if they experienced poor mental health again. Conclusions: This study identifies the need to develop tailored interventions at key stages in the journey to poor mental health. We propose a multifaceted approach that focuses on the prevention of social isolation, the promotion of mental health literacy for autistic adults and their families, and improving recognition of symptoms and nonverbal communication of poor mental health in families and professionals.

Electroencephalography (EEG) is a widely used non-invasive technique for monitoring brain activity, offering valuable insights into neurological disorders. Feature extraction methods based on signal processing approaches have been shown to be effective, but they tend to overlook the statistical properties of EEG signals. This study proposes a decile-based feature extraction method for EEG signal analysis, aimed at improving classification performance while maintaining simplicity and interpretability. The method was evaluated across multiple tasks, including the classification of Alzheimer’s disease (AD), frontotemporal dementia (FTD), Parkinson’s disease (PD), and seizure detection, using three machine learning models: Random Forest (RF), K-Nearest Neighbors (KNN), and LightGBM. Experimental results demonstrate that the decile-based approach, particularly when paired with RF and KNN, achieves competitive classification accuracy. Furthermore, the proposed method showed robustness to reduced channel counts, suggesting its potential relevance for low-cost, wearable EEG systems. While model performance varied across datasets, particularly for LightGBM, the results indicate that decile-based features provide a useful and interpretable representation for diverse EEG classification tasks. Further studies in larger and more heterogeneous EEG populations are needed to assess generalizability and establish the potential clinical applicability of early diagnosis and real-time monitoring of neurological conditions, especially in resource-constrained or ambulatory settings.

Introduction:Clinical practice guidelines (CPGs) have several limitations, namely: obsolescence, lack of personalization, and insufficient patient participation. These factors may contribute to suboptimal treatment recommendation compliance and poorer clinical outcomes. APPRAISE-RS is an adaptation of the GRADE heuristic designed to generate CPG-like treatment recommendations that are automated, updated, personalized, participatory, and explanatory using a symbolic AI approach. TDApp is a clinical decision support system (CDSS) that implements APPRAISE-RS for ADHD. Methods:Two clinical trials were conducted. In both studies a total of 33 and 32 ADHD patients, respectively, requiring treatment initiation or a major treatment change were enrolled. TDApp recommendations were compared to those of selected CPGs (American Academy of Pediatrics, National Institute for Health and Care Excellence, Spanish Health System, Canadian ADHD Resource Alliance, and the Australasian ADHD Professionals Association) CPGs. The diversity of treatment recommendations was analyzed using Blau's index. Concordance between TDApp and CPGs recommendations was assessed by calculating the proportion of patients for whom TDApp recommended one drug that was also endorsed by CPGs. Dendrograms were plotted to compare the distance between treatment recommendations as calculated using the NbN nomenclature. Results:The first study investigated eight methods that differed in how patient and clinician preferred outcomes were handled and the extent to which TDApp tailored the analysis of evidence. The method deemed most suitable was examined in the second study, which found that 50-75% of the patients received at least one favorable treatment recommendation. TDApp evaluated over 10 drugs, including recently marketed ones, with amphetamine derivatives emerging as the most frequently recommended interventions. TDApp generated 8-12 distinct treatment recommendations with a diversity index of 0.70-0.88, which was higher than those of CPGs. The proportion of patients for whom TDApp recommendations overlapped with at least one drug endorsed by CPGs ranged from 21.9% to 100%. Dendrogram analysis revealed that TDApp was positioned on one side of the tree, while CPGs clustered together on the opposite side. Conclusions:TDApp is an advanced prototype of an CDSS offering automated, participatory, personalized, and explanatory treatment recommendations for ADHD. It represents a promising alternative to CPGs for aiding clinicians and patients in shared treatment decision-making.

This work introduces a novel methodology for Electroencephalography (EEG) data analysis in the context of neurological diseases, emphasizing feature extraction through covariance matrices and their integration with Case-Based Reasoning (CBR). Departing from traditional techniques such as Fast Fourier Transform (FFT) and statistical analysis, we investigate the synergy between covariance matrices and CBR, highlighting their potential to improve the efficacy of EEG data analysis over conventional methods like Random Forest (RF) and Support Vector Machine (SVM). Covariance matrices analyze the relationships between channels, indirectly capturing interactions between brain regions, while CBR uses similarities in these relationship patterns across cases to make decisions, both techniques focusing on understanding the data through its interrelationships. Additionally, we incorporate Tangent Space Mapping (TSM) to make the covariance matrices more suitable for traditional classifiers by projecting them into a space that preserves their geometric properties. Empirical results on public EEG datasets show that CBR, using covariance matrices with TSM, achieves the best accuracy of 0.72 for Alzheimer's Disease (AD) and up to 0.83 for Parkinson's Disease (PD).

Organ donation is a multifaceted process involving the donor's family, medical professionals, and significant financial costs, all aimed at dramatically improving the recipient's quality of life. However, if an organ is deemed unsuitable for transplantation, the entire process may become ineffective. Assessing the viability of an organ prior to extraction is a critical step in ensuring the success of the transplant. This study explores the potential of AI models to predict organ suitability for donation based on the donor's clinical records prior to extraction. Using data from over 2,700 donors between 2015 and 2023 from medical centers within the Catalan Organization of Transplants (OCATT), we compared the performance of several machine learning algorithms, including Decision Trees (DT), Random Forest (RF), Gradient Boosting (GB), and Support Vector Machines (SVM). The RF model exhibited the highest accuracy in predicting organ viability, aligning closely with expert medical assessments, especially for kidneys and livers. Additionally, we analyzed the key features the model relies on to predict organ viability, providing deeper insights into organ health through medical records. Our findings lay the groundwork for developing a data-driven tool that could assist clinicians in evaluating organ viability more effectively, ultimately improving the transplant process.

Clinical practice guidelines (CPGs) are essential tools that facilitate the translation of the growing body of scientific evidence into clinical practice by providing clinicians with evidence-based recommendations. The first step of CPG development is the formulation of a clinical question involving an intervention of interest. For some interventions, the quantity and quality of the available scientific evidence can vary. This can significantly impact the treatment recommendations. In this work, we present a method for formulating clinical questions involving pharmacological interventions by considering groups of drugs with shared characteristics. This work focuses on drug grouping based on the treatment outcomes desired by both patient and clinician in addition to pharmacological features. To that end, a new method has been presented to learn distances among drugs that is personalized by considering the preferences of users, and an ensemble clustering method is designed to identify the most suitable grouping for each query. We demonstrate our approach in the context of attention deficit hyperactivity disorder (ADHD). Results demonstrate the feasibility of the approach.

Continuous-time Bayesian networks (CTBNs) are powerful tools for modelling and predicting complex disease trajectories in continuous-time scenarios. However, their application is often limited by a lack of individualisation in the results, if the covariates significantly influence a patient's diagnostic transitions. To address these challenges, we introduce the CTBN-PH model, which integrates CTBN models with Cox proportional hazards (Cox-PH) models. The proposed model combines the dynamic and probabilistic capabilities of CTBNs with the robust, covariate-driven risk estimation of Cox-PH models. By leveraging causal topologies learned from healthcare trajectories, the method dynamically adjusts transition intensities based on covariate effects, enabling efficient parameter learning in extensive databases. We validated the model using a dataset of over 2.1 million patients and found that it learned complex causal structures associated with multi-morbid conditions such as diabetes and hypertension. Performance comparisons with non-individualised and non-causally inferred networks highlight the model's effectiveness. Our model achieved an integrated Brier score (IBS) of 0.153 for predicting the onset of a single diagnosis over 25 years and an IBS of 0.04 for forecasting the inertia of the entire system over four years. Additionally, we explore the model's utility in simulating patient trajectories that are tailored to specific covariate-defined populations.

BACKGROUND AND OBJECTIVE:Hybrid forecasting methods aim to overcome the limitations of classical statistical approaches and deep learning models. While statistical methods provide interpretability, they often lack predictive power. Conversely, deep learning models achieve high accuracy but act as "black boxes." This study introduces the Comprehensive Cross-Correlation and Lagged Linear Regression Deep Learning (CCLR-DL) framework, combining statistical and deep learning techniques to enhance both forecasting accuracy and interpretability. Unlike existing hybrid methods that combine statistical filtering with deep learning, CCLR-DL integrates causal statistical selection with neural forecasting, producing interpretable predictors and consistently achieving higher accuracy than models without feature selection or other standard baselines. METHODS:The CCLR-DL framework integrates cross-correlation analysis, lagged multiple linear regression, and Granger causality testing with advanced deep learning architectures. This dual-phase approach first identifies causally significant predictors and then fits them into a deep learning model for multivariate time series forecasting. The framework was validated using a real-world dataset of clinical visits and diagnoses from 6.3 million individuals collected over 10 years. RESULTS:In the evaluated setting, the CCLR-DL framework outperformed baseline models, achieving an average Root Mean Square Error (RMSE) improvement of 19.8% over univariate models, 60.1% over no feature selection, and 51.9% over random selection. The causality phase ensured that all selected predictors demonstrated a significant Granger-causal (GC) relationship. Simpler recurrent architectures, particularly bidirectional Long Short-Term Memory units (BiLSTM), yielded the most accurate forecasts by effectively capturing nonlinear temporal dependencies. CONCLUSIONS:By addressing the challenges of both prediction accuracy and model transparency, the CCLR-DL framework offers a new approach for high-dimensional, multivariate time series forecasting. In healthcare settings, it may enable decision-makers to anticipate demand shifts with greater reliability, allowing earlier staff scheduling, more efficient resource allocation, and reduced waiting times. In our evaluation, it consistently outperformed baseline strategies, delivering measurable improvements that translate into thousands of patient visits being forecasted more accurately across large populations.

Childhood obesity is considered one of the main public health concerns. Research in the field of obesity detection and prevention is moving towards promising solutions thanks to the use of Artificial Intelligence applied to data from cohorts of children. Previous studies have analyzed the data without considering the temporal relationship between them. In this work, sequential pattern mining is used to characterize childhood obesity. Then, original data is represented based on these sequential patterns to feed a case-based reasoning system with the aim to predict childhood obesity. Experiments have been carried out on the data collected from 386 children from Girona and Figueres (Spain).

Childhood obesity is considered one of the main public health concerns. Research in the field of obesity detection and prevention is moving towards promising solutions thanks to the use of Artificial Intelligence applied to data from cohorts of children. Previous studies have analyzed the data without taking into account the relationship of data regarding when they are collected. In this work, frequent pattern mining is used to find the risk factors of childhood obesity, taking into account the relationship among the data gathered in different visits. The experiments carried out on the data collected from 386 children from Girona and Figueres (Spain) demonstrate the relevance of discriminant frequent patterns for childhood overweight prediction.

The rise of wearable EEG devices has opened the opportunity to develop new tools for neurological disorder monitoring, particularly for conditions like epilepsy. Machine learning plays a key role in processing the EEG signal towards an assessment of the person’s state, and eventually evaluating some condition risk. However, existing approaches often rely on raw EEG data, keeping a numerical representation of the information contained in the data. Conversely, in a previous work, we explored representing the EEG signals using sequential patterns. In this work, we analyze the potential of such a representation through several machine learning methods, including decision trees, support vector machines, k-nearest neighbors, and random forest. The experiments carried out with the CHB-MIT scalp EEG database of Physionet show the outperformance of random forest.

This work introduces a novel methodology for Electroencephalography (EEG) data analysis in the context of neurological diseases, emphasizing feature extraction through covariance matrices and their integration with Case-Based Reasoning (CBR). Departing from traditional techniques such as Fast Fourier Transform (FFT) and statistical analysis, we investigate the synergy between covariance matrices and CBR, highlighting their potential to improve the interpretability and efficacy of EEG data analysis over conventional methods like Random Forest (RF). Covariance matrices analyze the relationships between channels, indirectly capturing interactions between brain regions, while CBR uses similarities in these relationship patterns across cases to make decisions, both techniques focusing on understanding the data through its interrelationships. Furthermore, we explore the impact of using data windows in the analysis pipeline to assess their influence on covariance matrix feature extraction and subsequent classification performance in the context of neurological disease diagnosis. Empirical results on public EEG datasets show that CBR, using covariance matrices without temporal windows, achieved the best accuracy, with 0.64 for Alzheimer’s Disease (AD) and up to 0.83 for Parkinson’s Disease (PD), outperforming RF.

Background: Predicting the nocebo response in randomized controlled trials (RCTs) is crucial as it can help minimize its influence and improve the evaluation of the side effects of interventions for ADHD. The aim of this study is to determine the effect of covariates related to study design, intervention, and patients’ characteristics on the nocebo response in patients with Attention Deficit Hyperactivity Disorder (ADHD) using Metaforest, and, ultimately, to investigate Metaforest’s performance in predicting nocebo response in ADHD RCTs. Methods: This study is a secondary analysis of a previously published systematic review [1]. Nocebo response was defined as the proportion of patients experiencing at least one AE while receiving a placebo. We used Metaforest for investigating patient-, intervention, and study design-related nocebo response moderators in ADHD RCTs. Results: One hundred and five studies were included. Overall, 55.4% of patients experienced at least one AE while receiving placebo. However, between-study variability on nocebo response was very high, with nocebo response ranging from 4.2% to 90.2%, leading to high statistical heterogeneity (I2 = 88.3%). Older patients showed a higher nocebo response. The moderating effects of the year of publication, treatment length and gender were also significant. The predictive performance of the model was low-moderate (R2 test = 0, 1922; MSE = 0, 0408). Conclusion: Age was the most important nocebo response modifier, followed by year of publication, treatment length and gender. Metaforest lacked the capability to predict nocebo responses in future studies.