Abstracts Selected For Oral Communication
Presentation number: OC 1
Abstract number: 35-ISABS-2026
ARTIFICIAL INTELLIGENCE IN SKELETAL BIOLOGICAL PROFILING: TOWARD AN INTERPRETABLE AND AUTOMATED FORENSIC ANTHROPOLOGY
Jerković Ivan1, Bašić Željana1, Kružić Ivana1, Anđelinović Šimun2
1University of Split, Faculty of Forensic Sciences, Split, Croatia; 2University Hospital of Split, Split, Croatia
zbasic@unist.hr
The study aims to demonstrate a unified AI strategy for skeletal biological profiling across multiple anatomic elements and modalities. It addresses gaps in previous approaches, focuses on model performance and interpretability in real-world forensic applications, and bypasses the lack of documented skeletal collections in Europe. The main focus is biological profiling of the modern Croatian population using models developed from clinical MSCT images across 1,500+ specimens, including skulls, mandibles, hyoids, clavicles, and sterna. Deep learning segmentation enables automated extraction and organization of skeletal elements into structured digital datasets that mimic documented skeletal collections. AI models are trained on 2D skeletal images and 3D skeletal models, using convolutional neural networks to capture shape, size, and surface texture, while visualization tools (e.g., Grad-CAM activation maps, saliency maps) are systematically evaluated against domain expertise to confirm model reliance on anatomically meaningful regions. Sex estimation using fine-tuned VGG-16 and ResNet50V2 on 3D skull renders achieved accuracies of 93.5% and 92.5%, respectively (n=233). An adapted PointNet++ network applied to 3D hyoid point clouds achieved 88.7% test-set accuracy (MCC=0.77, n=202); mandibular models reached 92% test-set accuracy (n=254). Age estimation from medial clavicle images correctly classified the forensically critical 30-year threshold in 82.5–92.5% of cases. The models are deployed as open- access web applications for external validation and direct practitioner use. Ongoing work compares MSCT-derived models with smartphone photogrammetry and LiDAR-based alternatives to support field deployment without specialized equipment, with direct implications for forensic identification contexts where triage is required before considering confirmatory genetic evidence.
Keywords: artificial intelligence, forensic anthropology, MSCT, three-dimensional imaging, skeletal biological profiling
Presentation number: OC 2
Abstract number: 83-ISABS-2026
PREPARING FOR PRECISION MEDICINE AI DEPLOYMENT: A FUNDAMENTAL RIGHTS IMPACT ASSESSMENT FRAMEWORK FROM A CROATIAN GENERAL HOSPITAL
Bekić Marijo1, Vazdar Tomislav2, Valjalo Baldo1, Bekić Antun3
1General Hospital Dubrovnik, Dubrovnik, Croatia; 2Riskoria Advising & Professional Services d.o.o., Zagreb, Croatia; 3School of Medicine, Catholic University of Croatia, Zagreb, Croatia
marijob@bolnica-du.hr
Hospitals preparing to deploy AI-powered diagnostic imaging, clinical triage algorithms, and pharmacogenomic decision support – the precision medicine tools at the centre of this conference – face a regulatory prerequisite that remains largely unaddressed: the Fundamental Rights Impact Assessment (FRIA). Article 27 of the EU Artificial Intelligence Act (Regulation 2024/1689) requires every public hospital deploying high-risk clinical AI to assess its impact on patients’ right to health, non – discrimination, human dignity, and equitable access to care before first use, with compliance required by August 2, 2026. No operational methodology for conducting FRIA in hospital settings has been published. We developed a six-phase framework at General Hospital Dubrovnik (308 beds, 18 departments), classified as an essential entity under the Croatian Cybersecurity Act (NIS2 transposition). The framework addresses: (1) regulatory classification of AI systems under Annex III, (2) identification of affected patient populations and vulnerability factors, (3) mapping fundamental rights at risk per clinical AI category, (4) algorithmic bias assessment against hospital-specific demographics, (5) definition of human oversight mechanisms preserving clinician decisional authority, and (6) documentation and reassessment governance. By mapping Article 27 requirements against existing institutional roles (CISO, DPO, clinical department heads, Cybersecurity Board), we demonstrated that 78% of FRIA operational elements can be fulfilled through governance structures already mandated by NIS2 and GDPR, without establishing a dedicated AI ethics committee. This pre-deployment framework offers a replicable model for secondary hospitals across Europe seeking to move precision medicine AI from research validation to compliant clinical deployment.
Keywords: fundamental rights impact assessment, EU AI Act, precision medicine, clinical AI governance, hospital cybersecurity
Presentation number: OC 3
Abstract number: 10-ISABS-2026
DECODING BREAST CANCER THROUGH GENE EXPRESSION: PRELIMINARY RESULTS FROM THE CANCER TRANSCRIPTOMICS INTELLIGENCE FRAMEWORK (CTIF)
Bulić Luka1,2,3, Brenner Eva1, Brlek Petar1,3,4, Primorac Dragan1,3,5,6,7,8,9,10,11,12
1St. Catherine Specialty Hospital, Zagreb, Croatia; 2Algebra Bernays University, Zagreb, Croatia; 3School of Medicine, Josip Juraj Strossmayer University of Osijek, Osijek, Croatia; 4Department of Molecular Biology, Faculty of Science, University of Zagreb, Zagreb, Croatia; 5Faculty of Dental Medicine and Health, Josip Juraj Strossmayer University of Osijek, Osijek, Croatia; 6Eberly College of Science, The Pennsylvania State University, State College, PA, United States of America; 7School of Medicine, University of Split, Split, Croatia; 8The Henry C. Lee College of Criminal Justice and Forensic Sciences, University of New Haven, New Haven, CT, United States of America; 9Sana Kliniken Oberfranken, Coburg, Germany; 10School of Medicine, University of Rijeka, Rijeka, Croatia; 11School of Medicine, University of Pittsburgh, Pittsburgh, PA, United States of America; 12National Forensic Sciences University, Gandhinagar, India
luka.bulic0302@gmail.com
Breast cancer (BC) is a clinically and biologically heterogeneous disease characterized by marked variability in prognosis, treatment response, and molecular subtype. Hormone receptor status, including estrogen receptors (ER) and progesterone receptors (PR), as well as HER2 expression, play a central role in therapeutic decision-making and prognostics. Transcriptomic tumor profiling provides a comprehensive view of gene expression patterns, enabling deeper molecular characterization. In this study, machine learning (ML) was applied to high-dimensional transcriptomic data to predict BC clinical outcomes and immunophenotype, thereby advancing precision oncology and individualized risk stratification. This in silico study was conducted on transcriptomic BC data from TCGA, including 1215 samples and 20532 features. Based on transcriptomic features, unsupervised ML (UMAP + HDBSCAN) classified the samples into three distinct clusters. Supervised ML predictive modeling was conducted on raw transcriptomic features using the XGBoost algorithm, as well as on the three derived clusters using logistic regression (LR). Where clinical outcomes were concerned, XGBoost and LR achieved high accuracy scores for 5-year overall survival (88.1% vs. 89.3%), 5-year disease-specific survival (93.7% vs. 93.7%), 5-year disease-free interval (92.3% vs. 92.3%), and 5-year platinum-free interval (88.1% vs. 88.5%). Where hormone receptor status and HER2 expression were concerned, XGBoost and LR achieved high accuracy scores for ER positivity (92.6% vs. 89.6%), PR positivity (85.7% vs. 82.7%) and HER2 positivity (95.5% vs. 85.2%). Both XGBoost-based and LR-based predictive models performed well across all tasks, with XGBoost outperforming LR on immunophenotype prediction, and LR outperforming XGBoost on clinical outcome prediction. The results demonstrate a clear potential for BC clinical outcome and immunophenotype prediction through ML analysis of transcriptomic data.
Keywords: breast cancer, transcriptomic profiling, machine learning, artificial intelligence, outcome prediction
Presentation number: OC 4
Abstract number: 1-ISABS-2026
BEETS HEALTH: FRAMEWORK FOR EXPLORING THE FUTURE OF PREVENTIVE HEALTH USING WEARABLES, AI, AND MULTI-OMICS
Domljanovic Ivana1, Muratovic Tea1, Mileta Dino1,1
1Beets Health FlexCo, Vienna, Austria
ivana@beetshealth.com
Digital biomarkers derived from wearables offer scalable opportunities for preventive health, yet their utility is limited by fragmented interpretation and weak integration with established biomedical reference frameworks. We present Beets Health, an AI-driven multi-omics platform designed to support precision preventive health by integrating 75 physiological, behavioral, cognitive, environmental, and psychosocial variables into a unified and interpretable health profile. The platform aggregates continuous wearable-derived data spanning cardiovascular function, metabolic health, autonomic regulation, sleep, physical activity, cognitive performance, lifestyle behaviors, and environmental exposure. Variables are mapped to age- and sex-adjusted reference ranges and categorized into risk strata based on epidemiological evidence, clinical guidelines, and large population datasets. These features are computationally combined into four resilience domains – Calm, Fuel, Perform, and Recharge – to summarize multidimensional health patterns and generate a composite health score, with an associated confidence metric that reflects data completeness and signal stability. In addition to physiological measures, the platform incorporates psychosocial indicators, including mood, social connectedness, and reflective behaviors, which are increasingly recognized as biologically relevant contributors to cardiometabolic and neurocognitive risk. Beets Health is designed for iterative multi-omics expansion, enabling the future integration of blood-based and epigenetic biomarkers to improve alignment between wearable-derived digital phenotypes and laboratory-based measures. This scalable AI framework aims to support early risk identification, personalized prevention strategies, and longitudinal health monitoring, facilitating the translation of continuous real-world data into clinically meaningful preventive insights.
Keywords: multi-omics, wearable biosensors, blood biomarkers, systems biology, artificial intelligence
Presentation number: OC 5
Abstract number: 68-ISABS-2026
INTEGRATION OF DIETARY AND MULTI-OMICS DATA REVEALS MICROBIOME METABOLISM INTERACTIONS IN ANKYLOSING SPONDYLITIS
Karabekmez Muhammed Erkan1, Yarici Merve1, Torun Cündullah1, Alkaya Güneş2, Kaya Fatoş Nimet2
1Istanbul Medeniyet University, Istanbul, Turkey; 2Göztepe Prof. Dr. Süleyman Yalçın Şehir Hastanesi, Istanbul, Turkey
erkan.karabekmez@medeniyet.edu.tr
Ankylosing Spondylitis (AS) is a chronic inflammatory disease influenced by complex interactions between host genetics, environmental factors, and gut microbiota. While dietary modulation of the microbiome is increasingly recognized, the mechanistic links between diet, microbial metabolism, and immune regulation remain poorly understood. This study aims to leverage AI powered multi-omics integration to uncover microbiome diet host interactions and identify precision nutrition strategies in AS. We integrate metagenomic sequencing, dietary intake data derived from validated food frequency questionnaires, and genome-scale community metabolic models. The data integration frameworks are employed to link microbial composition, metabolic potential, and dietary patterns. Constraint-based modeling approaches, including flux balance analysis (FBA), flux variability analysis (FVA), and flux sampling, are used to simulate microbial metabolic responses to diet. Correlation and machine learning-based analyses are applied to identify associations between nutrients, microbial taxa, metabolic flux distributions, and inflammatory markers. Distinct microbial compositions and metabolic profiles were observed between AS patients and healthy controls. AI driven integration revealed key associations between specific dietary components, microbial metabolic pathways, and inflammation related features. Model based simulations suggested that diet dependent shifts in microbial metabolism may influence immune relevant metabolites, highlighting candidate pathways for intervention. This study demonstrates the potential of AI powered multi-omics integration combined with genome scale metabolic modeling to elucidate complex microbiome diet host interactions in AS. Our findings support the development of precision nutrition strategies targeting microbial metabolism to modulate inflammation. This integrative framework provides a scalable approach.
Keywords: ankylosing spondylitis, genome scale metabolic models, microbiota, precision nutrition
Presentation number: OC 6
Abstract number: 139-ISABS-2026
EXPLOITING THE LINEARITY OF JOINT EMBEDDING SPACES TO MINE CANCER PRECISION MEDICINE KNOWLEDGE
Malod-Dognin Noël1, Pržulj Nataša1,2,3
1School of Digital Public Health, Mohamed bin Zayed University of Artificial Intelligence, Abu Dhabi, United Arab Emirates; 2Department of Computer Science, University College London, London, United Kingdom of Great Britain and Northern Ireland; 3ICREA, Barcelona, Spain
Noel.Malod@mbzuai.ac.ae
Cancer is a leading cause of death worldwide. It is a multifactorial disease that results from multiple alterations, which are only partially captured by any biological data type studied in isolation (e.g., somatic mutation profiles or dysregulated genes). Hence, novel multi-omics data-fusion and analytics methods are needed to holistically mine the wealth of all available clinical and molecular data for new cancer precision medicine discovery. Network embedding is a cornerstone of modelling and analysis of such complex, biological datasets. In the field of NLP, it was observed that the embeddings of words capture semantic relationships linearly, allowing for efficient mining using simple linear vector operations rather than by using computationally intensive, black-box downstream analysis methods. Since then, this observation has been made in other fields, yielding the so-called linear representation hypothesis in vision language models. The question is whether this observation holds and can be exploited in the context of multi-omics data-fusion for cancer precision medicine. To assess this, we defined a Non-Negative Matrix Tri-Factorization based framework to jointly embed the multi-omics datasets of all cancer patients from TCGA (somatic mutation profiles and gene expression profiles of 9,134 cancer patients from 33 TCGA projects), the protein interactions from BioGRID and the drug data from DrugBank into the same embedding spaces. By comparing the performances of linear and non-linear methods for downstream classification and clustering tasks, we demonstrate that our joint embedding spaces are indeed linearly organized. Then, we define linear vector operations that enable prioritizing new pan-cancer or cancer-specific genes and drug repurposings, paving the way to new cancer therapies.
Keywords: AI/ML, multi-omics data-fusion, precision medicine, cancer gene prioritization, drug repurposing
Published: June 16th, 2026.
Copyright: © 2026 Authors of oral communication abstracts. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
