UK Biobank neck-to-knee body MRI scans of nearly 40 000 participants were automatically analyzed for body composition, age, sex, organ volumes, and abstract properties (eg, grip strength, pulse rat... UK Biobank (UKB) has recruited more than 500 000 volunteers from the United Kingdom, collecting health-related information on genetics, lifestyle, blood biochemistry, and more. Ongoing medical imag...

A quarter of the world population is estimated to have metabolic dysfunction-associated steatotic liver disease. Here, we aim to understand the impact of liver trait-associated genetic variants on fat content and tissue volume across organs and body compartments and on a large set of biomarkers. Genome-wide association analyses were performed on liver fat and liver volume estimated with magnetic resonance imaging in up to 27,243 unrelated European participants from the UK Biobank. Identified variants were assessed for associations with fat fraction and tissue volume in >2 million ‘Imiomics’ image elements in 22,261 individuals and with circulating biomarkers in 310,224 individuals. We confirmed four liver fat and nine liver volume previously reported genetic variants (p values <5 × 10^-8). We further found evidence suggestive of a novel liver volume locus, ADH4, where each additional T allele increased liver volume by 0.05 SD (SE = 0.01, p value = 3.3 × 10^-8). The Imiomics analyses showed that liver fat-increasing variants were specifically associated with fat fraction of the liver tissue (p values <2.8 × 10^-3) and with higher inflammation, liver and renal injury biomarkers, and lower lipid levels. Associations of liver volume variants with fat content, tissue volume, and biomarkers were more heterogeneous, for example the liver volume-increasing alleles at CENPW and PPP1R3B were associated with higher skeletal muscle volumes and were more pronounced in men, whereas the GCKR variant was negatively associated with lower skeletal muscle volumes in women (p values <2.8 × 10^-3). Liver fat-increasing variants were mostly linked to fat fraction of the liver and were positively associated with some adverse metabolic biomarkers and negatively with lipids. In contrast, liver volume-associated variants showed a less consistent pattern across organs and biomarkers. Liver fat and liver volume are common metabolic traits with a strong genetic component, yet the extent to which they exert organ-specific vs. systemic effects remains poorly defined. By integrating genome-wide association analyses and high-resolution neck-to-knee magnetic resonance imaging data through the Imiomics framework, this study reveals distinct genetic architectures for liver fat and liver volume, including sex-specific effects. These findings provide new insights into the biological, organ-level, tissue-specific, and systemic characteristics of steatotic liver disease and its genetic determinants. The results may inform the development of precision imaging genetic approaches, biomarker discovery, and stratified risk assessment strategies, while reinforcing the importance of incorporating sex-specific analyses in future research and clinical applications.

Machine learning applications for the drug discovery pipeline have exponentially increased in the last few years. An example of these applications is the biological Knowledge Graph. These graphs represent biological entities and the relations between them based on existing knowledge. Graph machine learning models such as Graph Neural Networks can be applied on top of knowledge graphs to support the development of novel therapeutics. Nevertheless, Graph Neural Networks present an improved performance at the expense of complexity, becoming difficult to explain their decisions. State-of-the-art explanation algorithms for Graph Neural Networks focus on determining the most relevant subgraphs involved in their decision-making while considering graph elements (nodes and edges) as independent entities and ignoring any communities these graphs could present. We explore in this work the idea that graph community structure in biological Knowledge Graphs could provide a better grasp of the decision-making of Graph Neural Networks. For that purpose, we introduce XP-GNN , a novel explanation technique for Graph Neural Networks in Knowledge Graphs. XP-GNN exploits the communities of nodes or edges in graphs to refine their explanations, inspired by cooperative game theory . We characterize XP-GNN in a basic example and in terms of scalability and stability. In two relevant use cases for the drug discovery pipeline, XP-GNN provides more relevant explanations than previous techniques, being evaluated quantitatively and by domain experts. At the same time, XP-GNN presents limitations on scalability and stability, which we will address. ACM Reference Format: Andrés Martínez Mora, Dimitris Polychronopoulos, Michaël Ughetto, and Sebastian Nilsson. 2024. Community-aware explanations in knowledge graphs with XP-GNN. In Proceedings of ACM Conference (Conference’17) . ACM, New York, NY, USA, 21 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn This work has been funded by AstraZeneca AB, Mölndal, Sweden and AstraZeneca Cambridge. Unfortunately, due to proprietary reasons from AstraZeneca AB, the data used in this work cannot be shared .

BACKGROUND Federated learning (FL) enables multi-institutional model training on clinical text without sharing raw data; however, gradient inversion methods can reconstruct sensitive information from shared model updates. The extent of such privacy leakage in FL applied to radiology reports, and the role of tokenizer design, remains unclear. OBJECTIVE To quantify gradient-based reconstruction of radiology report text in an FL setting and to compare privacy risk across three transformer tokenization strategies in a controlled, tokenizer-aware evaluation. METHODS Six FL clients trained a GPT-2–style transformer (117M parameters; sequence length 32) on two public radiology corpora comprising 368,751 diagnostic reports, 98,206 discharge summaries, and 1,500 MIMIC-CXR free-text reports. Models were trained using three tokenizers (GPT-2, RadBERT, LLaMA-2) with batch sizes of 64, 128, and 256. A curious-server threat model was assumed, and analytic gradient inversion was applied to recover text. Reconstruction fidelity was measured over five runs using exact sentence accuracy, S-BLEU, and ROUGE-L. RESULTS Exact sentence reconstruction ranged from 33% to 42% across tokenizers. At batch size 64, accuracy was 42.1% (GPT-2), 42.3% (RadBERT), and 39.4% (LLaMA-2), decreasing to 37.3%, 37.2%, and 34.3% at batch size 256. S-BLEU scores declined with increasing batch size (e.g., GPT-2: 0.44→0.33; RadBERT: 0.48→0.35; LLaMA-2: 0.39→0.30). RadBERT yielded higher reconstruction fidelity and greater recovery of clinical terms, but no tokenizer prevented leakage. CONCLUSIONS Substantial portions of radiology report text can be reconstructed from FL gradients even with larger batch sizes and domain-specific tokenizers. Tokenizer design influences leakage severity and should be incorporated into privacy evaluations for clinical language models. Integrating safeguards such as secure aggregation and differential privacy is necessary to meet HIPAA and GDPR requirements when deploying FL for radiology NLP. CLINICALTRIAL Not applicable.

Background: Federated learning (FL) enables multi-institutional model training on clinical text without sharing raw data; however, gradient inversion methods can reconstruct sensitive information from shared model updates. The extent of such privacy leakage in FL applied to radiology reports, and the role of tokenizer design, remains unclear. Objective: This study aimed to quantify gradient-based reconstruction of radiology report text in an FL setting and to compare privacy risk across 3 transformer tokenization strategies in a controlled, tokenizer-aware evaluation. Methods: Six FL clients trained a GPT-2–style transformer (sequence length 32) on 2 public clinical-text corpora comprising 368,751 diagnostic reports, 98,206 discharge summaries, and 1500 MIMIC-CXR (Medical Information Mart for Intensive Care Chest X-Ray) radiology reports. Models were trained using 3 tokenizers (GPT-2, RadBERT, and LLaMA-2) with batch sizes of 64, 128, and 256. An active malicious-server threat model was assumed, and analytic gradient inversion was applied to recover text. Reconstruction fidelity was measured over 5 runs using exact sentence accuracy, sentence-level bilingual evaluation understudy (S-BLEU), and recall-oriented understudy for gisting evaluation (ROUGE-L). Results: Exact sentence reconstruction ranged from 27% to 75% across tokenizers, datasets, and batch sizes. At batch size 64 on the discharge dataset, accuracy was 64.7% (GPT-2), 70% (RadBERT), and 67.5% (LLaMA-2), decreasing to 27.3%, 28.5%, and 27.5% at batch size 256. S-BLEU declined with increasing batch size (eg, discharge reports from 0.69 to 0.31). Reconstruction fidelity did not differ significantly across tokenizers (all but one of 27 comparisons nonsignificant; none significant after Holm correction), and approximately 75% of clinical concepts were represented in the reconstructed text (a corpus-level upper bound), regardless of tokenizer. Batch size was the dominant factor governing leakage. Conclusions: Under a worst-case malicious server that tampers with the shared model and observes unprotected per-client gradients (no secure aggregation or differential privacy), substantial portions of radiology-report text can be reconstructed, with up to approximately 75% of reconstructed 32-token sequences and 75% of clinical concepts (not direct patient identifiers) recovered from FL gradients. In a controlled ablation holding model architecture fixed, tokenizer choice, including domain-specific tokenizers, did not significantly affect leakage under the evaluated conditions, whereas batch size was the primary determinant, and no tokenizer significantly reduced the risk. Tokenizer selection should therefore not be treated as a privacy safeguard in this setting. Safeguards such as secure aggregation and differential privacy should therefore be evaluated as candidate protections for FL deployments that must satisfy Health Insurance Portability and Accountability Act (HIPAA) and General Data Protection Regulation (GDPR) requirements in radiology natural language processing (NLP); legal compliance additionally depends on organizational safeguards, risk assessment, and governance beyond the scope of this study.

Background/Objectives: Computer-aided vessel occlusion detection for acute ischemic stroke is typically developed on early-phase CT angiography (CTA). In late-phase CTA, enhanced venous contrast can mimic arterial occlusions and lead to incorrect occlusion detections. To counteract this, we propose a constraint-based post-processing framework that leverages anatomical and temporal vascular information from CT perfusion (CTP) scans. Methods: Vascular anatomy and relative contrast-arrival times were encoded as time–vessel maps. Three complementary constraints worked on these maps to remove incorrect detections based on distance-to-vessel truncation points, contrast-arrival time, and spatial priors. Constraints were fixed on a 32-case late-phase development subset and applied to an external held-out cohort of late-phase CTA (N = 354). To enable application without CTP, an auxiliary CTA-to-time–vessel map generator was designed to estimate maps directly from CTA. Results: In the external cohort, the framework reduced incorrect occlusions per scan from 3.9 (95% CI 3.6–4.2) to 2.1 (95% CI 1.9–2.3), preserving occlusion-level sensitivity at 80% (95% CI 69–91). The patient-level AUROC was 83 for the baseline versus 88 with post-processing. Improvements were also observed in detectors exposed to late-phase CTA, where post-processing reduced incorrect occlusions per scan from 2.8 to 1.9. As only 51 occlusion-positive cases were present in the external cohort, confidence intervals were wide. Consequently, these estimates should be interpreted with caution. Conclusions: Anatomically and temporally informed post-processing reduced incorrect occlusion detections in late-phase CTA without measurable loss of sensitivity, potentially improving robustness against CTA phase changes in external institutions.