PO094 | ARTIFICIAL INTELLIGENCE IN PULMONARY EMBOLISM PHENOTYPING: A SCOPING REVIEW

D. Santagata¹, A. Abenante¹, S. Cosci², D. Steiner³, S. Nopp³, A. Squizzato¹, M.P. Donadini¹, F. Dentali¹, C. Ay³, W. Ageno⁴ | ¹Department of Medicine and Surgery, University of Insubria, Varese and Como, Italy; ²Department of Internal Medicine, Ente Ospedaliero Cantonale (EOC), Bellinzona, Switzerland; ³Clinical Division of Haematology and Haemostaseology, Department of Medicine I, Comprehensive Cancer Center Vienna, Medical University of Vienna, Vienna, Austria; ⁴Department of Medicine, University of Padova, Padova, Italy

Background. Pulmonary embolism (PE) phenotyping aims to stratify patients according to clinical, radiological, and prognostic features to guide management. Artificial intelligence (AI) tools are increasingly applied to PE for risk stratification and outcome prediction, yet the scope, methodological quality, and clinical applicability of these approaches remain unclear.
Aims. To systematically map the existing literature on the application of AI in PE phenotyping, describe methodological characteristics and performance metrics, and identify gaps in current evidence.
Methods. A scoping review was conducted in accordance with the JBI methodology and the PRISMA-ScR checklist. Eligible studies included adult patients with imaging-confirmed or administratively coded PE and applied AI methods for phenotyping.
Results. After screening 2,074 records, 22 studies met inclusion criteria. Publications increased markedly in recent years, with 17 of the 22 studies published between 2023 and 2025. Reported sample sizes varied widely, ranging from 70 to 6,853. All studies were retrospective, with most conducted at a single center. Traditional machine learning models were used in 14 studies, deep learning approaches in 3, combined deep learning and traditional machine learning in 2, commercial imaging AI in 2, and quantitative imaging plus traditional machine learning in 1; radiomics-based features were used in 4 studies. Imaging-only inputs were most common, followed by clinical/laboratory data and multimodal models. Mortality prediction was the most frequent clinical task, followed by PE risk stratification or severity classification, composite adverse outcome prediction, functional marker assessment, and clot-burden evaluation. AI-derived model performance was most frequently reported using the area under the receiver operating characteristic curve (AUC), with values ranging from 0.63 to 0.97; 6 of 20 studies with reported AUCs achieved an AUC ≥0.90. Direct comparative evaluation against existing clinical or imaging-based scores was available in 9 studies, corresponding to 10 paired comparisons; AI-based models outperformed the comparator in 9 of 10 comparisons (Figure 1). External or temporal validation was conducted in a minority of studies.
Conclusions. This scoping review identified an arising publication trend in AI-phenotyping related articles. The literature is characterized by retrospective designs, frequent single-center cohorts, and substantial heterogeneity in study populations, data sources, and AI methodologies. Most studies focused on mortality prediction or risk stratification, while external validation and multimodal modeling were less frequently performed. Despite promising AI-model prediction capabilities and generally higher AUCs compared with traditional methods, the evidence reveals important methodological limitations, highlighting key research gaps and the need for standardized reporting to support interpretability and clinical implementation.