PO022 | ARTIFICIAL INTELLIGENCE–ASSISTED PREDICTION OF THROMBOTIC AND HEMORRHAGIC COMPLICATIONS IN PATIENTS TREATED WITH VITAMIN K ANTAGONISTS: A RETROSPECTIVE EXPLORATORY STUDY

G. Sottilotta¹, F. Misiti², G. Praticò¹, D. Megalizzi¹, F. Luise³ | ¹UOSD Microcitemie - Emostasi e Trombosi; ³Laboratorio Analisi Grande Ospedale Metropolitano di Reggio Calabria, Italy; ²Associazione Emofilici di Reggio Calabria, Italy

Background and Aims. Vitamin K antagonist (VKA) therapy remains widely used but is burdened by a substantial risk of thrombotic and hemorrhagic complications. Despite the availability of standardized international coagulation parameters, individualized therapeutic ranges, dedicated anticoagulation management software, and specialized anticoagulation clinics, adverse events continue to represent a major clinical challenge. Artificial intelligence (AI) tools, particularly large language models, may offer new opportunities for risk prediction. This study evaluated whether an AI chatbot, when guided by structured prompts, could predict thrombotic or hemorrhagic events in patients treated with VKAs.
Methods. We retrospectively analyzed 17 patients (4 males, 13 females; mean age 77.1 years, range 58–95) receiving VKA therapy who experienced thrombotic or bleeding complications. Sixteen patients were treated with warfarin and one with acenocoumarol. Indications for anticoagulation included atrial fibrillation (n=7), mechanical heart valve (n=4), deep vein thrombosis (n=5), and pulmonary embolism (n=1). A total of 19 adverse events were recorded: 10 hemorrhagic (7 intracranial, 3 gastrointestinal) and 9 thrombotic (5 ischemic strokes, 1 transient ischemic attack, 2 superficial venous thromboses, 1 myocardial infarction). One patient experienced both event types, and one had two thrombotic events. The mean time from therapy initiation to the adverse event was 4017.3 days, longer for thrombotic events (4972.5 days) than for hemorrhagic events (3157.6 days). Relevant comorbidities included erosive gastritis (n=1), renal insufficiency (n=3), and antiphospholipid antibody syndrome (APS) (n=1) among bleeding cases; APS (n=2) and malignancy (n=1) among thrombotic cases. CHATGPT® was instructed to predict the likelihood of thrombotic or hemorrhagic complications in a hypothetical population with identical demographic and clinical characteristics, including indication for anticoagulation, VKA type, and comorbidities, without disclosing that events had already occurred. Event timing was categorized into four intervals (0–5, 5–10, 10–15, >15 years). Prediction accuracy was assessed by comparing the AI assigned risk category—high (≥15%), intermediate (8–14.9%), or low (<8%)—with the actual occurrence of events within the corresponding time interval. These thresholds were derived from 5 year conversions of established bleeding risk models (HAS BLED, ORBIT, warfarin specific models) and applied uniformly to both bleeding and thrombotic outcomes.
Results. CHATGPT predictions demonstrated high or very high concordance in 12 of 19 cases (63.1%), intermediate concordance in 6 cases (31.6%), and poor concordance in only 1 case (5.3%). Overall, predictions were fully correct in 14 events and partially correct in 4, yielding an overall reliability of 94.7%.
Conclusions. This exploratory analysis suggests that AI based chatbots may achieve a high degree of accuracy in predicting thrombotic and hemorrhagic complications in patients treated with VKAs. Although preliminary and limited by the small sample size, these findings support the potential role of AI tools in enhancing individualized risk stratification. Larger prospective studies, including healthy control groups and more diverse clinical profiles, are warranted to refine predictive performance and to integrate AI assisted risk assessment into routine anticoagulation management.