| Is Part of |
CEUR workshop proceedings: IVUS 2026: proceedings of the 31st international conference on information society and university studies (IVUS 2026), Kaunas, Lithuania, May 12, 2026 / edited by Audrius Lopata,Tomas Krilavičius, Ilona Veitaitė.. Aachen : CEUR-WS. 2026, vol. 4317, p. 1-11.. ISSN 1613-0073 |
| Abstract [eng] |
and multiple treatment options. Many studies report that AI-based clinical decision support systems (CDSS) can help in prostate and urology cancer care, but they commonly rely on well-structured or heavily curated inputs and do not fully exploit free-text documentation in electronic health records (EHRs) [1,2,6,10]. In this paper, baseline experiments are presented using real-world, de-identified EHR data from 23,110 urology oncology patients. The data include structured fields and a large amount of Lithuanian clinical text. Across all experiments, the prediction target is a clinical procedure (target exam/procedure), while the dataset formulations differ in how patient information is represented. Variant A uses a patient-level multilabel formulation based on structured fields and free text. Variants B and C predict the next target procedure from sequential clinical history, with Variant C additionally including examination results. Text was used with basic preprocessing only, without advanced clinical NLP or text structuring. Four classical models were compared: logistic regression, support vector machines (SVM), Random Forest, and XGBoost. Accuracy, macro-F1, and weighted-F1 are reported due to strong class imbalance. Random Forest achieved the best overall performance (up to 0.84 accuracy and 0.82 weighted-F1), while macro-F1 remained below 0.49, indicating weak performance on rare labels. Logistic regression performed consistently poorly, and XGBoost results were inconsistent across formulations. Overall, these baselines highlight that modelling directly from raw Lithuanian free-text EHR data is limited by noise, sparsity, and long-tail imbalance. Future work will focus on lightweight clinical text structuring using an NLP pipeline with entity extraction and normalization, as well as imbalance-aware modelling to improve rare-label coverage and support more reliable decision-support use. |
