A Novel Machine Learning Methodology for the Systematic Extraction of Chronic Kidney Disease Comorbidities from Abstracts

Eszter Saghy ^1* Mostafa Elsharkawy ¹ Frank Moriarty ² Sándor Kovács ¹ István Wittmann ¹ Antal Zemplenyi ¹

• ¹ University of Pécs, Pécs, Hungary
• ² Royal College of Surgeons in Ireland, Dublin, County Dublin, Ireland

Background: Chronic Kidney Disease (CKD) is a global health concern and is frequently underdiagnosed due to its subtle initial symptoms, contributing to increasing morbidity and mortality. A comprehensive understanding of CKD comorbidities could lead to the identification of risk-groups, more effective treatment and improved patient outcomes. Our research presents a two-fold objective: developing an effective machine learning (ML) workflow for text classification and entity relation extraction and assembling a broad list of diseases influencing CKD development and progression. Methods: We analysed 39,680 abstracts with CKD in the title from the Embase library. Abstracts about a disease affecting CKD development and/or progression were selected by multiple ML classifiers trained on a human-labelled sample. The best classifier was further trained with active learning. Disease names in question were extracted from the selected abstracts using a novel entity relation extraction methodology. The resulting disease list and their corresponding abstracts were manually checked and a final disease list was created. Findings: The SVM model gave the best results and was chosen for further training with active learning. This optimised ML workflow enabled us to discern 68 comorbidities across 15 ICD-10 disease groups contributing to CKD progression or development. The reading of the ML-selected abstracts showed that some diseases have direct causal effect on CKD, while others, like schizophrenia, has indirect causal effect on CKD. Interpretation: These findings have the potential to guide future CKD investigations, by facilitating the inclusion of a broader array of comorbidities in CKD prognostic models. Ultimately, our study enhances understanding of prognostic comorbidities and supports clinical practice by enabling improved patient monitoring, preventive strategies, and early detection for individuals at higher CKD development or progression risk.