Supplementary Material for "Extracting Chronic Kidney Disease Comorbidities from Abstracts using Advanced Machine Learning Techniques: A Comparative Analysis"

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 help with identifying risk-groups, providing more efficient therapies, and increasing patient outcomes. Our study has dual aims: to establish a robust machine learning (ML) process for identifying comorbidities from abstracts, and to compile an extensive list of conditions that influence CKD's onset and progression. Methods: We analysed 39,680 abstracts with CKD in the title downloaded from the Embase library. Seven machine-learning classifiers were compared in identifying abstracts with information of a disease affecting CKD development and/or progression. The top-performing classifier was then further trained with active learning. The relevant disease names were extracted from the chosen abstracts using a novel entity relation extraction technique. The corresponding abstract of each disease was manually reviewed and a final comorbidity list was established. Findings: The SVM classifier proved to be the most effective and was selected for further active learning training. Our machine learning (ML) pipeline helped to identify 71 comorbidities across 15 ICD-10 disease groups that play a role in the onset or progression of CKD. A review of the selected abstracts revealed that certain diseases have a direct causative impact on CKD, while others, such as schizophrenia, have an indirect effect. Interpretation: These insights could steer future research into CKD by promoting the consideration of a wider range of comorbidities in CKD prognostic models. Our study ultimately boosts understanding of prognostic comorbidities and aids in clinical practice by improving patient tracking, preventative strategies, and early detection for individuals at elevated risk of CKD onset or progression. Funding: This research is a part of the first author's PhD project. No additional funding was received for this study.背景：慢性肾脏病（Chronic Kidney Disease，CKD）是一项全球性健康问题，因其初始症状隐匿而常被漏诊，进而导致发病率与死亡率持续攀升。全面掌握慢性肾脏病的共病情况，有助于识别高危人群、优化治疗方案，从而改善患者结局。本研究具有双重目标：一是构建一套稳健的机器学习（Machine Learning，ML）流程，用于从文献摘要中识别共病；二是编制一份可全面涵盖影响慢性肾脏病发病与进展的疾病清单。 方法：本研究从Embase数据库中检索并下载了标题含慢性肾脏病的39680篇摘要，展开分析。对比了7种机器学习分类器在识别携带影响慢性肾脏病发生及/或进展的疾病相关信息的摘要时的性能表现。随后选取性能最优的分类器，通过主动学习算法开展进一步训练。采用新型实体关系抽取技术，从筛选出的摘要中提取相关疾病名称。对每一种疾病对应的摘要进行人工审校，最终确立共病列表。 结果：支持向量机（Support Vector Machine，SVM）分类器被证实为性能最优的模型，被选中用于后续的主动学习训练。本机器学习流程共识别出15个ICD-10疾病分类下的71种与慢性肾脏病发病或进展相关的共病。对筛选出的摘要进行审校后发现，部分疾病对慢性肾脏病具有直接致病作用，而另一些疾病如精神分裂症则仅产生间接影响。 解读：本研究结果可为未来慢性肾脏病的相关研究提供方向，推动在慢性肾脏病预后模型中纳入更多类型的共病因素。本研究最终加深了学界对预后相关共病的认知，并通过优化高危慢性肾脏病发病或进展人群的患者追踪、预防策略与早期筛查手段，为临床实践提供助力。 资助：本研究为第一作者博士阶段研究项目的一部分，未获得额外研究经费支持。