PO.BCS01.16 · 生物信息与计算
基于机器学习的筛查工具用于在实验室检测有限的慢性肾脏病患者中检测多发性骨髓瘤
Machine learning-based screening tool for multiple myeloma detection in chronic kidney disease patients with limited laboratory testing
作者与单位 Authors & Affiliations
摘要 Abstract
中文摘要
背景:多发性骨髓瘤(MM)和慢性肾脏病(CKD)具有相似的临床表现,包括肾功能损害和贫血,使得在普通门诊和肾脏病门诊中诊断MM具有挑战性。许多CKD患者携带未确诊的MM,但医疗保险限制和公共卫生系统的制约限制了专项检查(血清蛋白电泳、游离轻链、免疫球蛋白)的开具。这种诊断延迟可能导致不可逆的肾脏损伤和治疗延误。我们开发了一个使用常规实验室参数的机器学习模型,以识别需要MM特异性检查或血液科转诊的CKD患者。
方法:我们分析了来自一家三级中心的4,759例CKD患者(591例MM,4,168例非MM对照;患病率12.4%)。仅使用普通诊疗中可获得的常规检查,我们构建了包括CBC(血红蛋白、WBC、血小板)、生化(总蛋白、白蛋白、钙、LDH、肌酐)和尿液检查(蛋白、白蛋白、肌酐)在内的特征。测试了三个特征集:RANK1(20个特征)、RANK2(16个特征)和RANK3(4个最简特征)。我们使用分层5折交叉验证比较了XGBoost、随机森林和逻辑回归,同时使用非平衡数据和SMOTEENN重采样数据。
结果:XGBoost取得了最佳性能:准确率95.1%,F1分数0.798,ROC-AUC 0.962,PR-AUC 0.889。最简的4特征模型(血小板、WBC、尿蛋白、LDH)保持了临床上有用的性能(F1 0.596,PR-AUC 0.627)。SHAP分析揭示血红蛋白/WBC比值、血小板计数、白蛋白/总蛋白比值和总蛋白为最重要的预测因子,在无需专项检测的情况下捕捉了MM特征性的血液系统抑制和副蛋白血症。降维可视化(PCA、UMAP、t-SNE)显示了MM与非MM CKD患者之间明显的聚类模式,证实了尽管临床表现存在重叠,其特征空间仍具有内在的可分离性。
结论:这一基于ML的筛查工具展示了仅使用常规实验室数据识别需要MM特异性检测的高风险CKD患者的潜力。然而,这项回顾性单中心研究存在局限性,包括潜在的选择偏倚和缺乏外部验证。在临床实践中实施之前,需要开展前瞻性多中心验证研究,以评估其真实世界的临床效用、最佳决策阈值以及对诊断及时性的影响。
查看英文原文 English abstract
Background: Multiple myeloma (MM) and chronic kidney disease (CKD) share similar clinical manifestations including renal impairment and anemia, making MM diagnosis challenging in general and nephrology clinics. Many CKD patients harbor undiagnosed MM, but healthcare insurance restrictions and public health system constraints limit ordering of specialized tests (serum protein electrophoresis, free light chains, immunoglobulins). This diagnostic delay can result in irreversible renal damage and delayed treatment. We developed a machine learning model using routine laboratory parameters to identify CKD patients requiring MM-specific workup or hematology referral.
Methods: We analyzed 4,759 CKD patients (591 MM, 4,168 non-MM controls; 12.4% prevalence) from a tertiary center. Using only routine tests available in general practice, we engineered features including CBC (hemoglobin, WBC, platelets), biochemistry (total protein, albumin, calcium, LDH, creatinine), and urine studies (protein, albumin, creatinine). Three feature sets were tested: RANK1 (20 features), RANK2 (16 features), and RANK3 (4 minimal features). We compared XGBoost, Random Forest, and Logistic Regression using stratified 5-fold cross-validation. with both unbalanced and SMOTEENN-resampled data.
Results: XGBoost achieved optimal performance: accuracy 95.1%, F1-score 0.798, ROC-AUC 0.962, PR-AUC 0.889. The minimal 4-feature model (platelet, WBC, urine protein, LDH) maintained clinically useful performance (F1 0.596, PR-AUC 0.627). SHAP analysis revealed hemoglobin/WBC ratio, platelet count, albumin/total protein ratio, and total protein as top predictors, capturing MM's characteristic hematologic suppression and paraproteinemia without requiring specialized assays. Dimensionality reduction visualization (PCA, UMAP, t-SNE) demonstrated distinct clustering patterns between MM and non-MM CKD patients, confirming inherent feature space separability despite overlapping clinical presentations.
Conclusions: This ML-based screening tool demonstrates potential for identifying high-risk CKD patients warranting MM-specific testing using only routine laboratory data. However, this retrospective single-center study has limitations including potential selection bias and lack of external validation. Prospective multicenter validation studies are needed to assess real-world clinical utility, optimal decision thresholds, and impact on diagnostic timeliness before implementation in clinical practice.
利益披露 Disclosure
W. Wu, None..
M. Yang, None..
C. Yang, None.