Scholarlink

<p>High free light chain (iFLC) multiple myeloma is associated with renal impairment and poor outcomes. Existing risk models are largely derived from Western populations and do not account for dynamic biomarkers in resource-limited settings. We conducted a retrospective pilot study for model development and internal validation in a cohort of 122 patients with myeloma and baseline FLC &gt; 1000&#xa0;mg/L. A risk score was developed using explainable machine learning (XGBoost with SHAP analysis) to identify patients at high risk of mortality. Percentage change in creatinine, eGFR and FLC levels were derived as dynamic biomarkers. Permutation adjusted log-rank tests were used to identify optimal risk-score thresholds. Clustering analysis was performed to identify latent patient subgroups. A total of 122 patients were included, with a median age of 63 years (IQR 55–69) and M: F ratio of 69:53. Median baseline iFLC was 3000&#xa0;mg/L (IQR 1011–7413), creatinine 2&#xa0;mg/dL (IQR 1.01–6.93), and eGFR 47 mL/min/1.73&#xa0;m² (IQR 9–89). Renal impairment was present in 69 patients (56.5%) at baseline. After a median follow up of 15.5 months, patients with renal impairment had significantly inferior median OS (NR vs. 29.96 months, <i>p</i> = 0.033) and PFS (NR vs. 15.21 months, <i>p</i> = 0.012). XGBoost identified ΔFLC, age, and ΔCreatinine as the strongest predictors of OS, contributing weighted points of 5, 3, and 2, respectively to give a total risk score of 10. A cutoff score of 5 stratified patients into high- and low-risk groups, with the high-risk group showing a median OS of 3.45 months compared to NR (HR 5.55, 95% CI 1.14–26.93, <i>p</i> = 0.033). Clustering further revealed four distinct patient subtypes based on age and renal dynamics, demonstrating significant OS differences (range 3.45 months – NR across clusters, χ² = 9.54, <i>p</i> = 0.02). SHAP analysis clearly illustrated the relative prognostic importance of these variables using summary and box plots. This exploratory study successfully developed and internally validated an explainable machine learning risk score for early post treatment response based risk assessment that highlights dynamic biomarkers as critical predictors of early mortality in high FLC myeloma. The findings support the feasibility of using XAI to identify high-risk subgroups based on local data and external validation in larger, prospective, and multicenter cohorts is planned before clinical implementation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Model Development and Internal Validation of a Machine Learning Risk Score for High Free Light Chain Myeloma

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Suvir Singh,
  • Reshma Roshan,
  • Kunal Jain,
  • Imran Nazir,
  • Akanksha Chhabra,
  • Javid Rasool,
  • Ekta Bansal

摘要

High free light chain (iFLC) multiple myeloma is associated with renal impairment and poor outcomes. Existing risk models are largely derived from Western populations and do not account for dynamic biomarkers in resource-limited settings. We conducted a retrospective pilot study for model development and internal validation in a cohort of 122 patients with myeloma and baseline FLC > 1000 mg/L. A risk score was developed using explainable machine learning (XGBoost with SHAP analysis) to identify patients at high risk of mortality. Percentage change in creatinine, eGFR and FLC levels were derived as dynamic biomarkers. Permutation adjusted log-rank tests were used to identify optimal risk-score thresholds. Clustering analysis was performed to identify latent patient subgroups. A total of 122 patients were included, with a median age of 63 years (IQR 55–69) and M: F ratio of 69:53. Median baseline iFLC was 3000 mg/L (IQR 1011–7413), creatinine 2 mg/dL (IQR 1.01–6.93), and eGFR 47 mL/min/1.73 m² (IQR 9–89). Renal impairment was present in 69 patients (56.5%) at baseline. After a median follow up of 15.5 months, patients with renal impairment had significantly inferior median OS (NR vs. 29.96 months, p = 0.033) and PFS (NR vs. 15.21 months, p = 0.012). XGBoost identified ΔFLC, age, and ΔCreatinine as the strongest predictors of OS, contributing weighted points of 5, 3, and 2, respectively to give a total risk score of 10. A cutoff score of 5 stratified patients into high- and low-risk groups, with the high-risk group showing a median OS of 3.45 months compared to NR (HR 5.55, 95% CI 1.14–26.93, p = 0.033). Clustering further revealed four distinct patient subtypes based on age and renal dynamics, demonstrating significant OS differences (range 3.45 months – NR across clusters, χ² = 9.54, p = 0.02). SHAP analysis clearly illustrated the relative prognostic importance of these variables using summary and box plots. This exploratory study successfully developed and internally validated an explainable machine learning risk score for early post treatment response based risk assessment that highlights dynamic biomarkers as critical predictors of early mortality in high FLC myeloma. The findings support the feasibility of using XAI to identify high-risk subgroups based on local data and external validation in larger, prospective, and multicenter cohorts is planned before clinical implementation.