Scholarlink

Objective <p>To develop and internally validate an interpretable machine-learning model using routine precollection variables to predict mononuclear cell (MNC) collection efficiency (CE) during leukapheresis for CAR-T manufacturing.</p> Methods <p>This retrospective study included 206 consecutive patients who underwent leukapheresis between 2022 and 2025. Candidate variables were screened with Pearson correlation, Boruta, recursive feature elimination, and variance-threshold filters. Six supervised algorithms were trained with nested cross-validation, using an inner fivefold loop for tuning and an outer tenfold loop repeated three times for performance estimation. Bootstrap resampling (1000 iterations) generated confidence intervals, and SHAP analysis was used for interpretation.</p> Results <p>Extreme gradient boosting (XGBoost) showed the best continuous prediction performance, with an <i>R</i><sup>2</sup> of 0.67 (95% CI: 0.64–0.70), Pearson <i>r</i> = 0.82, RMSE = 12.4%, and MAE = 9.8%. In the secondary triage analysis, observed high CE was defined as CE ≥ 60% (33/206, 16.0%). The predicted CE score discriminated high CE with an AUC of 0.84. A predicted-CE threshold of ≥ 50% yielded sensitivity of 84.8%, specificity of 97.1%, PPV of 84.8%, and NPV of 97.1%. The leading predictors were precollection lymphocyte count, lymphocyte/WBC ratio, age, body surface area, and monocyte count.</p> Conclusion <p>Routine precollection variables estimated MNC CE with good internal validity and may support individualized leukapheresis planning before CAR-T manufacturing.</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

Machine Learning-Based Prediction of Mononuclear Cell Collection Efficiency for CAR-T Manufacturing

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Wei Xie,
  • Lin Liu,
  • Jin-hui Shu,
  • Meng-yi Du,
  • Heng Mei

摘要

Objective

To develop and internally validate an interpretable machine-learning model using routine precollection variables to predict mononuclear cell (MNC) collection efficiency (CE) during leukapheresis for CAR-T manufacturing.

Methods

This retrospective study included 206 consecutive patients who underwent leukapheresis between 2022 and 2025. Candidate variables were screened with Pearson correlation, Boruta, recursive feature elimination, and variance-threshold filters. Six supervised algorithms were trained with nested cross-validation, using an inner fivefold loop for tuning and an outer tenfold loop repeated three times for performance estimation. Bootstrap resampling (1000 iterations) generated confidence intervals, and SHAP analysis was used for interpretation.

Results

Extreme gradient boosting (XGBoost) showed the best continuous prediction performance, with an R2 of 0.67 (95% CI: 0.64–0.70), Pearson r = 0.82, RMSE = 12.4%, and MAE = 9.8%. In the secondary triage analysis, observed high CE was defined as CE ≥ 60% (33/206, 16.0%). The predicted CE score discriminated high CE with an AUC of 0.84. A predicted-CE threshold of ≥ 50% yielded sensitivity of 84.8%, specificity of 97.1%, PPV of 84.8%, and NPV of 97.1%. The leading predictors were precollection lymphocyte count, lymphocyte/WBC ratio, age, body surface area, and monocyte count.

Conclusion

Routine precollection variables estimated MNC CE with good internal validity and may support individualized leukapheresis planning before CAR-T manufacturing.