Predicting responsiveness to fixed-dose methylene blue in adult patients with septic shock using interpretable machine learning: a retrospective study.

Journal: Scientific reports
PMID:

Abstract

This study aimed to develop an interpretable machine learning model to predict methylene blue (MB) responsiveness in adult patients with refractory septic shock and to identify key factors influencing MB responsiveness using the SHapley Additive exPlanations (SHAP) approach. We retrospectively analyzed data from 416 adult patients with refractory septic shock who received MB treatment at Xiangya Hospital of Central South University between June 2018 and October 2022. MB responders were defined as patients who, within 6 hours after MB administration, exhibited either a reduction in average norepinephrine equivalence (NEE) of ≥ 10% or an increase in mean arterial pressure of ≥ 10 mmHg without an associated increase in NEE. The incidence of MB responders was 38.2%(n=159). Statistical and machine learning methods were used for feature selection, resulting in two datasets (ST and ML). Each dataset was randomly divided into a training set (75%) for model development and a testing set (25%) for internal validation. Prediction models were developed using logistic regression, support vector machine (SVM), random forest, light gradient boosting machine (LightGBM), and explainable boosting machine (EBM). The models were evaluated regarding discrimination, calibration, and clinical benefit. The SVM model trained on the ML dataset demonstrated the best predictive performance, with an area under the curve (AUC) of 0.74 (95% CI 0.62-0.84), 76% accuracy, 36% sensitivity, and 94% specificity. Although the model's sensitivity was low, its high specificity and the safety profile of MB underscore its clinical relevance. The model showed superior net benefit within a 24-85% threshold probability, as determined by decision curve analysis. The SHAP analysis identified the average NEE dose within 6 hours before MB initiation as the most important factor influencing MB responsiveness (P<0.01), with higher doses positively correlating with a greater likelihood of response. Lactate levels were identified as the second most important factor. The optimal model was externally validated in an independent cohort from the same institution, achieving an AUC of 0.75 and an accuracy of 74%.

Authors

  • Shasha Xue
    School of Computer Science and Engineering, Central South University, Changsha, 410083, China.
  • Li Li
    Department of Gastric Surgery, Sichuan Clinical Research Center for Cancer, Sichuan Cancer Hospital & Institute, Sichuan Cancer Center, Affiliated Cancer Hospital of University of Electronic Science and Technology of China, Chengdu, China.
  • Zhuolun Liu
    Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education/Beijing), Department of Radiation Oncology, Peking University Cancer Hospital & Institute, Beijing Cancer Hospital & Institute Beijing, Beijing, China.
  • Feng Lyu
    Department of Computer Science and Engineering, Central South University, Changsha, China.
  • Fan Wu
    Department of Product Design, Dalian Polytechnic University, Dalian 116034, China.
  • Panxiao Shi
    School of Computer Science and Engineering, Central South University, Changsha, 410083, China.
  • Yongmin Zhang
    School of Computer Science and Engineering, Central South University, Changsha, 410083, China. zhangyongmin@csu.edu.cn.
  • Lina Zhang
    Intensive Care Unit, XiangYa Hospital, Central South University, Changsha, China.
  • Zhaoxin Qian
    Department of Critical Care Medicine, Xiangya Hospital, Central South University, Changsha, 410008, China.