目的 構(gòu)建基于誤差反向傳播神經(jīng)網(wǎng)絡(luò)（BPNN）的重癥感染患者利奈唑胺血藥濃度預(yù)測模型，為利奈唑胺個體化給藥提供依據(jù)。方法 納入2022年7月—2024年12月天津市第四中心醫(yī)院重癥醫(yī)學(xué)科（ICU）接受利奈唑胺治療的113例重癥感染患者，采用液相色譜-串聯(lián)質(zhì)譜（LC-MS/MS）法測定利奈唑胺血藥濃度，通過Boruta算法篩選與利奈唑胺血藥濃度相關(guān)的特征變量，構(gòu)建BPNN模型并開發(fā)可視化預(yù)測操作界面，結(jié)合SHAP法對模型的特征變量進(jìn)行解釋分析。結(jié)果BPNN模型驗(yàn)證結(jié)果顯示，預(yù)測值與測定值相關(guān)系數(shù)（R²）為0.85，平均絕對誤差（MAE）為1.065 mg·L^-1，84.2%樣本預(yù)測誤差≤2 mg·L^-1。SHAP分析顯示，血清丙氨酸轉(zhuǎn)氨酶（ALT）、總膽紅素（TB）、白蛋白（ALB）及聯(lián)用P-糖蛋白（P-gp）抑制劑對利奈唑胺血藥濃度呈正向貢獻(xiàn)，而肌酐清除率（CLCr）、身體質(zhì)量指數(shù)（BMI）、C反應(yīng)蛋白（CRP）呈負(fù)向貢獻(xiàn)；可視化預(yù)測操作界面實(shí)現(xiàn)輸入特征變量后一鍵生成血藥濃度預(yù)測值。結(jié)論 構(gòu)建的BPNN模型對利奈唑胺血藥濃度預(yù)測能力良好，SHAP提示多因素影響利奈唑胺血藥濃度，可視化預(yù)測界面實(shí)現(xiàn)對利奈唑胺血藥濃度的實(shí)時預(yù)測。

Objective To construct a prediction model for linezolid blood concentration in critically ill patients based on the Back Propagation Neural Network (BPNN), to support personalized medication.Methods A total of 113 critically ill patients treated with linezolid in the Intensive Care Unit (ICU) of Tianjin 4th Center Hospital from July 2022 to December 2024 were included in the study. Linezolid blood concentration was determined by liquid chromatography-tandem mass spectrometry (LC-MS/MS). The Boruta algorithm was used to select feature variables related to linezolid blood concentration, and a BPNN model was constructed along with a visual prediction interface. The SHAP (Shapley Additive Explanations) method was used for interpretative analysis of the model's feature variables.Results The validation results of the BPNN model showed that the correlation coefficient (R²) between the predicted and measured values was 0.85, with a mean absolute error (MAE) of 1.065 mg·L^-1, and 84.2% of the samples had prediction errors of ≤ 2 mg·L^-1. SHAP analysis showed that serum alanine aminotransferase (ALT), total bilirubin (TB), albumin (ALB), and the use of P-glycoprotein (P-gp) inhibitors had a positive impact on linezolid blood concentration, while creatinine clearance rate (CLCr), body mass index (BMI), and C-reactive protein (CRP) negatively impacted it; The visual prediction interface allows for one-click generation of blood concentration predictions after inputting feature variables.Conclusion The constructed BPNN model shows strong predictive ability for linezolid blood concentration. SHAP analysis suggests that multiple factors influence linezolid blood concentration, and the visual prediction interface allows for real-time predictions of linezolid blood concentration.

R978

天津市衛(wèi)生健康科技項(xiàng)目（TJWJ2022QN038）