基于TabPFN的围手术期镇痛镇静药物相关高钾血症风险预测模型研究

Prediction of Perioperative Analgesic and Sedative-Associated Hyperkalemia Risk Based on TabPFN

关键词:  围手术期  高钾血症  TabPFN  风险预测  药物不良反应

Key words:  perioperative period  hyperkalemia  TabPFN  risk prediction  adverse drug reaction

摘要:

目的 围手术期高钾血症严重威胁患者生命安全，在慢性肾脏病（CKD）等高危人群危害更为显著。本研究旨在利用表格数据基础模型（TabPFN），构建围手术期镇痛镇静药物相关高钾血症风险预测模型，并挖掘驱动该风险事件的关键药物因子，为优化围手术期患者用药管理提供工具。方法 回顾性收集武汉市某三级甲等医院2022年7月至2025年11月围手术期患者临床数据。将2023年1月至2025年11月的524例患者按3:1比例划分为训练集与内部验证集，另选取2022年下半年123例患者作为外部验证集。纳入包括患者基线特征、实验室检查及详细的围术期用药（如NSAIDs、镇痛类、镇静类等）等67个潜在预测因子。采用基于Transformer架构的TabPFN算法进行模型拟合，并通过DALEX的RMSE DROPOUT方法进行关键因子筛选后，再次引入TabPFN完成回归模型建立。结果 TabPFN模型在内部验证集中展现出优异的预测性能，受试者工作特征曲线下面积（AUC）为0.898，精确率-召回率曲线下面积（AUPRC）为0.791，校准曲线显示预测概率与实际风险拟合良好。特征重要性分析显示，除了血肌酐、基础血钾及AKI分期等病理生理指标外，药物因素中含钾药物、利尿剂、甲苯磺酸瑞马唑仑剂量以及非甾体抗炎药（NSAIDs）对高钾血症风险具有显著贡献。其中，瑞马唑仑剂量位列麻醉镇静类药物风险因子的首位。结论 基于TabPFN模型构建的围手术期镇痛镇静药物相关高钾血症风险预测模型，具备良好的区分度与稳健性。DALEX工具为模型关键因子识别提供筛选依据与可解释分析，明确了甲苯磺酸瑞马唑仑与NSAIDs等药物的致高钾血症的风险。该模型建立作为临床药师个体化药学监护的量化工具，推动围术期用药安全管理从“被动应对”向“主动预防”转变。

Abstract:

Objective Perioperative hyperkalemia poses a serious threat to patient safety, with particularly significant risks in high-risk populations such as those with chronic kidney disease (CKD). This study aims to utilize the Tabular Pattern-Based Neural Network (TabPFN) to construct a predictive model for hyperkalemia risk associated with perioperative analgesic and sedative medications. It further seeks to identify key drug factors driving this risk event, thereby providing a tool for optimizing medication management in perioperative patients. Methods Clinical data of perioperative patients from a tertiary Grade-A hospital in Wuhan were retrospectively collected from July 2022 to November 2025. A total of 524 patients from January 2023 to November 2025 were divided into a training set and an internal validation set at a 3:1 ratio, while 123 patients from the second half of 2022 served as an external validation set. The study incorporated 67 potential predictors, including baseline characteristics, laboratory tests, and detailed perioperative medication data (e.g., NSAIDs, analgesics, sedatives, etc.). The TabPFN algorithm, based on the Transformer architecture, was employed for model fitting,with key factor screening conducted via RMSE DROPOUT of the DALEX methods. Results The TabPFN model demonstrated superior predictive performance in the internal validation set, with an Area Under the Receiver Operating Characteristic Curve (AUC) of 0.898 and an Area Under the Precision-Recall Curve (AUPRC) of 0.791. The calibration plot indicated a good fit between predicted probabilities and actual risks. Feature importance analysis revealed that, in addition to pathophysiological indicators such as serum creatinine, baseline potassium, and AKI stage, drug-related factors—specifically potassium-containing drugs, diuretics, remimazolam tosilate dosage, and non-steroidal anti-inflammatory drugs (NSAIDs)—contributed significantly to the risk of hyperkalemia. Notably, the dosage of remimazolam ranked highest among anesthesia/sedation-related risk factors. Conclusion A perioperative hyperkalemia risk prediction model for analgesic and sedative drugs, constructed based on the TabPFN model, demonstrates excellent discriminatory power and robustness. The DALEX tool provides screening criteria and interpretable analysis for identifying key model factors, clarifying the hyperkalemia risk associated with drugs such as remimazolam tosylate or NSAIDs. This model serves as a quantitative tool for clinical pharmacists to conduct individualized pharmaceutical monitoring, driving the shift in perioperative medication safety management from “reactive response” to “proactive prevention”.