Leveraging large language models for preoperative prevention of cardiopulmonary bypass-associated acute kidney injury

Abstract

Background: Acute kidney injury (AKI) usually occurs after cardiopulmonary bypass (CPB) and threatens life without timely intervention. Early assessment and prevention are critical for saving AKI patients. However, numerical data-driven models make it difficult to predict the AKI risk using preoperative data and lack preventive measures. Large language models (LLM) have demonstrated significant potential for medical decision-making, offering a promising approach.

Objective: For preoperative assessment and prevention of CPB-associated AKI (CPB-AKI).

Methods: Clinical variables were converted into text through prompt engineering and a LLM was used to extract information hidden in the semantics of subtle changes. A multimodal fusion model, fuzing semantic and numerical information, was proposed to assess the AKI risk before surgery. We then used a structural equation model to analyze the impact of preoperative features and intraoperative interventions on CPB-AKI prevention.

Results: A total of 2,056 patients who underwent CPB were enrolled from the intensive care unit of Sir Run Run Shaw Hospital between 2014 and 2022, with 40.5% progressing to AKI. Our model performed better with an area under the receiver operating characteristic curve of 0.9201 compared with the baseline models. The structural equation model's chi-square to degrees of freedom ratio was 0.46, less than 2.0. We discussed how the preoperative prediction model could optimize intraoperative interventions to prevent CPB-AKI.

Conclusions: The prediction model can predict CPB-AKI risk earlier after fuzing the clinical characteristics and their semantics. Preoperative assessment and intraoperative interventions provide decision-making to prevent CPB-AKI.

Keywords: Cardiopulmonary bypass; acute kidney injury; large language model; multimodal fusion model; preoperative assessment and prevention.

Figure 1.

Schemetic of our approach. We propose a multimodal fusion model based on LLM to evaluate AKI risk using preoperative clinical characteristics. Then, a structural equation was constructed to optimize intraoperative interventions. The risk of preoperative assessment was a latent variable in the structural equation. Abbreviations: FFN, feedforward neural network; Q, query of attention; K, key of attention; V, value of attention.

Figure 2.

Flowchart outlining patient enrollment and triplet sampling.

Figure 3.

Curves of assessment. (a) The receiver operating characteristic curve (ROC). (b) the comparison of AUROC pairwise using DeLong test with Benjamini-Hochberg correction (p-values). (c) the precision-recall curve (PRC). (b) the decision curve analysis (DCA). DCA has two extreme curves. One means that all samples are negative, none is treated, and the net benefit is 0; the other means that all samples are positive and treated, and the net benefit is a backslash with a negative slope.

Figure 4.

The importance of preoperative clinical indicators. Abbreviations: PT, prothrombin time; HCT, hematocrit; Na, serum sodium; Cl, serum chloride; Ca, serum calcium; SCr, serum creatinine; ALB, albumin; P, serum phosphate; LDH, lactate dehydrogenase. The p-value less than 0.05 in the structural equation model indicates that the corresponding path or relationship is statistically significant.