Synthetic Data-Driven Early Prediction Framework for Acute Kidney Injury in Patients Receiving Vancomycin and Ceftazidime/Avibactam.

Abstract

Background: The nephrotoxic risks of combining ceftazidime/avibactam (AVI) with vancomycin (VAN) remain underexplored, despite both agents independently being linked to acute kidney injury (AKI). This study assessed the risk of AKI associated with concurrent VAN and ceftazidime/avibactam (VAN-AVI) therapy and developed synthetic data models to enable early prediction of AKI.

Methods: We conducted a retrospective analysis using electronic health record data from hospitalized adults between 2015 and 2022. The incidence of AKI was compared among patients receiving VAN-AVI or VAN in combination with piperacillin/tazobactam (VAN-TPZ) versus VAN monotherapy. AKI was defined as a composite of de novo and recurrent AKI (i.e., patients had a prior diagnosis of AKI within the preceding 6 months and experienced a new AKI event after 7 days of VAN-AVI initiation). To address sample size imbalance, we applied inverse probability of treatment weighting (IPTW) and generated synthetic datasets using Conditional Tabular Generative Adversarial Networks (CTGAN) and Tabular Variational Autoencoders (TVAE). These synthetic datasets were subsequently used to augment machine learning (ML) models aimed at the early prediction of AKI in patients treated with VAN-AVI combination therapy.

Results: Among the 92 patients receiving VAN-AVI combination therapy, only four (4.3%) patients experienced new-onset AKI, and 66 (71.7%) patients had a recurrent AKI. After applying IPTW, VAN-AVI was associated with a higher risk of AKI Hazard Ratio (HR) = 3.47; 95% Confidence Interval (CI): 1.97-6.11, followed by VAN-TPZ (HR = 1.96; 95% CI: 1.37-2.81), compared to VAN alone. Synthetic data analyses conducted over 1000 iterations supported these findings, with mean HRs for VAN-AVI of 3.80 using TVAE and 4.45 using CTGAN. ML models augmented with synthetic data outperformed those using original data alone. For 30-day AKI prediction, F1-scores improved across all models, with the highest performance observed in the augmented XGBoost and logistic regression classifier (F1 = 0.80).

Conclusion: This study introduces a novel approach that integrates IPTW with synthetic data generation to evaluate drug-associated AKI risk in small-sample cohorts. Although our findings demonstrate a lower incidence of de novo AKI in the VAN-AVI group, the use of synthetic data and augmented ML models significantly improved early AKI prediction. These findings support the potential utility of synthetic data frameworks for scalable drug safety evaluations, although further validation is warranted.