Vancomycin treatment outcomes and machine learning-based identification of heterogeneous vancomycin-intermediate Staphylococcus aureus infections.

OBJECTIVES: Heterogeneous vancomycin-intermediate Staphylococcus aureus (hVISA) infections remain underdiagnosed with unclear clinical impact. This multicenter study evaluated vancomycin treatment outcomes and developed a machine learning (ML)-based diagnostic model. METHODS: We prospectively enrolled 200 patients with MRSA infections from 20 Chinese hospitals between February 2012 and July 2021. Population analysis profile-area under the curve (PAP-AUC) testing was performed on frozen isolates to confirm hVISA phenotype, and confirmed strains underwent whole-genome sequencing. Vancomycin therapeutic drug monitoring and pharmacokinetic/pharmacodynamic (PK/PD) analysis was performed. A ML model was developed for hVISA prediction. RESULTS: Of 200 MRSA-infected patients, 117 (58.5%) were identified with hVISA infections, including 98 (98/175, 56.0%) adults and 19 (19/25, 76.0%) pediatric patients. Despite higher vancomycin doses, adult patients with hVISA infection achieved lower PK/PD indices than vancomycin-susceptible S. aureus (VSSA). In adults receiving vancomycin monotherapy, hVISA infection was associated with significantly higher treatment failure than VSSA infection (34.6% vs. 5.0%, P = 0.028). The predominant hVISA clones were ST5, ST239 and ST764, with ST5 and ST239 showing higher treatment failure rates than ST764. Among 10 ML models, LASSO-logistic regression performed best with AUC of 0.834 (95% CI 0.711-0.930) on test set, identifying vancomycin MIC, tracheal intubation, and prior penicillin use as key predictors. CONCLUSIONS: High prevalence of hVISA among MRSA isolates and its impact on vancomycin monotherapy warrant early identification. The ML model enables identification of high-risk patients, supporting a shift toward phenotype- and sequence type-guided personalized dosing.