Letter to the Editor: Bacterial extracellular vesicle as a predictive biomarker for postoperative delirium status after spinal surgery: a prospective cohort study

We read with great interest the prospective cohort study by Jee et al showing that preoperative circulating bacterial extracellular vesicles (BEVs) stratify postoperative delirium (POD) risk in older adults undergoing spine surgery, with lower BEV diversity and Gammaproteobacteria enrichment in POD and an externally tested machine-learning classifier derived from significant BEV features. Notably, clinically interpretable partial-dependence thresholds for Acinetobacter and Moraxellaceae may inform preoperative counseling and pathway design[1]. We report in accordance with internationally endorsed AI reporting standards, notably the TITAN guideline[2]. We commend the authors for focusing on a preoperative, implementable biomarker and for including an independent validation set. To further strengthen analytical rigor and accelerate translation, we respectfully offer four concise suggestions: (1) Internal validation, calibration, and clinical utility reporting; (2) Compositionality and absolute quantification; (3) Low-biomass safeguards for circulating BEV 16S workflows; (4) From predicted function to orthogonal confirmation. Because statistical screening of features preceded model fitting, nested cross-validation with embedded feature selection is recommended to mitigate optimism in small, high-dimensional biomedical datasets[3,4]. Beyond accuracy, transparent reporting per contemporary guidance (e.g., AUROC/AUPRC with 95% CIs, calibration via Brier score and calibration plots, and decision-curve analysis to quantify net clinical benefit) would allow readers to appraise discrimination, reliability, and usefulness[3,5]. In the external cohort, providing sensitivity, specificity, PPV, and NPV would also address class-imbalance concerns. Taxonomic EV profiles expressed as relative abundances are inherently compositional; apparent differences can arise from the closure constraint. We suggest complementing univariate tests with composition-aware approaches (e.g., log-ratio modeling or ANCOM-class methods) and, where sample volume allows, adding absolute quantification (qPCR or spike-in–normalized counts)[6–8]. Such analyses would corroborate proposed thresholds for Acinetobacter or Moraxellaceae and reduce the risk of spurious cut-points. Serum/plasma BEV 16S assays operate in a low-biomass regime prone to reagent and batch contaminants, often within Proteobacteria. We encourage explicit detailing (and, ideally, sharing) of extraction blanks/no-template controls, batch randomization, and contaminant-aware filtering to reinforce that the Gammaproteobacteria signals – especially Acinetobacter – are biological rather than exogenous[8]. The functional inferences are appropriately framed as hypothesis-generating. Where feasible, shotgun metagenomics of BEVs and/or untargeted metabolomics could directly assess proposed mediators (e.g., succinate vs short-chain fatty acids, pyruvate, 2-oxoglutarate), while targeted protein/glycan profiling (e.g., OmpA on Acinetobacter vesicles) would link taxa to mechanistic payloads relevant to neuroinflammation and mitochondrial stress—moving from association to actionable biology[9]. Two deployment-focused clarifications could expedite bedside use. First, quantifying the incremental value of BEVs over conventional assessments (e.g., ASA class, Charlson Comorbidity Index, cognitive screens) using reclassification metrics would position BEVs within existing preoperative clinics. Second, prospectively protocolizing thresholds (e.g., a pre-specified Acinetobacter cut-point) and embedding them in perioperative pathways would allow real-world testing of risk-adapted care bundles (delirium prevention, monitoring intensity, anesthesia choices). In summary, Jee et al provide compelling evidence that preoperative BEV profiling may enhance POD risk stratification beyond routine factors in a well-characterized geriatric spine cohort. Implementing nested validation with calibration and decision-curve analysis, composition-aware statistics with absolute quantification, low-biomass safeguards, and orthogonal functional assays would, in our view, consolidate analytical rigor and hasten integration into perioperative decision-making.