Researchers at the Affiliated Hospital of Nantong University reviewed consecutive adults with Sepsis-3-defined sepsis who received clinician-directed mNGS between January 2022 and November 2025. Of 251 patients screened, 142 were excluded because essential clinical information was unavailable, leaving 109 patients for the overall analysis. Fifty-eight were classified as clinically actionable, meaning that antimicrobial treatment was documented as being modified in response to the mNGS findings, while 51 were classified as non-actionable. A further 37 patients lacked the paired post-report C-reactive protein measurement required for the complete variable-selection procedure, so the final model-development cohort contained 72 patients, including 39 actionable and 33 non-actionable cases.

The mNGS tests detected at least one microorganism in 98 of the 109 patients, compared with 50 detections by conventional culture. The detection yield was 94.8% in the actionable group and 84.3% in the non-actionable group, whereas culture detected organisms in 53.4% and 37.3% of those respective groups. The sequencing assay also produced a descriptively broader organism spectrum, including 12 fungal taxa and 18 DNA or RNA viral taxa not recovered by culture. These comparisons were made at the patient level rather than through uniformly matched specimens, and the clinical reference assessment incorporated mNGS results, so the estimated concordance does not demonstrate that mNGS was independently superior for diagnosis.

Clinical actionability was defined through a prespecified process by two experienced intensive-care physicians who had not cared for the included patients. They reviewed the sampling site, clinical syndrome, imaging, conventional microbiology, treatment history, biological plausibility and the contemporaneous rationale in medical records; disagreements were resolved by consensus. Detection alone did not qualify a result as actionable. Among the 58 actionable cases, 39 involved antimicrobial escalation, eight de-escalation, eight antifungal modification, two antiviral modification and one discontinuation. Among the 51 non-actionable cases, therapy remained unchanged in 49, while two patients underwent modifications not documented as attributable to mNGS. The endpoint therefore measured a pragmatic treatment change rather than whether the change was appropriate or beneficial.

Six prespecified changes measured between the pre-mNGS assessment and the first eligible reassessment after report availability were evaluated: white blood cell count, neutrophil percentage, log-transformed procalcitonin, log-transformed C-reactive protein, Sequential Organ Failure Assessment score and Acute Physiology and Chronic Health Evaluation II score. These follow-up measurements were obtained before any antimicrobial modification classified as the outcome, although clinicians could see the individual laboratory results while making treatment decisions. The timing was not standardized, and the interval between the report and repeat testing could not be reliably reconstructed for all patients. Among the six candidates, only the change in white blood cell count, designated dWBC, passed the univariable screening threshold and was retained by the subsequent least absolute shrinkage and selection operator procedure.

The baseline model used information available before mNGS and included age, diabetes, baseline white blood cell count and baseline log-transformed procalcitonin. The dynamic model added dWBC and was intended for use only after the sequencing report became available. In the 72-patient development cohort, the apparent area under the receiver operating characteristic curve rose from 0.744 for the baseline model to 0.816 after adding dWBC, while the apparent Brier score fell from 0.199 to 0.173. However, a paired DeLong comparison did not reach statistical significance, with P = 0.088, so the numerical improvement did not establish that the dynamic model was superior. In the broader cohort, dWBC remained positively associated with actionability in univariable analysis, but it was no longer statistically significant after adjustment for the four baseline variables in a post hoc refitting, with an adjusted odds ratio of 1.06, 95% confidence interval 0.99–1.13 and P = 0.092.

Internal assessment used 1,000 nonparametric bootstrap resamples, with all resamples successfully fitted. Because the analysis refitted the finalized predictor sets without repeating univariable screening and variable selection, the correction was conditional on those choices and could underestimate total development optimism. The optimism-corrected AUCs were 0.697 for the baseline model and 0.768 for the dynamic model, while the corrected Brier scores were 0.228 and 0.207, respectively. Corrected calibration slopes were 0.783 and 0.732, indicating some residual overfitting. An exploratory decision-curve analysis based on uncorrected predictions suggested greater net benefit for the dynamic model at parts of the examined threshold range, but it was not optimism-corrected and did not establish clinical utility or patient benefit.

The investigators also tested whether the relationship between dWBC and actionability differed according to baseline white blood cell count. Among patients with a baseline count of 12 × 10⁹/L or less, the adjusted odds ratio for dWBC was 1.15, with a 95% confidence interval of 0.95–1.38 and P = 0.144. Above that threshold, the adjusted odds ratio was 1.16, with a 95% confidence interval of 1.01–1.34 and P = 0.034. Although statistical significance appeared only in the higher-count stratum, the point estimates were nearly identical and their confidence intervals overlapped. A formal continuous interaction test was not significant, with P = 0.494, providing no evidence that the association varied according to baseline white blood cell count. The subgroup analysis was exploratory because the strata were small.

Several limitations constrain interpretation. The study was small, retrospective and confined to one center, and the exclusion of 142 screened patients and the further restriction to 72 complete cases may have introduced selection bias. Differences in sex and specimen distribution also limit generalizability, while requiring an eligible post-report measurement before treatment modification may have created measurement-availability or survivor bias. Respiratory specimens were more common in the actionable group, making anatomical sampling site a potential confounder that the small model could not assess reliably. Treatment changes were attributed retrospectively, and incomplete documentation, hindsight bias, shared institutional practice and the absence of a counterfactual prevented causal conclusions. The authors also did not assess treatment appropriateness, broader effects on infection control or public health, cost-effectiveness, or patient-centered outcomes. Prospective multicenter validation with standardized repeat measurements and a prespecified post-report pathway is therefore needed before the model could support clinical decisions.

Subject of Research: Development of a post-report sepsis model linking peri-report white blood cell changes to mNGS-associated antimicrobial modification.

Article Title: Predicting clinically actionable metagenomic next-generation sequencing in sepsis through dynamic inflammatory trajectories

Article References: Ding, X., Bao, P., Huang, X., Xu, R., Tang, X., Wu, Y., Cui, X., & Lu, Y. (2026). Predicting clinically actionable metagenomic next-generation sequencing in sepsis through dynamic inflammatory trajectories. Frontiers in Microbiology, 17, Article 1949602. https://doi.org/10.3389/fmicb.2026.1949602

Image Credits: AI Generated

DOI: 10.3389/fmicb.2026.1949602

Keywords: sepsis, metagenomic sequencing, white blood cell count, clinical prediction model, antimicrobial stewardship, infectious diseases, diagnostic actionability, inflammatory trajectories, intensive care, Predicting, clinically, actionable