Routine Blood Tests Outperform CRP for Distinguishing Bacterial From Viral Infection in Children
Mihaela Demireva, Zhecho Mitev, Djuna Chinareva-Klimentova, Svetoslav Ivanov, Georgi Nalbantov, Dimitar Mitev
cs.LG
Quick verdict
CBC+age ML models beat CRP-only rules (AUC 81.7% vs 57.4%, n=906 children); plausible but 74% case attrition and label noise weaken the single-center claim.
Summary
This retrospective single-center study (Zdraveto Hospital, Sofia) asks whether routine complete blood count (CBC) parameters plus CRP and age can distinguish bacterial from viral infection in 906 children aged 2-14 evaluated between 2022 and 2026. From 5,347 screened patients, 3,612 had confirmed single-type infection, but only 956 had complete CBC and CRP results; the 50 most recent formed a held-out set, leaving 906 (482 viral, 424 bacterial) for modeling. The authors trained L2-penalized logistic regression and XGBoost, each with and without CRP, tuned hyperparameters by stratified 5-fold cross-validation, and reported bootstrap 95% CIs (1,000 resamples) on out-of-fold predictions. XGBoost with CRP performed best (AUC 81.7%, sensitivity 70.8%, specificity 79.2%); dropping CRP cost only about one AUC point (80.8%). All models beat a CRP-only rule thresholded at 22 mg/L (AUC 57.4%, sensitivity 30.9%, specificity 85.6%). SHAP ranked WBC, lymphocyte count, and monocyte percentage as most informative. On the 50-patient holdout, models kept AUC around 74-75% versus 52.8% for the CRP rule. Main limitations: roughly 74% of infected children were excluded for missing labs with no missingness analysis; the viral reference panel covered only a few pathogens; sensitivity and specificity were compared at unmatched operating points; and validation is single-center with no code, calibration, or decision-curve analyses.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- All four trained models (LR and XGBoost, with and without CRP) beat a CRP-only 22 mg/L decision rule in AUC (best: 81.7% vs 57.4%).
- Adding CRP to CBC-plus-age features improves XGBoost AUC by only ~1 point (81.7% to 80.8%) and logistic regression marginally, implying CBC carries most of the discriminative signal.
- WBC, absolute lymphocyte count and monocyte percentage are the most influential features per SHAP, more informative than CRP in this pediatric cohort.
- The advantage persists on a temporal 50-patient holdout (AUC 74.1-74.8% vs 52.8% for the CRP rule).
Strengths
Clean CRP ablation isolates the paper's core question
The same stratified 5-fold splits train LR and XGBoost with and without CRP (Table 2): XGBoost 81.7% vs 80.8% AUC, LR 79.8% vs 79.7%, so CRP's marginal contribution is measured directly rather than asserted.
Honest treatment of the CRP baseline
Section 2.3 explicitly flags that the 22 mg/L cutoff was chosen on the whole dataset ('overoptimistic'), and the class overlap in CRP is documented via histograms (Fig S1) — an unusually candid baseline setup.
Uncertainty is reported properly
Temporal holdout plus interpretable output
Weaknesses
Severe selection attrition, unexamined
Imperfect, narrow reference standard
Unmatched operating points and untested significance
No decision-analytic evaluation despite clinical framing
Reproducibility gaps and small inconsistencies
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.21332, cs.LG) claims that routine blood tests — CBC parameters plus age — distinguish bacterial from viral infection in children better than CRP alone. Cohort: 906 children aged 2-14 (482 viral / 424 bacterial) from Zdraveto Hospital, Bulgaria, 2022-2026. Four models (L2 logistic regression and XGBoost, each with/without CRP) plus a CRP-only baseline at a 22 mg/L cutoff. Central claims: (1) all trained models beat the CRP baseline in AUC (best 81.7% vs 57.4%); (2) CRP adds only ~1pp AUC to XGBoost (81.7% to 80.8%) and, per the Discussion, 0.2pp to LR — though Table 2 shows 79.8% vs 79.7%, a 0.1pp discrepancy; (3) SHAP identifies WBC, LYM and MON% as the most informative features. The clinical motivation (antibiotic stewardship) is legitimate, and the comparison target — the CRP-only decision rule — is exactly the practice the authors want to improve.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Can routine blood tests tell whether a child's infection is bacterial or viral?
In this Bulgarian study of 906 children aged 2-14, an XGBoost model using complete blood count, CRP and age reached AUC 81.7% (sensitivity 70.8%, specificity 79.2%) in cross-validation — clearly better than a CRP-only rule (AUC 57.4%) but still too imprecise to replace clinical judgment.
What CRP cutoff separates viral from bacterial infection in this paper?
The authors chose 22 mg/L, the value maximizing accuracy on their whole dataset. That rule caught only 30.9% of bacterial infections while correctly identifying 85.6% of viral ones, illustrating why CRP alone is a weak discriminator in this cohort.
Which blood count values were most useful for distinguishing bacterial from viral infection?
SHAP analysis of the best model ranked white blood cell count (WBC), absolute lymphocyte count (LYM) and monocyte percentage (MON%) as most influential; higher WBC and LYM pushed predictions toward bacterial infection, and CRP ranked below these features.
Is the study's data or code publicly available?
No. The data come from a single Bulgarian hospital (Zdraveto, 2022-2026) and the paper contains no code or data availability statement, so the reported numbers cannot currently be independently re-derived.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures (Fig 1 SHAP, Fig 2 ROC, Fig S1 CRP histograms) are not analyzable from the extracted text; figure-level statements rely on captions and body text.
- arXiv v1 preprint (18 Sep 2026), cs.LG category, apparently not yet peer reviewed; supplementary Tables S1/S2 are quoted in the text but no separate SI files were seen.
- Exact hyperparameter search ranges, threshold-selection rules for ML models, and scaling fit scope are absent from the text, limiting verification of reported metrics.
- Minor editorial sloppiness in the source ('best mode', 'MPW', 'the four mentioned models' before they are enumerated) suggests light proofreading; unlikely to affect the results themselves.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.