machine learning prediction of bacterial infections in infants aged 1 to 90 days using routine non-invasive paraclinical markers versus CSF culture gold standard

Source article: Machine learning models for predicting neonatal bacterial infections: a retrospective cohort study

Bacterial infections represent a critical threat to neonatal health, accounting for approximately 25% of neonatal mortality globally. Timely and precise diagnosis in infants aged 1 to 90 days is essential to facilitate rapid intervention and prevent severe complications. This study aimed to develop and evaluate machine learning (ML) models for the early, non-invasive prediction of bacterial infections using routine clinical data, maximizing clinical interpretability for point-of-care triage. Data from 306 infant…

By the Truvace deskAugust 16, 20262 min read1 sourcePeer-reviewedTRV-2026-0783· cite & verify

Trace impact readingPositive state

P 68

Both sides are scored from claims and sources, not community votes.

G 74

In brief

Researchers retrospectively analyzed 306 infants aged 1 to 90 days hospitalized between 2014 and 2022 in Khorasan Razavi, Iran, using CSF culture via lumbar puncture as the gold standard, to train nine machine learning classifiers on routine non-invasive paraclinical markers with nested cross-validation and SHAP interpretation.

The work matters because bacterial infections account for about 25% of neonatal mortality globally and lumbar puncture is invasive; the models offer an objective risk-stratification aid at point-of-care triage, but the documented sensitivity-specificity trade-off and optimism gaps mean it remains uncertain whether the tool can reduce unnecessary procedures without missing infections in broader populations.

Main points

Retrospective analysis of 306 infants aged 1 to 90 days hospitalized 2014-2022 with CSF culture via lumbar puncture as gold standard (158 infectious, 148 non-infectious).
Nine classifiers evaluated via leakage-safe nested cross-validation (5-folds x 2 repeats outer, threefold inner) with QuantileTransformer normalization and SHAP interpretation.
HistGBM had highest raw AUROC 0.786 (95% CI: 0.757-0.812) but severe training optimism gap 0.214; L2-regularized logistic regression selected for minimal gap 0.083 and Brier score 0.208.

The gain

L2-regularized logistic regression trained on routine non-invasive paraclinical markers achieved stable discrimination with minimal generalization gap and robust calibration for early prediction of bacterial infections in infants aged 1 to 90 days.

The problem

When tuned for ≥95% sensitivity screening, the models suffered a steep parallel decline in specificity and cannot safely eliminate the need for lumbar puncture, with top non-linear models showing severe training optimism.

The rundown

The study used 306 infants from a Social Security Organization hospital in Khorasan Razavi, Iran, labeled by CSF culture via lumbar puncture, and limited predictors to routine non-invasive paraclinical markers from the EHR normalized with QuantileTransformer.

Evaluation used a leakage-safe nested cross-validation framework and SHAP values for global and local interpretability, with training-to-validation audits showing top-tier models clustered at outer-CV AUROC 0.74-0.76 and predictive signal distributed across urinary markers, age, and metabolic indicators.

What this doesn’t fix

High-sensitivity operation forces steep specificity loss, so model cannot be used as standalone rule-out to eliminate lumbar punctures and must remain a risk-tiering aid.