Kenya study finds protein biomarkers poorly distinguish bacterial, viral childhood pneumonia

In the internal validation dataset, the full model achieved an area under the receiver operating characteristic curve, or AUC, of 0.61, with 50% sensitivity and 68% specificity. The researchers concluded that the model did not distinguish the two groups with enough accuracy to guide clinical management, although performance improved in a separate analysis that excluded children whose infection could not be classified. The nested case-control study analyzed data from 457 children aged 2 to 59 months who met the 2013 World Health Organization definition of pneumonia and were admitted to Kilifi County Referral Hospital.

The data originated from a long-term surveillance program of respiratory syncytial virus and other childhood lower respiratory tract infections conducted between January 2007 and August 2020. The surveillance program sampled children younger than five years who were febrile or presented with sepsis or pneumonia. For this secondary analysis, children born prematurely, younger than one week or with a history of congenital heart disease were excluded.

Because the investigators used an existing dataset, the participants did not take part in the design or conduct of the predictive analysis. Investigators classified 228 children as having probable viral pneumonia because multiplex real-time PCR detected one or more viral pathogens associated with pneumonia in nasopharyngeal or oropharyngeal swabs, with no evidence of another infection. The remaining 229 children were placed in a possible bacterial group.

This group included 93 children with a positive blood culture, 12 with bacterial-virus co-infections and 124 whose cause remained unknown after blood culture, viral PCR and malaria testing. The broader classification reflected situations in which bacterial infection could not be excluded, but it also introduced uncertainty into the outcome used to train and evaluate the model. Indeed, 124 of the 229 children in the possible bacterial group, or 54% of that group, had an unknown etiology.

Among the 228 probable viral cases, the most frequently detected pathogens were respiratory syncytial virus A in 39 children, parainfluenza virus 3 in 36, human metapneumovirus in 35, influenza A in 34 and adenovirus in 28. The most common bacterial isolates from blood cultures were Escherichia coli and Streptococcus pneumoniae, each detected in 17 children, followed by Staphylococcus aureus in 11 and Haemophilus influenzae in eight. Severe pneumonia was present in 63% of the study population, including 72% of the probable viral group and 54% of the possible bacterial group.

This difference in severity was relevant because exploratory analyses found that the predictive performance of several biomarkers varied between severe and non-severe pneumonia. The researchers tested C-reactive protein, or CRP, procalcitonin, or PCT, and six less commonly evaluated plasma proteins: angiotensinogen, or AGT; serpin family A member 1, or SERPINA1; serpin family A member 3, or SERPINA3; paraoxonase 1, or PON1; histidine-rich glycoprotein, or HRG; and lipopolysaccharide-binding protein, or LBP. The six less commonly assessed proteins had been selected after an earlier discovery analysis within the surveillance cohort quantified 418 plasma proteins.

A modified Poisson regression model assessed associations while accounting for age, and backward selection identified variables for the multivariable model. The dataset was divided into training and validation subsets, with 70% used for training and 30% for validation. The investigators also repeated the complete model-building procedure through five-fold cross-validation stratified by etiology.

When possible bacterial cases were compared with probable viral cases, AGT had the strongest individual discriminatory performance, but its AUC was only 0.60, with a 95% confidence interval of 0.53 to 0.66. PCT and SERPINA1 each had an AUC of 0.58, while LBP had an AUC of 0.56. CRP, HRG, PON1 and SERPINA3 did not distinguish the groups beyond chance in the primary analysis.

A threshold of at least 7.9144 mean fluorescent intensity units for AGT had 49% sensitivity, 72% specificity, a positive predictive value of 64% and a negative predictive value of 58%. A SERPINA1 threshold of at least 8.9741 mean fluorescent intensity units had 29% sensitivity, 91% specificity, a positive predictive value of 75% and a negative predictive value of 86%. The high specificity of some individual markers came with correspondingly low sensitivity, and the biomarker distributions overlapped substantially between the two clinical groups.

In age-adjusted analyses of the training data, lower chest-wall indrawing, cough, convulsions, crackles, AGT and SERPINA1 were associated with pneumonia etiology. Once age, chest-wall indrawing and the other candidate predictors were considered together, however, only chest-wall indrawing remained significantly associated with etiology, and it was more common in the probable viral group. The model containing age and chest-wall indrawing produced an AUC of 0.72 in the training dataset.

That result declined to 0.61 in validation, while five-fold cross-validation gave a mean AUC of 0.64, with individual fold values ranging from 0.54 to 0.68. Chest-wall indrawing was retained in three of the five folds, cough and SERPINA1 in two folds each, and crackles and LBP in one fold each. No predictor was retained in four or more folds, indicating that variable selection was not stable across resamples.

The validation calibration slope was 0.5688, and the investigators described the predictions on unseen observations as overly extreme. A separate analysis excluded children with infections of unknown etiology and bacterial-virus co-infections, leaving more clearly defined bacterial and viral groups. In that restricted setting, chest-wall indrawing, cough, crackles and AGT at or above 7.9144 mean fluorescent intensity units were significant predictors after adjustment for age.

The resulting model achieved an AUC of 0.91 in the training set and 0.78 in the validation set. Validation sensitivity was 61%, and specificity was 96%. The model had a calibration intercept of 0.0696 and a calibration slope of 0.6915, indicating that its predicted risks were miscalibrated in the validation data.

The researchers said the improved performance suggested that unknown infections had attenuated the performance of the broader model. The restricted analysis was conducted in a selected subset with clearer laboratory-defined outcomes, rather than in the full population used for the primary model. Two previously published approaches were also applied without modification to the validation dataset.

A model using CRP of at least 5.70 milligrams per litre, fever of at least 38 degrees Celsius and rainy season, developed in hospitalized Indonesian children, achieved an AUC of 0.59, with a 95% confidence interval of 0.46 to 0.71. A second model using CRP of at least 72 milligrams per litre, fever of at least 38 degrees Celsius and the presence or absence of rhinorrhea reached an AUC of 0.74, with a 95% confidence interval of 0.64 to 0.85, a sensitivity of 55% and a specificity of 86%. These external comparisons showed mixed performance, but neither published approach nor the full biomarker model offered strong and consistent discrimination across the validation measures reported in the study.

Several limitations constrain interpretation. A negative blood culture did not exclude bacterial infection, and PCR detected only a defined panel of viruses, leaving other respiratory pathogens, including human bocavirus, unassessed. Nasopharyngeal or oropharyngeal swabs served as a proxy for the site of lung infection, potentially misclassifying the cause of pneumonia.

Biomarker results were recorded as mean fluorescent intensity rather than conventional concentrations, limiting direct comparison with studies that use concentration-based thresholds. The hospital-based population was also more severely ill than many children treated in outpatient settings, so the findings cannot be generalized to all children with pneumonia. The authors concluded that future work could address weaknesses in infection classification, stratify analyses by disease severity and further examine potential variation in biomarker performance by age.

Their overall finding was that the broad combination of clinical observations and protein markers did not provide sufficiently accurate internal validation for the primary model intended to distinguish possible bacterial from probable viral pneumonia.

Subject of Research: Internal validation of clinical and plasma-biomarker models for distinguishing possible bacterial from probable viral childhood pneumonia in Kilifi, Kenya

Article Title: Predictive modelling to differentiate possible bacterial and viral cases of childhood pneumonia in Kilifi, Kenya using protein markers and clinical data

Article References: Matuli, C., Waeni, J. M., Gicheru, E. T., Sande, C. J., & Gallagher, K. (2026). Predictive modelling to differentiate possible bacterial and viral cases of childhood pneumonia in Kilifi, Kenya using protein markers and clinical data. Frontiers in Medicine, 13, Article 1943481. https://doi.org/10.3389/fmed.2026.1943481

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1943481

Keywords: childhood pneumonia, bacterial pneumonia, viral pneumonia, plasma biomarkers, predictive modelling, internal validation, Kilifi County, Kenya, paediatric diagnostics, clinical signs, Predictive, modelling