Clinical Impact of 16S rRNA Sequencing in Pediatric Infections: A Real-World Experience from a Tertiary Children’s Hospital

Presenter: Seungjae Lee (Korea, Republic of)

This retrospective study evaluated the diagnostic and clinical impact of 16S ribosomal ribonucleic acid (rRNA) sequencing in pediatric patients with suspected infections. The study aimed to assess the real-world value of 16S rRNA sequencing as an adjunct to conventional culture, particularly in situations where rapid antimicrobial decisions are required. The analysis included patients aged 18 years or younger who underwent clinically indicated 16S rRNA sequencing at Seoul National University Children’s Hospital between January 2020 and May 2025. Clinical specimens were tested using both conventional culture and 16S rRNA sequencing. Pathogens identified by either method were categorized as Group A (clinically actionable pathogens), Group B (confirmatory causative pathogens), or Group C (low clinical relevance or likely contaminants). A total of 301 patients contributed 359 clinical specimens. Overall, 16S rRNA sequencing detected pathogens in 121 samples (33.7%), compared with 94 samples (26.2%) identified by conventional culture. Detection rates varied according to specimen type. The highest detection rates with 16S rRNA sequencing were observed in abscess specimens (68.9%) and bronchoalveolar lavage samples (46.6%), while the lowest rate was observed in joint fluid specimens (11.1%). Treatment modifications based on 16S rRNA sequencing results occurred in 26 cases (7.2%), mainly among patients with culture-negative infections. In monomicrobial infections, discordance between 16S rRNA sequencing and culture differed significantly across pathogen groups (p<0.001). Discordance rates were 72.7% for Group A pathogens, 15.8% for Group B pathogens, and 46.4% for Group C pathogens. Among discordant cases, treatment modification occurred more frequently in Group A (93.8%) and Group B (100%) than in Group C (30.8%) (p<0.05).

Overall, 16S rRNA sequencing provided additional diagnostic information beyond conventional culture and had the greatest clinical impact in culture-negative infections, although its utility varied according to specimen type and pathogen category.

Evaluation of Serum C-Reactive Protein and Serum Procalcitonin in Infections in Children with Febrile Neutropenia during Cancer Chemotherapy

Presenter: Ajeitha Loganathan (India)

This study assessed whether C-reactive protein (CRP) or procalcitonin (PCT) is more useful for identifying infections in children presenting with febrile neutropenia (FN). A total of 103 FN episodes in 51 children were evaluated. Most episodes (72%) presented with fever alone. Microbiologically documented infections (MDI) accounted for 23% of episodes, clinically documented infections (CDI) for 7%, and fever of unknown origin (FUO) for 70%. Bloodstream infection was confirmed in 15.5% of episodes. Children with documented infections had higher CRP levels than those with FUO. Median CRP levels were 29.5 mg/L in the MDI group and 29.7 mg/L in the CDI group, compared with 7.89 mg/L in the FUO group (p=0.01). In contrast, PCT levels were similar across the three groups and did not show a significant association with infection. CRP levels were also significantly higher in children with bacteraemia, whereas PCT did not distinguish between those with and without bloodstream infection. Analysis of diagnostic performance showed that CRP consistently performed better than PCT. CRP was more accurate in identifying documented infections overall (AUC 0.68 vs 0.53) and showed notably better performance for detecting bacteraemia (AUC 0.82 vs 0.65). PCT showed limited value, particularly for identifying clinically documented infections (AUC 0.43).

Overall, CRP was a more useful biomarker than PCT for detecting infection and bacteraemia in children with febrile neutropenia.

Pathogen Identification and Antibiotic Prescription in Febrile Children Presenting to Hospital (2016–2026): The Perform and Diamonds International Prospective Cohort Studies

Presenter: Irene Rivero Calle (Spain)

This analysis examined antibiotic prescribing patterns in children presenting with febrile illnesses and explored factors associated with antibiotic initiation in patients with viral infections. The study used data from two prospective, multicentre, international studies: Personalised Risk Assessment in Febrile Illness to Optimise Real-Life Management (PERFORM; 2016–2020) and Diagnosis and Management of Febrile Illness using RNA Personalised Molecular Signature Diagnosis (DIAMONDS; 2020–2026). Children younger than 18 years with suspected infection were enrolled in the studies. For this analysis, patients adjudicated as having either bacterial or viral infections were included. Among 18,825 recruited participants, 3,382 children with bacterial infections and 3,225 with viral infections were analysed. Overall, a causative pathogen was identified in 3,601 patients, representing 54.5% of illness episodes. Pneumonia was associated with a particularly high rate of antibiotic initiation, with 78.9% of patients receiving antibiotics. However, pathogen detection in pneumonia was relatively low, with a bacterial or viral pathogen identified in only 53.9% of cases (n=691). Bronchiolitis and central nervous system (CNS) infections had the highest proportions of viral pathogens identified as the cause of illness, accounting for 79.5% and 51.8% of cases, respectively. Despite this, antibiotic use remained common, with antibiotics initiated in 56.7% of children with bronchiolitis and 63.2% of those with CNS infections. Multivariate regression analysis identified several factors associated with increased odds of antibiotic initiation in children with viral infections. These included the need for intensive care, the presence of at least one comorbidity, and age younger than 3 months. In contrast, C-reactive protein (CRP) levels and neutrophil counts were not associated with antibiotic initiation after adjustment for other clinical factors.

Overall, the findings suggest that opportunities exist to improve the precision of antibiotic prescribing in paediatric emergency settings, particularly among children with pneumonia or CNS infections, critically ill patients, young infants, and those with underlying comorbidities.

Prediction of Septic Shock in Neonates with Sepsis using Machine Learning: A Prospective Cohort Study

Presenter: Shiv Sajan Saini (India)

This prospective cohort study evaluated the use of machine learning to identify neonates with sepsis who are at increased risk of developing septic shock. Early recognition of shock in neonatal sepsis remains challenging, and the study aimed to determine whether routinely available clinical parameters could improve risk stratification. Consecutive inborn neonates with clinical features suggestive of sepsis were enrolled. Demographic information, maternal risk factors, and baseline clinical characteristics were recorded. Septic shock was defined using predefined clinical and hemodynamic criteria. The dataset was divided into a derivation cohort (70%) and a validation cohort (30%). A random forest classification model was developed using baseline variables, and a classification and regression tree (CART) model was subsequently created to provide a simpler bedside risk assessment tool. A total of 1,167 neonates were included, representing 1,286 sepsis episodes. Among the enrolled neonates, 199 (17%) developed septic shock. The random forest model demonstrated good predictive performance in the validation cohort, with an area under the receiver operating characteristic curve (AUC) of 0.73. The three most important predictors of septic shock were birth weight, systolic blood pressure, and gestational age. The CART model identified distinct risk groups. Neonates with a birth weight of at least 967 g and a systolic blood pressure of at least 49 mmHg represented the largest low-risk group, accounting for 76% of the cohort. Only 9% of neonates in this group developed shock. In contrast, neonates with a birth weight of at least 967 g, particularly those weighing less than 1,274 g, and a systolic blood pressure below 49 mmHg had a substantially higher risk, with 61% developing shock. Similarly, among neonates with early-onset sepsis and a birth weight below 870 g, 61% developed shock.

Overall, the study showed that machine learning-based risk stratification can identify clinically relevant shock-risk phenotypes in neonatal sepsis. Birth weight and early systolic blood pressure were key predictors and may help support earlier identification and management of high-risk neonates.

Diagnostic Accuracy of Host-Protein Bacterial/Viral Test in Febrile Children with Community Acquired Pneumonia

Presenter: Tobias Tenenbaum (Germany)

Community-acquired pneumonia (CAP) in children is often difficult to diagnose accurately because chest X-rays (CXR) cannot reliably differentiate bacterial pneumonia from viral disease. This uncertainty frequently leads to antibiotic use, even when the infection may not be bacterial. This analysis evaluated the performance of MeMed BV® (MMBV), a host-response blood test designed to distinguish bacterial infections, including bacterial co-infections, from viral or other non-bacterial illnesses. Data from seven previously published studies conducted between 2008 and 2020 were pooled and analyzed. The study included children younger than 18 years with suspected infection who had undergone chest X-ray evaluation. A total of 1,048 children were included. Based on expert adjudication, 56.8% had viral infections and 31.2% had bacterial infections, while 12.0% had equivocal MMBV results. More than half of all children (51.3%) received antibiotics. Among 281 children with radiological evidence of CAP, 57.7% had bacterial infection and 32.7% had viral infection. Despite this, 93.2% of children with CAP were prescribed antibiotics. MMBV demonstrated high diagnostic accuracy in the overall study population, achieving an area under the curve (AUC) of 0.95 in the pooled analysis and 0.91 in the meta-analysis. Performance was even higher among children with CAP, with AUC values of 0.97 and 0.95 in the pooled and meta-analyses, respectively. In children with CAP who had non-equivocal test results, MMBV showed a sensitivity of 95.2% and a specificity of 97.7% for distinguishing bacterial from viral infection.

Overall, MMBV accurately differentiated bacterial and viral infections in children undergoing chest X-ray evaluation and in those with CAP. The findings suggest that the test could help guide antibiotic prescribing and reduce unnecessary antibiotic use in pediatric pneumonia.

Machine Learning Model for the Prediction of Severe Disease Among Children Hospitalized for Acute Respiratory Infection

Presenter: Rohini R. Datta (Canada)

This study evaluated whether information available at hospital admission could be used to identify children with respiratory infections who are at risk of developing severe disease. Early recognition of high-risk patients may help guide triage and management decisions. Researchers analyzed data from a multicenter Canadian cohort of children aged 0 to less than 18 years who were hospitalized with acute respiratory infections between 2022 and 2023. Severe disease was defined as the need for non-invasive or invasive mechanical ventilation, extracorporeal membrane oxygenation, renal replacement therapy, cardiac arrest, or death. A total of 2,586 children were included, with a median age of 2.5 years. Overall, 21.3% of patients developed severe disease. Machine learning models were developed using information available at admission, including demographics, presenting symptoms, vital signs, and laboratory findings. Among the tested algorithms, extreme gradient boosting showed the best performance when using demographic and symptom data alone, achieving an area under the receiver operating characteristic curve (auROC) of 0.76 (95% confidence interval [CI]: 0.72–0.80). Age was identified as the most important predictor of severe disease. Model performance improved when patients were analyzed by age group. The highest predictive accuracy was observed in infants younger than 1 year, with an auROC of 0.89 (95% CI: 0.84–0.93). The corresponding auROC values were 0.79 (95% CI: 0.73–0.85) for children aged 1 to less than 5 years and 0.78 (95% CI: 0.70–0.85) for those aged 5 to less than 18 years. Adding laboratory markers further improved model performance. The voting classifier achieved an auROC of 0.78 (95% CI: 0.68–0.87), with a sensitivity of 0.89 (95% CI: 0.76–1.00) and a specificity of 0.47 (95% CI: 0.41–0.53).

Overall, the findings demonstrate that machine learning models using information available at admission can help identify children at risk of severe respiratory illness, with age-specific models showing the best predictive performance.

An Artificial Intelligence Classifier as a Screening Tool to Rule Out Otitis Media in Children

Presenter: Simo Nuuttila (Finland)

This study evaluated whether an artificial intelligence (AI) system could identify children without acute otitis media (AOM) using images of the tympanic membrane. Because AOM is one of the most common bacterial infections in childhood and is frequently misdiagnosed, tools that improve diagnostic accuracy could help reduce unnecessary medical visits and treatment. Researchers analyzed 793 tympanic membrane images obtained from 100 children aged 6–35 months. The images were collected in a primary care setting as part of a randomized, double-blind study conducted in Finland between 2006 and 2009. Several versions of an AI classifier were developed and tested to distinguish normal ears from those with abnormalities suggestive of otitis media. All four AI models demonstrated very high sensitivity for detecting an abnormal ear, ranging from 96% to 100%. The area under the receiver operating characteristic curve (AUC), a measure of overall diagnostic performance, ranged from 0.83 to 0.92 across the different models. After image processing methods were adjusted to improve image quality, the best-performing model achieved a sensitivity of 92% while improving specificity to 73%. This indicated that the classifier maintained a high ability to identify abnormal ears while becoming better at correctly recognizing healthy ears.

Overall, the findings suggest that AI-based analysis of tympanic membrane images may be useful for ruling out acute otitis media in children. Such an approach could potentially reduce unnecessary physician visits for children with normal ear examinations, although further studies are needed to evaluate its effectiveness in real-world, parent-led settings.

Diagnostic Accuracy and Beyond: A Sputum-Independent T Cell Activation-Based Assay for Paediatric Tuberculosis

Presenter: Alia Razid (Germany)

This prospective diagnostic accuracy study evaluated a simplified version of the T cell activation marker tuberculosis (TAM-TB) assay for diagnosing tuberculosis (TB) in children. The study enrolled children younger than 15 years with suspected TB across five low- and middle-income countries. Among 726 children with valid TAM-TB results, 191 (26%) had confirmed TB, 218 (30%) had unconfirmed TB, and 235 (32%) were classified as unlikely TB. Using culture-confirmed TB versus unlikely TB as the reference standard, the assay demonstrated a sensitivity of 63.7% (95% confidence interval [CI]: 54.1–72.4) and a specificity of 86.8% (95% CI: 81.8–90.6). Among 248 children who tested negative by both culture and Xpert MTB/RIF Ultra, 77 (31%) had a positive TAM-TB result. Of these, 55 (71%) had recent TB exposure, 47 (61%) were classified as unconfirmed TB, and 30 (39%) as unlikely TB. Notably, children in the TAM-TB-positive unlikely TB group showed disease trajectories similar to those with TB, and most had strong evidence of recent TB exposure. Among these 30 children, 10 (33%) received isoniazid preventive treatment and showed a significantly greater decline in interferon-gamma (IFNγ)-positive CD38 expression by one month.

Overall, the simplified TAM-TB assay demonstrated good specificity and may help identify children with TB who are missed by conventional microbiological tests. The assay also showed potential as a biomarker for monitoring treatment response.

Performance of a Multimodal Deep Learning Tool (“Strepapp”) Compared with Established Clinical Scoring Systems for Pediatric Gas Pharyngitis

Presenter: Rana Hamdy (United States of America)

This prospective study evaluated StrepApp, a multimodal artificial intelligence (AI)-based tool designed to assess suspected Group A Streptococcus (GAS) pharyngitis in children. StrepApp combines throat images with clinical information and was compared with commonly used clinical scoring systems, including the FeverPAIN and McIsaac scores. Children aged 3–17 years undergoing GAS testing at eight pediatric urgent care or emergency department sites in the United States were enrolled. Patients who had received antibiotics within the previous 48 hours were excluded. Throat images and clinical data were collected using a dedicated mobile application. A total of 4,789 children were included, and 41.6% tested positive for GAS. For the deep learning analysis, images from 1,137 patients were used, including 527 for training, 131 for validation, and 479 for testing. Clinical factors positively associated with GAS infection included fever, difficulty swallowing (dysphagia), recent exposure to GAS, scarlatiniform rash, and enlarged or tender cervical lymph nodes. Factors negatively associated with GAS included age greater than 15 years, cough, congestion or rhinorrhea, previous tonsillectomy, and abdominal pain. The diagnostic performance of each approach was assessed using the area under the receiver operating characteristic curve (AUC). FeverPAIN and McIsaac scores each achieved an AUC of 0.61. A symptom-based regression model developed from the study cohort achieved an AUC of 0.68. StrepApp demonstrated the highest performance, with an AUC of 0.91.

Overall, StrepApp showed substantially better discrimination of GAS pharyngitis than established clinical scoring systems, suggesting that AI-based analysis of throat images combined with clinical data may improve the evaluation of children with suspected GAS pharyngitis.

Deep Learning-Based Tonsil Size Estimation for Improved Oropharyngeal Examination in Pediatric Patients

Presenter: Rana Hamdy (United States of America)

This study developed an artificial intelligence (AI)-based computer vision tool to standardize tonsil size assessment using the Brodsky grading scale. Accurate and consistent evaluation of tonsil enlargement is important for diagnosing and monitoring head and neck infections in children, but traditional grading can vary between clinicians. Using a database of 6,635 oropharyngeal images, researchers created an automated system that identifies key anatomical landmarks, defines a standardized region of interest containing the tonsils, and estimates tonsil size on the Brodsky scale (grades 0–4+). The system also included an automated quality control process to exclude inadequate images. The anatomical landmark detection model achieved an accuracy of 97.7% across four key structures used to define the oropharyngeal region of interest: the uvula, tonsils, pharynx, and hard/soft palate. For tonsil size estimation, images from 28 patients were evaluated. Of these, 15 images met predefined quality criteria and were analyzed. Agreement between the AI-generated tonsil grade and clinician-assigned grade was observed in 11 of 15 cases (73%). The four disagreements occurred in borderline cases where grading was less clear.

Overall, the AI-based system was able to standardize image assessment and estimate tonsil size with good agreement to clinician grading, suggesting potential value in improving the consistency and reproducibility of tonsil size documentation.

44^th Annual Meeting of the European Society for Paediatric Infectious Diseases (ESPID 2026), June 1-5, Bologna and Online