médical IA

Pathog-PDx: a machine learning system to identify 22 pediatric respiratory pathogens from the electronic health record (Su 2026, npj Digital Medicine)

Published on May 30, 2026 · 12 min read

Dubin Su, Qun Chen, Ruizhi Xu and colleagues, led from the First Affiliated Hospital of Xiamen University (Wanshan Ning, corresponding co-author) with co-authors from Zhengzhou University, Nanjing University, Shenzhen Second People's Hospital and the University of Illinois Urbana-Champaign, publish in npj Digital Medicine on 29 May 2026 Pathog-PDx, a machine learning system (machine learning: a family of techniques in which the model learns a decision rule from labeled examples) that combines 42 clinical and laboratory features from the electronic health record to distinguish 22 pathogen subtypes responsible for respiratory infections in hospitalized children. Development cohort of 134,500 children across three clinical centers and two databases, prospective independent validation on 1,338 children, mean AUC (area under the ROC curve, a discrimination metric between positive and negative cases bounded between 0.5 and 1.0) of 0.88 across the 22 pathogens, 0.95 for influenza virus with sensitivity 0.88 and specificity 0.86, explicit handling of co-infections, and public deployment of a web-based decision support system. Four caveats nonetheless: all development centers are Chinese, the human clinical comparator is absent from the abstract, the CC BY-NC-ND license closes academic adaptation, and the gold standard used to label 22 classes deserves separate discussion — molecular biology with variable precision across pathogens.

The context

Acute respiratory infections are the leading cause of pediatric morbidity worldwide and the second cause of childhood mortality under the age of five according to WHO. When a feverish, symptomatic child is admitted, the treatment decision depends on the causative pathogen: antibiotics for Streptococcus pneumoniae or Mycoplasma pneumoniae, targeted antivirals for influenza or SARS-CoV-2, supportive care alone for respiratory syncytial virus (RSV) or a common rhinovirus. In practice the clinician decides before test results come back: multiplex PCR and culture usually take between 6 and 48 hours to return an actionable result, while the efficacy window of oseltamivir against influenza ideally closes within the first 48 hours. It is precisely this gap between decision and diagnosis that Pathog-PDx targets.

Earlier approaches have focused on one or a handful of pathogens — typically binary models trained to recognize RSV bronchiolitis, pneumococcal pneumonia, or Covid-19 — based on bedside clinical variables or radiology images. The leap to 22 pathogen subtypes simultaneously, with explicit handling of co-infections (a single child can host two or three agents at once), is much more recent and still rare in the published literature. The bet behind Pathog-PDx is that a stable, multi-organ signal aggregated from the electronic health record (EHR) — complete blood count, inflammatory markers, basic biochemistry, coded clinical signs, comorbidities — carries enough information to pre-orient diagnosis before specific microbiology tests resolve it.

The method

The study is led by Wanshan Ning (Institute for Clinical Medical Research, First Affiliated Hospital of Xiamen University), with co-corresponding authors Yungang Yang (Department of Pediatrics, same institution), Yaping Guo (School of Basic Medical Sciences, Zhengzhou University), and Jingjing Yang (Pulmonary and Critical Care Medicine, Xiamen). The co-authors share eleven affiliations, almost all Chinese (Xiamen, Nanjing, Zhengzhou, Shenzhen, Children's Hospital of Fudan University - Xiamen Branch), plus one affiliation at the Siebel School of Computing and Data Science of UIUC (Jiajun Fan). Published 29 May 2026 in npj Digital Medicine, DOI 10.1038/s41746-026-02818-9, under CC BY-NC-ND 4.0 (non-commercial, no derivatives) — a point we return to. Exclusively Chinese public funding (National Key R&D Program of China 2021ZD0201300 and 2022YFC2704300, NSFC 32400532 and 32570802, Fujian, Henan, and Xiamen programs). The authors declare no competing interests. The manuscript is released in an unedited "Article in Press" version, hence subject to revision.

The system is called Pathog-PDx (Pathogen Diagnostic System for Pediatric Respiratory Infections). It takes as input 42 variables from the EHR — child demographics, clinical signs coded on admission, history and comorbidities, complete blood count and routine biochemistry results, inflammatory markers (likely CRP, procalcitonin) — and outputs a probability for each of the 22 pathogen classes plus a joint probability for co-infections. The authors claim an interpretable architecture: without the full manuscript, one can reasonably assume a gradient boosting backbone (XGBoost or LightGBM) coupled to Shapley-value explanations (SHAP), now a standard pattern for tabular models in healthcare. The cited comparators are "conventional models" — presumably multinomial logistic regression, random forest, and a clinical-rules baseline.

The development dataset covers 134,500 hospitalized children, aggregated across three Chinese clinical centers and two databases. The independent prospective validation set covers 1,338 children from a fourth facility — a strong methodological point because this cohort was unseen by the model during training and was collected after the model was frozen. The gold standard used to label the 22 pathogens is not detailed in the abstract but plausibly combines, according to current Chinese practice, respiratory multiplex PCR, bacterial culture, and serology depending on the agent — a mixture whose analytical sensitivity varies significantly from one pathogen to another.

The results

On the prospective validation cohort, Pathog-PDx reaches a mean AUC of 0.88 across the 22 pathogens. For influenza virus, taken as a clinically frequent example, the AUC is 0.95, sensitivity 0.88, and specificity 0.86. The authors state that the model outperforms conventional approaches on both single and mixed infections — the central practical difficulty in pediatrics, where RSV and a bacterial superinfection often coexist. The system has been deployed as a public web service at pathogpdx.zzu.edu.cn, hosted on Zhengzhou University servers, and provides for each case a prediction "ahead of conventional test results" with an explicit goal of therapeutic pre-orientation.

Clinical translation. To anchor the numbers on 1,000 children admitted with a febrile respiratory syndrome of whom about 100 truly have influenza (typical seasonal prevalence in East Asia): at the announced influenza performance, the model would correctly detect 88 of the 100 true flu cases (12 false negatives), at the cost of 126 false positives among the 900 non-flu cases (sensitivity 0.88, specificity 0.86). Across all 22 classes, a mean AUC of 0.88 indicates strong global discrimination but almost certainly masks substantial per-pathogen variability: rare agents with weak biological signatures (adenovirus, parainfluenza type 4, certain seasonal coronaviruses) are likely classified less well than influenza or pneumococcus. For a pediatrics ward handling a seasonal surge, the tool replaces neither multiplex PCR nor clinical judgment, but it can credibly guide empirical prescription of oseltamivir or decisions to place a child in respiratory isolation while awaiting biological confirmation.

What works well

The prospective independent validation actually exists, and that is rare. Most published clinical ML models in this space settle for internal validation by cross-validation or a retrospective validation on a separate site. Here, 1,338 children were prospectively enrolled into a cohort specifically built to evaluate the model after training — the model had to classify patients it had never seen, in a period subsequent to the collection of the development cohort. This is the most demanding evaluation procedure short of a randomized controlled trial. Add to that the public deployment of a working web service: the leap "from paper to usable prototype" is concretely cleared, which remains a minority outcome in clinical ML literature.

The multi-class ambition with explicit co-infection handling attacks a real clinical problem. Binary "RSV yes/no" models published in series over the past decade do not help the clinician who must rank risk between RSV, influenza, metapneumovirus, pneumococcus, and Mycoplasma in the same child. A model that outputs a joint probability over 22 classes and explicitly addresses co-infections changes the practical utility: the clinician gets a ranked probability table that matches the way she actually reasons. This labeling and evaluation discipline is technically harder than binary approaches and deserves credit.

The input variables are available everywhere. The 42 features used by Pathog-PDx are routine EHR fields: demographics, coded clinical signs, complete blood count, CRP, basic biochemistry. None require specialized imaging, biopsy, or expensive molecular panels. This means that, subject to recalibration and external revalidation, the model is in principle transposable to a European or North American academic hospital — even to lower-resource centers with minimal standard lab capability. The variable choice is thus consistent with a generalization ambition, even if that generalization remains to be demonstrated.

What works less well

All development centers are Chinese — clear population bias. The three clinical centers and the two databases providing the 134,500 training children are Chinese (Xiamen, Zhengzhou, Shenzhen, Children's Hospital of Fudan University - Xiamen Branch). Pediatric respiratory epidemiology varies substantially from continent to continent: the prevalence of Mycoplasma pneumoniae went through an exceptional surge in China in 2023-2024 that shifted class distributions, the bacterial/viral ratio depends on vaccination coverage (PCV13, MMR, influenza) and therefore on the health system, and certain regionally endemic pathogens (typically avian metapneumovirus or specific seasonal coronaviruses) are over-represented in given countries. The failure mode here is the classic population bias: the model learns a conditional distribution "clinical signs + biology → pathogen" that is not universal. Proving it works on a feverish child in Marseille, Boston, or Nairobi would require non-Chinese external validation — absent here.

The human comparator is absent from the abstract. Pathog-PDx is compared to "conventional models" — likely multinomial logistic regression and random forest — which it outperforms. The real clinical question is different: with the same 42 variables, does a senior pediatric inpatient physician reach equivalent discrimination? An infectious-disease pediatric consultant on top? Without that human reference, the reported numbers flatter the tool against algorithmic baselines but do not let one decide whether pre-orientation by Pathog-PDx adds incremental value over standard clinical judgment. This is the failure mode of comparator bias by omission: the most relevant baseline — an experienced clinician with the same data — is invisible. Compounding this, the abstract is silent on per-class performance: a mean AUC of 0.88 across 22 pathogens nearly mechanically implies that at least a handful of subtypes fall below 0.75, and identifying which ones (likely rare agents with weak biological signatures) is decisive for deciding what to trust in practice.

The CC BY-NC-ND license closes academic adaptation and surfaces a gold-standard question. The manuscript is published under CC BY-NC-ND 4.0 — not the standard Nature open-access CC BY 4.0. The NC blocks commercial use (legitimate); the ND blocks derivative works, complicating academic adaptation of the code, weights, or even figures for replication or transposition to another setting. Add a substantive methodological question: the gold standard used to label the 22 classes is not detailed in the abstract. Yet the analytical sensitivity of respiratory multiplex PCR varies across pathogens (excellent for influenza, weaker for Mycoplasma), bacterial culture systematically underestimates fragile agents, and late serology only covers certain diagnoses retrospectively. If some classes are labeled with a less reliable standard, that is label noise injected into training, which can either degrade performance on the affected agent, or — worse — create an illusion of good classification by teaching the model the biological signature of the test rather than that of the pathogen. Without extracting these details from the full manuscript, a shortcut learning failure mode remains to be formally ruled out.

What this changes

For the AI-in-health research community, Pathog-PDx fits a recent wave returning to tabular methods for differential-diagnosis questions, after several years of heavy investment in imaging. The merit is pushing the standard to 22 classes with co-infections: a binary RSV yes/no benchmark no longer suffices; one now expects a ranked table of probable pathogens. Groups working on adjacent problems — neonatal sepsis, pediatric emergency sepsis, meningitis differential diagnosis — will find here a pipeline pattern (routine EHR variables, prospective validation on an independent cohort, deployment as a web service) they can replicate. The remaining question is replication by a European or North American team on their own data, which will tell whether the signal is portable or was specific to the Chinese respiratory epidemiology of the training period.

For pediatricians, the most credible near-term use is not replacing multiplex PCR but therapeutic pre-orientation in the first hours of hospitalization — the moment the clinician decides whether to give oseltamivir, whether to isolate an RSV child, whether to empirically cover bacterial co-infection. An AUC of 0.95 on influenza is a credible figure for that use case. Across other classes, lacking per-pathogen detail, prudence treats Pathog-PDx as a hypothesis aid, not a diagnosis. No system of this kind is currently approved by Haute Autorité de Santé in France, CE-marked as Software as a Medical Device, or cleared by the FDA in the United States. Real-world clinical use outside China will require local revalidation and regulatory certification, which are not the subject of this article.

For patients and their families, the useful takeaway is that early identification of the pathogen behind a child's bronchiolitis or pneumonia is becoming technically faster thanks to models that exploit the routine tests already available on admission. This does not replace the clinical pathway: a feverish child who is worsening must be seen by a physician, and the treatment decision remains human. Until such a tool is validated and certified for use outside China, the right reflexes remain prudent antibiotic use (antibiotics do nothing on viruses), annual influenza vaccination for at-risk children, and RSV immunization, now recommended for infants in several European countries.