In the BIOTOPE cohort study, a machine learning score predicted 7-day hospitalisation and/or death in 2- to 59-month-old Malawian children with WHO-defined pneumonia seen in primary care, with AUROC 0.87 and precision-recall AUC 0.57 in external testing. Hospitalisation occurred in 14.3% of the training cohort and 12.1% of the testing cohort, and the ML model outperformed existing pneumonia risk scores.

Why It Matters To Your Practice

• Primary care clinicians often must decide quickly which children with pneumonia may worsen after the initial visit.

• This study suggests an ML-based score may improve early risk stratification beyond traditional pneumonia tools, especially when applied in low-resource outpatient settings.

• The model was developed for children meeting WHO pneumonia criteria, making the findings most relevant where WHO-based assessment guides frontline care.

Clinical Implications

• A better-performing triage score could help identify children who need referral, closer follow-up, or earlier escalation despite an initially ambulatory presentation.

• The outcome was hospitalisation and/or death within 7 days, so the model is most applicable to short-term disposition decisions rather than long-term prognosis.

• Because only one death occurred, this study supports admission-risk prediction more than mortality prediction.

Insights

• The training cohort included children from nine primary care centres recruited in 2022-2023; external testing used two centres with a 2016 cohort, providing a meaningful real-world generalisability check.

• WHO danger signs were present in 17.6% of the training cohort and 15.9% of the testing cohort.

• Calibration was reasonable but not perfect, with expected calibration error 0.16, underscoring the need for local validation before implementation.

The Bottom Line

• An ML score outperformed traditional pneumonia risk models for predicting 7-day hospitalisation in young children with WHO-defined pneumonia in Malawian primary care.

• For clinicians interested in AI, the practical takeaway is not replacement of judgment, but potentially better triage support in settings where early deterioration is hard to predict.

• Further external validation and clinical impact studies are still needed before routine adoption.