In resource-constrained community settings, identifying which febrile children require referral remains a major unmet need. Current World Health Organization (WHO) danger signs have limited accuracy, resulting in missed severe illness and unnecessary referrals. Here we developed and validated clinical prediction models to support referral decisions using data from 3,405 children aged 1-59 months presenting with community-acquired acute febrile illnesses to seven hospitals across Bangladesh, Cambodia, Indonesia, Laos and Vietnam. Cambodian data were held out for external validation. The model using simple clinical parameters (sensitivity 74.7% (95% confidence interval (CI): 59.4-88.1); specificity 99.1% (95% CI: 97.7-99.7)) outperformed WHO criteria (sensitivity 55.5% (95% CI: 39.4-72.7); specificity 82.6% (95% CI: 77.1-87.6)) for identification of children at risk of severe disease (death or organ support within 2 days). Including either pulse oximetry or the host biomarker soluble TREM1 (sTREM1) increased sensitivity to 88.9% (95% CI: 76.7-97.8; pulse oximetry) and 89.2% (95% CI: 76.9-97.5; sTREM1), respectively. The pulse oximetry-based model achieved these gains with a threefold reduction in referral rates. These approaches appear cost-effective (pulse oximetry incremental cost effectiveness ratio (ICER) = $26.28; sTREM1 ICER = $196.46) and could improve triage for febrile illness in low-resource settings by enabling more accurate referral decisions. They warrant evaluation in community-based trials.