Background Children with sepsis in low- and middle-income countries (LMICs) face high rates of in-hospital mortality and prolonged hospitalisation, yet most clinical prediction models for poor outcomes require laboratory tests or imaging that are unavailable at primary care facilities in these settings. We aimed to develop and evaluate a point-of-care machine learning model for predicting poor clinical outcome in children under five years admitted with suspected sepsis in a sub-Saharan African LMIC setting, using only features collectable at first clinical contact without laboratory or imaging infrastructure. Methods We used a synthetic research dataset derived from a real-world prospective cohort of 3,837 children aged 2 to 60 months admitted with suspected sepsis across six hospitals in Uganda, made available through the 2024 Pediatric Sepsis Data Challenge (Borealis/Sepsis CoLab). After exclusion of one record with a data entry error, 2,685 records were included. The composite outcome, defined as in-hospital death or length of stay exceeding five days, was present in 911 cases (33.9%). We selected 45 point-of-care clinical features collectable without laboratory or imaging resources, including vital signs, clinical signs, nutritional anthropometry, symptom history, and comorbidity status. Five additional features were engineered: WHO age-specific tachypnoea flag, severe hypoxia flag (SpO₂ <90%), severe acute malnutrition by MUAC (<115 mm), moderate acute malnutrition by MUAC (115-125 mm), and Blantyre Coma Scale total score. Four models were trained on an 80% development set with stratified five-fold cross-validation and Optuna hyperparameter optimisation, and evaluated on a 20% holdout test set. SHAP (SHapley Additive exPlanations) analysis using a LinearExplainer was performed on the best-performing model to characterise feature contributions. Results Logistic regression achieved the highest test set AUC-ROC of 0.686 (95% CI 0.637–0.734) after post-hoc Platt scaling calibration, with AUC-PR of 0.557 (95% CI 0.486–0.637) and Brier score of 0.199 (95% CI 0.182–0.217). The Hosmer-Lemeshow test confirmed good calibration after Platt scaling (chi-squared = 7.48, df = 8, p = 0.486) with an expected-to-observed ratio of 1.027. All four models converged to a narrow pre-calibration AUC-ROC range of 0.680 to 0.689, indicating a data-determined performance ceiling. At the Youden J-optimised threshold of 0.33, the calibrated model achieved sensitivity of 0.571, specificity of 0.769, and F1 score of 0.565. SHAP analysis identified severe acute malnutrition (mean |SHAP| = 0.193), weight (0.128), WHO tachypnoea (0.111), and moderate acute malnutrition (0.107) as the four strongest predictors globally. These results are directly comparable to the best published Bangladesh-specific treatment failure model (AUC-ROC 0.691 on derivation; Mamun et al., PLOS Global Public Health, 2023), which was derived in a hospital setting with clinical monitoring infrastructure. Conclusions A logistic regression model using exclusively point-of-care clinical features achieves discrimination equivalent to published hospital-derived LMIC sepsis and treatment failure models, without requiring any laboratory tests or imaging. Nutritional status, as measured by MUAC and weight, dominates prediction alongside respiratory signs, supporting the inclusion of a brief nutritional assessment in any LMIC triage tool for unwell children. The model is suitable for further prospective validation and potential adaptation into a paper-based scoring system for deployment at primary care and field-level facilities. Keywords: pediatric sepsis, clinical prediction model, point-of-care, LMIC, machine learning, logistic regression, SHAP, malnutrition, Uganda, triage