OBJECTIVE: To determine the prevalence of hypoxaemia among under-five children admitted to hospital with clinical severe pneumonia, and to assess the performance to diagnose hypoxaemia of models based on clinical signs. METHODS: We conducted a hospital-based survey in a district hospital from Southern Mozambique. RESULTS: A total of 825 children were recruited after obtaining an informed consent. The prevalence of hypoxaemia on admission was 27.9%, and 19.8% of these children died (OR compared to non-hypoxaemic children 3.22, 95%CI 1.98 - 5.21, p<0.001). The model with larger area under the ROC curve (AUC-ROC) to predict hypoxaemia included cyanosis or thoracoabdominal breathing or respiratory rate >/= 70 breaths per minute. None of the models performed well when tested in different case scenarios of oxygen availability through mathematical modelling, with over 50% of hypoxaemic children not receiving oxygen even in favourable case scenarios. CONCLUSIONS: Clinical signs alone or in combination are not suitable to diagnose hypoxaemia. The use of pulse oximeters should be strongly encouraged. This article is protected by copyright. All rights reserved.