Background Rapid stratification of the risk of cardiac arrest is essential in the assessment of patients with isolated accidental hypothermia. Traditional methods based on measurement of core temperature are unreliable in the field. Behavioural observations have been used as predictors of core temperature and thus indirect predictors of cardiac arrest. This study aims to quantify the direct relationship between observed conscious level and cardiac arrest. Methods Retrospective case report analysis identified 114 cases of isolated accidental hypothermia meeting inclusion criteria. Level of consciousness in the acute assessment and management phase was classified using the AVPU system with an additional category of “Alert with confusion”; statistical analysis then related level of consciousness to incidence of cardiac arrest. Results All patients who subsequently suffered cardiac arrest showed some impairment of consciousness (p <  < .0001), and the risk of arrest increased directly with the level of impairment; none of the 33 fully alert patients arrested. In the lowest impairment category, Alert confused, a quarter of the 12 patients went on to arrest, while in the highest Unresponsive category, two thirds of the 43 patients arrested. Where core temperature was available (62 cases), prediction of arrest by consciousness level was at least as good as prediction from core temperature. Conclusions This study provides retrospective analytical evidence that consciousness level is a valid predictor of cardiac arrest risk in isolated accidental hypothermia; the importance of including confusion as a criterion is a new finding. This study suggests the use of consciousness alone may be at least as good as core temperature in cardiac arrest risk prediction. These results are likely to be of particular relevance to the management of accidental hypothermia in the pre-hospital and mass casualty environment, allowing for rapid and accurate triage of hypothermic patients.