The study received Research Ethics Approval from Medical University of Silesia (no PCN/CBN/0052/KB/32/23). It is designed as a retrospective, observational, and multicenter study of AH patients. We used individual patient data collected for the study by Podsiadło et al. [5]. The data was collected from the International Hypothermia Registry (IHR), Hypothermia Life Support in Poland (HELP) Registry (https://rejestrhipotermii.ujk.edu.pl/), and the hospitals involved in that study. The data have been updated up to 1st April 2023.

The primary outcome was survival to hospital discharge.

Adult patients > 18yo < 90yo with accidental hypothermia, core temperature of ≤ 28 °C, and preserved spontaneous circulation at patient discovery and at rewarming commencement were included in the analysis. All patients underwent non-ECLS rewarming.

Exclusion criteria were: hypothermia associated with asphyxia (drowning, avalanche victims); cardiac arrest without return of spontaneous circulation; severe trauma with hemorrhagic shock and other non-hypothermia-related hemodynamic instability; implanted pacemaker. Also, patients with terminal illnesses and receiving palliative treatment were excluded from analysis.

The following data were collected: Patient age, gender, comorbidities, circumstances of hypothermia development (indoors/outdoors), vital signs at hospital admission (core temperature, heart rate, blood pressure, ventricular arrhythmias), occurrence of cardiac arrest with return of spontaneous circulation (ROSC) at any time of patient’s management before rewarming, mechanical ventilation before rewarming, and laboratory tests on admission (arterial blood gases with no temperature correction, acid–base balance, potassium and lactate concentration). When the non-invasive blood pressure was reported as „unmeasurable”, we assigned the value of 30 mmHg to such cases because this is the lowest systolic blood pressure measured by cardiac monitors commonly used in emergency departments.

Initially, we compared the surviving patients with the deceased group to identify variables associated with unfavorable outcome. The distribution of the variables was assessed using the Shapiro–Wilk test and QQ plots. Differences between groups were assessed with the Pearson’s chi-squared test, Mann–Whitney U test, or Student's t- test, depending on the variable’s distribution. In the descriptive statistics, variables are presented as mean and 95%CI or median and IQR. Qualitative variables are presented as absolute values and percentages.

We also performed a post-hoc analysis assessing the relationship between comorbidities (Charlson Comorbidity Index, CCI) and cooling circumstances to check whether the latter parameter could substitute for CCI [6]. Further, we developed a multivariable logistic regression. The potential risk factors were chosen based on previously published research in hypothermic patients: cooling circumstances, age, gender, core temperature (Tc), heart rate (HR), systolic blood pressure (SBP), diastolic blood pressure (DBP), mean arterial pressure (MAP), alpha-stat arterial blood gases at admission (temperature uncorrected), lactate and potassium concentrations, CCI, catecholamines administration, mechanical ventilation. We converted the PaCO2 originally measured with the alpha-stat method to pH–stat using the following formula [7]:

$$pH \, PaCO2 = alpha \times PaCO2 \times 10EXP(0.021(temperature - 37)) \,$$

Spearman correlation coefficients were determined, and only variables with correlations < 0.7 were included in the analysis. Univariate logistic regression was performed, based on which the independent variables with the highest OR/value of the Wald test were selected at the level of significance 0.25. We conduct a purposeful selection of variables as per Bursac et al. [8]. In the binominal regression model, significance of variables was determined at the 0.1 alpha level, while confounding was defined as a change in the remaining parameter of more than 20%. When covariates were non-significant and not cofounders, they were eliminated from the model. Model evaluation was based on the Hosmer–Lemeshow test, and Negelkerke R Square. The comparison of the models was based on the AUC and the coordinates of the ROC curve.

Finally, we calculated the cut-off values of risk factors with their sensitivity and specificity values. For statistical analysis we used StatsDirect 3.3.5 (StatsDirect LTD, Wirral, UK).