A single-center, retrospective observational study was conducted at Ramathibodi Hospital, a tertiary care and university hospital in Bangkok. This study was approved by the ethics committee of Ramathibodi Hospital, Mahidol University (COA. MURA2020/1791). In addition, the Ethics Committee waived each patient’s need for informed consent.

Patients aged 65 years and older who visited the ED at Ramathibodi Hospital between January 1, 2018, and December 31, 2019, were included in this study. Patients who were in cardiac arrest at the time of arrival, trauma patients, patients who denied resuscitation, patients who transferred from the ED, and patients with missing data were excluded from the analysis. Data between January 2018 and June 2019 were used to find factors associated with 28-day mortality and construct the prediction model for mortality or receiving lifesaving interventions. Data from July 2019 was used for validating the predictive model.

We collected the data from the Ramathibodi Hospital database via electronic medical records. The study variables were recorded for all eligible patients, including the baseline characteristics and potential clinical risk factors for mortality. Clinical factors included age, sex, comorbidities, ESI triage, vital signs at triage (heart rate [HR], systolic blood pressure [SBP], respiratory rate [RR], oxygen saturation [SpO2], and body temperature [BT]), the initial fraction of inspired oxygen (FiO2) at triage, NEWS at ED arrival, mental change, and diagnosis.

The primary outcome aimed to find variables related to 28-day mortality. The secondary outcome was to develop a screening risk assessment score for 28-day mortality and lifesaving intervention in the ED for geriatric patients.

The lifesaving intervention was defined as using an invasive mechanical ventilator, noninvasive positive pressure ventilation, and the use of vasopressors in the ED. The NEWS consists of seven physiological variables: SBP, HR, RR, BT, SpO2, use of supplemental oxygen, and level of consciousness. Furthermore, NEWS has been categorized into four levels, and high risk is a key threshold to emergency response (Additional file 1). The ESI is a five-level ED triage algorithm based on the acuity of patients’ problems and the number of resources used. ESI level-1 indicates a patient who needs immediate lifesaving intervention. ESI level-2 is defined by a patient who has high-risk conditions of deterioration. ESI level-3 indicates a patient with an urgent condition who must use more than one resource in the ED. ESI level-4 indicates who has non-urgent conditions and needs to use only one resource in ED. Finally, ESI level-5 indicates who has non-urgent conditions and does not require using the resource in ED (Additional file 2).

We calculated the sample size required to analyze predictor risks for 28-day mortality. Our previous hospital data on 28-day mortality in older patients who visited the ED showed 28-day mortality in 15 (4.03%) of 372 patients. Our calculation revealed that 1000 patients were required to provide an adequate sample size for the primary outcome in this study (95% confidence level, 5% alpha error).

For secondary outcome, the prevalence of lifesaving intervention in geriatrics was 5.2% [16]. We assume that the new model has eight candidate factors and approximately 25% of the variability of the Cox-Snell R squared statistic (R2cs). The sample size for the new predictive model was calculated by PMSAMPSIZE command [17] and needed 860 and 46 events for the development cohort.

All statistical analyses were performed using Stata version 16.1 (StataCorp, College Station, TX, USA). Frequencies or percentages described the categorical variables and mean, and standard deviation (SD) or median and interquartile ranges (IQR) were used to describe continuous variables, as appropriate. Exploratory analysis was done for all potential predictors of 28-day mortality using univariable logistic regression. Odds ratio (OR) with a P value was reported separately for each predictor variable. Then, multivariable logistic regression with stepwise model analysis was performed to identify independent predictors of 28-day mortality. Predictor variables with a P value of > 0.1 were sequentially eliminated from a logistic regression model.

The measure of each variable’s performance of 28-day mortality and lifesaving intervention was reported as the area under the receiver operating characteristic curve (AUROC) with a 95% confidence interval (CI). Predictors in the new model were selected by the highest AUROC, and multicollinearity was checked by the variance inflation threshold in the new model. Each final predictor was assigned scores based on each item’s logistic regression coefficient. Finally, the scores were calculated to generate the lifesaving intervention and mortality prediction score, referred to as the “Screening Tool Risk Score Assessment in the Emergency Department for Geriatric (S-TRIAGE)” in this article.

The predictive model was evaluated for its prognostic performance in terms of discrimination and calibration. A measure of discrimination was reported as AUROC. A measure of calibration was reported as Hosmer–Lemeshow goodness of fit statistics. The cutoff prediction scores for 28-day mortality and requiring lifesaving intervention were determined using sensitivity, specificity, and predictive probability. A calibration plot comparing score predicted risk versus observed risk was presented with 1000 replication bootstrapping. The predictive score was validated with internal validation and compared with the development set by chi-square. A P value of < 0.05 was considered to indicate statistical significance.