Can Prediction Models Triage Trauma Patients More Accurately Than Clinicians?
Can Prediction Models Triage Trauma Patients More Accurately Than Clinicians?
Using a prospective cohort design, the aim of this study is to compare the performance of clinicians with prediction models in triaging trauma patients.
Title Can prediction models triage trauma patients more accurately than clinicians?
Background Nearly 5 million people die because of trauma each year. Hence trauma, defined as external injury combined with the body's associated response, is a major contributor to health issues globally. Trauma's time sensitive nature differentiates it from other health conditions. Therefore, prioritisation is crucial to treat the most urgent trauma cases first. This prioritisation is generally referred to as triage. Henceforth, formalized and non-formalized triage systems are distinguished between, where the former is defined as explicit triage consciously performed according to a priori determined criteria, and the latter as decisions on prioritisation by clinicians subconsciously taken in the absence of formalized triage.
There are two conceptually different methods currently used to triage trauma patients. Triage may be done by clinicians. Triage performed by clinicians may be affected by human factors such as tiredness and experience, yet may be viewed as the most morally appropriate method to triage trauma patients. Algorithms, such as statistical algorithms, may also be constructed for triage purpose. To predict the possibility of a certain outcome using statistical methods is called prediction modelling. Algorithms exclude human factors when triaging patients. However, algorithmic triaging may seem morally inappropriate due to the absence of human connection between the clinician and the patient.
Formalised triage systems are frequently used in trauma care in high income countries. Such systems are believed to have improved trauma care and hence trauma patient outcome. However, more than 90 per cent of all deaths due to trauma occur in low- and middle income countries, where formalised triage systems are uncommon to non-existing. Formalized triage systems may help strengthen trauma care in low - and middle income countries. However, research is lacking on whether algorithms or clinicians perform the most accurate in triaging patients.
Aim To compare the performance of prediction models with clinicians in trauma patient triage.
Method Study design Prospective cohort study
Setting A prospective cohort study will be conducted in India, a lower middle income country. India accounts for 20 per cent of global trauma deaths per year, hence efforts to strengthen trauma care in India are urgently needed. Data will be collected from four urban hospitals: one in Mumbai, one in Kolkata, one in Delhi, and one in Nellore. The collection of data began on the 11th of July, however only from the hospital situated in Kolkata. Once ethical permission has been formally provided data will be collected from the remaining hospitals and will continue until the desired study size is reached (see study size). Data is collected by full-time project officers at the mentioned hospitals.
Participants Individuals presenting to the emergency department (ED) because of trauma at any of the participating hospitals. The individuals are included in the study once they arrive to the ED. The project officers are the ones including individuals. The participating individuals are visited by the project officers if they are still hospitalized after 30 days. Otherwise, the follow up is done per phone.
Variables
Models:
RTS (Systolic blood pressure (SBP), Glasgow coma scale (GCS), and respiratory rate (RR)) The Gerdin et al. model (SBP, heart rate (HR), and GCS) GAP (age, SBP, and GCS) KTS (SBP, GCS, RR, and number of serious injuries) Clinical judgement (Clinicians assign a triage category to trauma patients by categorising them into four color coded groups. The groups are green, yellow, orange, and red. The colors are intended to represent trauma severity and how soon patients needs to be treated, with green corresponding to the least severe patient and red to the most severe patient.)
Other descriptive variables:
Sex, mechanism of injury
Outcome variables:
Mortality within 30 days
Study size 200 trauma patients with mortality as outcome and all patients that survived during the same time period. Mortality within 30 days is assumed to be 15 %, based on past studies. Thus, 1350 patients is the desired study size. The estimated time needed to collect the desired data is two months (four project officers, one per hospital. A project officer collect data from the first ten patients per shift, 5 days per week. Thus, a data collector collect data from 50 patients per week and from 200 patients per month. The total collected data per month will then be from 800 patients. Two months of collected data will be from 1600 patients. Consequently, it will take about two months to reach the desired study size.)
Quantitative variables Quantitative variables will be treated as in the original studies. In other words, the quantitative variables will be categorised or non linear methods will be used as they were in the original studies.
Statistical methods The performance of prediction models and clinicians will be evaluated and compared in terms of discrimination, calibration and reclassification. First, the scores obtained from the prediction models will be used to assign patients to green, yellow, orange or red categories. Discrimination will then be assessed by calculating the area under the receiver operating characteristics curve for both the original scores and the color groupings based on the scores. Calibration will be assessed by calculating the mortality in each color group, assuming a linear association between increased severity and higher mortality. Finally, reclassification will be assessed by comparing what patients end up in which group, depending on whether the grouping is based on prediction models or clinicians. Only patients with complete data will be analysed. All analyses will be conducted using the R statistical environment, using standard 95% confidence intervals and a 5% significance level.
Inclusion Criteria:
Exclusion Criteria: