This app demonstrates the behavior of a risk assessment instrument reflecting a patient's risk for deterioration at the time of initial evaluation by an ambulance crew on scene. The instrument is based on methods described in our research article A validation of machine learning-based risk scores in the prehospital setting . This user inferface contains no code to estimate risk scores, but rather calls to openTriage , an open-source API for estimating risk scores for use in clinical decision support systems. Please note that these scores have not been validated outside of Region Uppsala in Sweden, and that this tool is provided for demonstation purposes only. The software is provided 'as is' under the terms of the GPLv3 license without assumption of liability or warranty of any kind. Put plainly: If you use this on actual patients outside of the context in which the models have been validated, you could kill people.

A patient with median values for each predictor included in the models is described by default. Modify the model parameters in the sidebar and see how the risk assessment instrument reacts. If no choice is made for the multiple choice values, a missing/other value is assumed. Multiple choice options are sorted in order of descending average variable importance across all models.

The raw score is displayed for the patient at the top of the screen, and the relative position of the score with respect to all scores in a test dataset is displayed. The Likelihoods of each component outcome in the score and their percentile ranks are displayed beneath the graph. Finally, average SHAP values for each variable across the component outcomed are displayed to explain how the model arrived at the final score.

The API this front-end system employs may be accessed via a POST request to https://opentriage.net/predict/amb_refer. The API expects a JSON file with a specific format. You can download a test payload based on the currently selected predictors here:

Download test data

Diagnostics based on 37516 validation samples at score: