A Reusable Software Architecture for Deploying MATLAB Machine Learning Models in Android and iOS Mobile Apps, and Interactive Websites

Abstract

A novel and reusable software architecture was developed for productionizing a hospital trauma team activation model built in MATLAB. Previously, models were prototyped in MATLAB, then reimplemented, retrained, and retuned in Python and revalidated for deployment. This new architecture directly deploys the MATLAB model, expediting the release timeline and reducing cost. The roadblock preventing direct deployment of MATLAB models up to this point has been MATLAB’s incompatibility with mobile ARM CPUs found in smartphones and tablets. This new architecture uses a server with an x86-64 CPU to conduct all model computations in a MATLAB instance running on the server, rather than attempting to transplant the MATLAB model to mobile devices. It leverages a RESTful API on the server to accept prediction requests containing validated and formatted patient data, and the Java Engine API to interact with the MATLAB Engine in which the model is run on the patient data. Three frontends, two apps and a website, accept patient data in a form containing number fields and selectable options, send prediction requests to the RESTful API upon form completion, and display the model output and corresponding activation recommendation to the user.