Creation of an Interactive Prognostic Model for Myxoid Liposarcoma Overall Survival Using Machine Learning Methods

Abstract

Background: Prior studies have evaluated multiple risk factors for poor prognosis in rare orthopaedic malignancies like myxoid liposarcoma. However, the ability to combine multiple risk factors in a provider friendly method to accurately predict patient outcomes remains limited. Questions:

How well do machine learning models predict overall survival in myxoid liposarcoma cases obtained from a national registry? Do these models retain predictive accuracy when validated on an institutionally derived dataset? To produce a publicly available and provider-friendly user interface for model interaction. Patients and Methods: Ten predictive machine learning models for five-year overall survival were developed using a training dataset from the Surveillance, Epidemiology, and End Results (SEER) national registry consisting of 628 patient cases with myxoid, mixed, or round cell liposarcoma. Survival outcomes out to five years were available for included cases. The four best models were further evaluated on validation datasets from both the national database and an institutionally derived dataset (n=70) of myxoid liposarcoma cases. The most predictive model was then deployed in a clinician friendly interface on the public website https://mgh-ortho.shinyapps.io/MyxoidApp/. Results: How well do machine learning models predict overall survival in myxoid liposarcoma cases obtained from a national registry? The most predictive model evaluated was support vector machine. When evaluated on the SEER validation set it showed an area under the receiver operator characteristic curve (AUC) of 81.5% sensitivity of 77.5%, positive predictive value (PPV) of 76.9%, and negative predictive value (NPV) of 85.7%. Notably the random forest, averaged neural network, and oblique random forest models all had AUC’s above 80%. Do these models retain predictive accuracy when validated on an institutionally derived dataset? Out of the four models with the highest AUCs from the validation set, when evaluated on an institutional dataset of 70 patients the random forest model had the highest AUC of 78.2%. Its accuracy was 85.7%, PPV was 86.4%%, and NPV was 75.0%. To produce a publicly available and provider-friendly user interface for model interaction. An interactive version of the random forest model was deployed on the website https://mgh-ortho.shinyapps.io/MyxoidApp/ for clinical or research uses. Conclusion: Machine learning models were developed to estimate overall survival in myxoid liposarcoma patients. The model performed at a high level in terms of operational characteristics. A user- friendly model interface was deployed online which can easily be utilized for clinical or research use.