Prognostic Survival Model Of Myxoid Liposarcoma Patients Using Machine Learning

Predicting outcomes for oncology patients is largely generalizable, often by cancer stage, and individualization of outcome prediction remains limited. Machine learning applications offer a resource to individualize patient outcome predictions in a manner usable for clinicians. We sought to develop an interactive application to predict survival outcomes in soft tissue sarcoma patients. The national oncology registry and an institutional database were utilized for creation and validation, respectively, of a 5-year survival prediction model of myxoid liposarcoma patients. Multiple machine learning algorithms were applied to the national registry, and the most accurate algorithms for predicting patient survival were then tested and validated on an institutional database. The most predictive model was then deployed on an interactive, online application. When developed and tested on a national database, four out of ten predictive models displayed area under the receiver operator characteristic curve (AUCs) of over 80%. The best model, a support vector machine, had a positive predictive value (PPV) of 76.9% and a negative predictive value (NPV) of 85.7%. When these 4 models were validated on the institutional dataset, the best performing model was a random forest with an AUC of 78.2% (95% confidence interval [CI], 64.1-92.2), PPV of 86.4%, and NPV of 75.0%. An interactive web-application is now available online for interfacing with the predictive model. Using machine learning models, individual outcome prognostication that was traditionally best defined using clinical experience and gestalt can now be quantified and modeled. Our publicly available survival model can be utilized by researchers and clinicians to estimate 5-year survival probabilities in patients with myxoid liposarcoma diagnosis.