Machine Learning–Based Prediction of 6-Month Postoperative Karnofsky Performance Status in Patients with Glioblastoma: Capturing the Real-Life Interaction of Multiple Clinical and Oncologic Factors

Abstract

Ability to thrive after invasive and intensive treatment is an important parameter to assess in patients with glioblastoma multiforme (GBM). Karnofsky Performance Status (KPS) is used to identify those patients suitable for postoperative radiochemotherapy. The aim of the present study is to investigate whether machine learning (ML)-based models can reliably predict patients' KPS 6 months after surgery. A cohort of 416 patients undergoing surgery for a histopathologically confirmed GBM were collected from a multicentric database and split into a training and hold-out test set in an 80:20 ratio. Worsening of KPS at 6 months after surgery (compared with preoperative KPS) occurred in 138 patients (33.2%). Relevant preoperative, intraoperative, and immediately postoperative variables were selected by a recursive features selection algorithm (Boruta) and used to build 2 ML-based predictive models. A random forest classifier and a random forest regressor were trained to predict 6 months postoperative KPS as a categorical (worsening vs. stable/improving) and continuous variables; they achieved, respectively, an area under the curve of 0.81 (95% confidence interval, 0.76-0.84) and a mean absolute error of 4.4 (95% confidence interval, 4.0-4.7). Leveraging the predictive value resulting from the combination of independent variables, the random forest classifier outperformed conventional statistics (area under the curve improvement of+21%). Two robust ML-based prediction models were successfully trained and internally validated. Considerable effort remains to improve the interpretation of the results when these predictions are used in a patient-centered care context.