Vorhersage des postoperativen Resektionsstatus bei Meningeomen durch maschinelles Lernen auf der Grundlage klinischer Merkmale

The extent of achievable resection in the surgical treatment of meningiomas is of high importance for the further prognosis and treatment planning. Subtotal resections are often associated with recurrent tumor growth and thus the risk of disease progression. For further treatment planning, it is therefore important to determine as early as possible whether a meningioma can be completely resected or not. The aim of our study is to predict the postoperative resection status of meningiomas using machine learning based on clinical features.

Our retrospective, IRB-approved study is based on a cohort of 138 patients. A gross total meningioma resection was performed in 107 cases and a subtotal resection in the remaining 31 cases. Approximately 20 clinical features, such as the location of the tumor, its shape and volume, were analyzed with regard to the predictability of the achievable postoperative resection status. The features to be included in the models were selected using recursive feature elimination. A total of 14 machine learning algorithms, including a neural network, were trained and tested to predict postoperative resection status. All models were developed a total of 100 times, each time with new training data. The model performance was subsequently determined for each of the 100 runs using new independent test data.

We achieved our best results using a neural network containing only 4 clinical features. This model exhibits a mean AUC of 88.7 %, a mean accuracy of 87.4 %, a mean sensitivity of 92.8 %, a mean specificity of 68.5 %, a mean Cohen's kappa of 62.6 % and finally a mean positive and negative predictive value of 91.3 % and 75.1 % respectively. Very good results were also achieved using a generalized linear model. Thus, our models show a very high performance in the preoperative prediction of the achievable postoperative resection status of meningiomas. We were able to identify clinical features that are of particular importance regarding the achievable postoperative resection status. In addition, we have determined in which cases it is particularly difficult to make a precise preoperative statement.

Our results show that the achievable postoperative resection status of meningiomas can be predicted very accurately using machine learning algorithms based on only a few clinical features. In the near future, such machine learning based methods may enable physicians to accelerate further treatment planning.