Detailed paper information

Back to list

Paper title Machine learning algorithm to predict early complications after brain tumor surgery
Paper code P12
Authors
  1. Christiaan Hendrik Bas van Niftrik Universitätsspital Zürich/ Universität Zürich Speaker
  2. Frank van der Wouden University of California
  3. Victor Staartjes Universitätsspital Zürich/ Universität Zürich
  4. Jorn Fierstra Universitätsspital Zürich/ Universität Zürich
  5. Martin Stienen Universitätsspital Zürich
  6. Martina Sebök Universitätsspital Zürich/ Universität Zürich
  7. Tommaso Fedele Universitätsspital Zürich/ Universität Zürich
  8. Johannes Sarntheim Universitätsspital Zürich/ Universität Zürich
  9. Oliver Bozinov Kantonsspital St. Gallen
  10. Niklaus Krayenbühl Universitätsspital Zürich
  11. Luca Regli UniversitätsSpital & Universität Zürich
  12. Carlo Serra Universitätsspital Zürich/ Universität Zürich
Form of presentation Poster
Topics
  • SSNS-Neurosurgery
Abstract text Introduction:
Reliable preoperative identification of patients at high risk of developing early postoperative surgical complications occurring(EPC) within 24 hours after brain tumor surgery can improve patient safety and postoperative management. Statistical analysis using machine learning algorithms might generate models that better predict EPC than conventional statistical methods.
Methods:
Adult patients with suspected brain tumor, undergoing elective neurosurgery between June 2015 and May 2017 were retrospectively selected from an ongoing prospective database. EPC were categorized based on the Clavien-Dindo classification grading score and labeled 0 (no complication) or 1 (complication). This allowed us to train supervised machine learning algorithms, because these labels would “guide the learning process”. Different machine learning algorithms (Random Forest, Neural Network, Support Vector Machine and Gradient Boosting) were used to predict EPC using preoperatively available patient and surgical variables, being beforehand unclear which algorithm would work best for our data.. The models were created using a training dataset (random 80% of the cohort) and overfitting was prevented using a validation dataset (consisting of a random 50% sample of the residual 20% of the dataset).The performance of each model was derived by examining classification performance metrics on an out-of-sample test dataset (last 10% of the cohort). As a comparison, we also created a prediction model using conventional statistical methods.
Results:
EPC occurred in 174 (26%) of 668 patients included in the analysis. Gradient boosting machine learning algorithms provided the model best predicting the probability of an EPC. The model scored an accuracy of 0.70 with an AUC of 0.73 and a sensitivity and specificity of 0.80 and 0.67 on the test-set. The conventional statistical model showed inferior predictive power (test-set: accuracy: 0.59; AUC: 0.64; sensitivity 0.75; specificity of 0.53).
Conclusion:
Using gradient boosting machine learning algorithms, it was possible to create a prediction model superior to conventional statistics. Pathology and surgical related variables proved to be more powerful predictors of EPC than patients’ characteristics, contrarily to what was found with conventional statistic methods. With the future addition of more patients to our training data this model will increase in predictive power.