Article
Predicting functional outcome after aneurysmal subarachnoid hemorrhage using machine learning
Search Medline for
Authors
Published: | June 9, 2017 |
---|
Outline
Text
Objective: The pathogenesis leading to poor functional outcome after aneursysmal subarachnoidal hemorrhage (aSAH) is multifactorial and not fully understood. Microcirculatory dysfunction, assumed to play a key role, can be indirectly quantified with CT perfusion (CTP) imaging. We evaluated a machine learning approach, based on easily determinable features on admission, including CTP, to predict functional outcome.
Methods: Out of 614 consecutive subarachnoid hemorrhage patients (2009-2015), 351 were included (aSAH, ≥1 CTP within 72h of ictus and documented modified Rankin scale (mRS) scores after 6 months; n=5 excluded due to corrupt CTP datasets). The dataset was split in a training (n=264) and test dataset (n=87). A Random Forest model was trained on age, sex, WFNS and modified Fisher scores on admission, bihemispheric MTT and rCBF maximum, mean, maximum-mean and standard deviation (sd) to predict dichotomized mRS (≤2; >2). Feature importance was derived from the trained model. The model performance was evaluated on the test dataset and Receiver-Operator-Characteristic (ROC) analysis was performed.
Results: Sensitivity and specificity when predicting a mRS score >2 in the test dataset were 68.8% and 81.8%, respectively. Accuracy was 77.0%. Positive and negative predictive value was 68.8% and 81.8%, respectively. The area under the ROC curve was 0.83. MTTSD, WFNS and age were found to be most important.
Conclusion: A Random Forest model trained on easily determinable features on admission can predict functional outcome after aSAH with good accuracy, sensitivity and specificity. Our results underline the role of CTP. A novel finding is the importance of MTTSD.