Article
Identification of predictors for an unfavourable outcome after surgical treatment of chronic subdural haematomas in 755 patients using machine-learning
Identifikation von Prädiktoren für ein schlechtes Behandlungsergebnis nach der chirurgischen Evakuation von chronischen Subduralhämatomen in 755 Patienten mittels maschinellen Lernens
Search Medline for
Authors
Published: | June 26, 2020 |
---|
Outline
Text
Objective: Chronic subdural hematomas (cSDH) are expected to become the most frequent neurosurgical disease by the year 2030. Although often perceived as a "benign" condition, considerable rates of mortality and poor outcome have been reported. We therefore evaluated factors associated with an unfavorable outcome after surgical treatment of cSDH patients by developing a predictive model using machine-learning.
Methods: Consecutive patients treated for cSDH with surgical evacuation between 2006-2018 at a single institution were retrospectively analyzed. Potential demographical, clinical, imaging and laboratory predictors were assessed and a decision-tree predicting unfavorable outcome (GOS 1-3) was subsequently developed using the Classification and Regression Tree (CART) algorithm. Hereby, the complexity parameter was set at 0.02 and at least 25 observations were required at every split or node. Out-of-sample model performance was evaluated using repeated cross-validation (5-fold with 200 repetitions).
Results: 755 eligible patients were analyzed. The median age was 75 (IQR 68-81) years and 69% were males. The mortality rate was 1.6% and the rate of unfavorable outcomes was 14.3%. The developed decision-tree to predict unfavorable outcome had 5 splits and included the following 4 clinical variables (in descending order of calculated importance): GCS, comorbidities, Hb, and age. After cross-validation, the following model performance metrics were obtained: a model accuracy of 0.88 (0.85-0.90), sensitivity of 0.35 (0.19-0.51), and specificity of 0.96 (0.94-0.99).
Conclusion: GCS, comorbidities, Hb, and age were identified as the most important clinical predictors for an unfavorable outcome in cSDH patients after surgery. The developed model was simple and still displayed a high accuracy and very high specificity, the sensitivity was however rather low. Our results might help clinicians to better assess the prognosis in patients with cSDH.