Article
Outcome prediction in aneurysmal subarachnoid haemorrhage is not improved by modern machine learning applications using traditional clinico-radiological features
Die Vorhersagegenauigkeit des Patientenoutcomes der aneurysmatischen Subarachnoidalblutung wird durch die Anwendung von modernen machine learning-Techniken mit traditionellen klinischen und radiologischen Kriterien nicht verbessert
Search Medline for
Authors
-
Meike Unteroberdörster
- Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland
-
Vince Istvan Madai
- Charité – Universitätsmedizin Berlin, CLAIM - Charité Lab for AI in Medicine, Berlin, Deutschland
-
Esra Zini
- Technical University Dublin, Dublin, Ireland; Charité – Universitätsmedizin Berlin, CLAIM - Charité Lab for AI in Medicine, Berlin, Deutschland
-
Sophie-Charlotte Brune
- Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland
-
Stefan Wolf
- Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland
-
Peter Vajkoczy
- Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland
-
Dietmar Frey
- Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland; Charité – Universitätsmedizin Berlin, CLAIM - Charité Lab for AI in Medicine, Berlin, Deutschland
-
Nora Franziska Dengler
- Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland
| Published: | June 26, 2020 |
|---|
Outline
Text
Objective: Numerous clinical, radiological and combined grading scales exist to describe the severity of aSAH and predict outcome. However, prognostic accuracy of scales is limited. We aimed to test whether the application of modern machine learning techniques by using traditional clinical and radiological features improves the predictive performance.
Methods: The single-institution database of 378 patients with aneurysmal subarachnoid hemorrhage was analyzed by inclusion of traditional clinico-radiological parameters: age, sex, pupillary state, presence of intracerebral hemorrhage (ICH), intraventricular hemorrhage (IVH), and/or subdural hemorrhage (SDH), aneurysm location (dichotomized anterior/posterior circulation) with respect to outcome prediction using the modified Rankin scale (mRS). Three different machine learning models were trained: ElasticNet logistic regression, Catboost boosted tree model and multilayer persepton (MLP) as a neuronal layer example. Data was randomly assigned to training and test purposes in a 4:1 relation. Models were tuned with 5 cross-validations, and the analysis was repeated in 50 shuffles. Variable ranking was calculated by weights for ElasticNets, by shap values for Catboost and by Taylor-decomposition for MLP.
Results: Prognostic ability of poor patient outcome (mRS 3-6) was similar between traditional ElasticNets logistic regression (mean area-under-the-curve 0.74) and modern machine learning methods (Catboost: 0.77 und MLP=0.76). All three models identified GCS (AUC 0.7) as the most important predictor. Other variables of importance were age, ICH, and aneurysm location.
Conclusion: Modern methods like boosted tree and neuronal networks do not inevitably lead to better performances in outcome prediction using traditional clinico-radiological features. This may imply that new biomarkers are needed to reach clinically relevant performance values.
Figure 1 [Fig. 1]
Figure 2 [Fig. 2]
