RSS-Feed abonnieren
DOI: 10.1055/s-0040-1703312
Outcome Prediction of Acute Intracranial Hemorrhage based on Computed Tomography: Comparison of Conventional Semantic Assessments and AI-backed Evaluation of High-end Image Features
Publikationsverlauf
Publikationsdatum:
21. April 2020 (online)
Zielsetzung Intracranial hemorrhage (ICH) requires prompt diagnosis and treatment to optimize patient outcomes. We hypothesized that imaging-based machine learning algorithms can automatically analyze non-contrast computed tomography scans (NECT) and predict clinical outcome of ICH patients with higher discriminatory power than conventional assessments of clinical data and imaging markers.
Material und Methoden 312 NECTs with acute spontaneous ICH between 2014–2019 were retrospectively analyzed from the database at a tertiary university hospital. Clinical patient outcome (modified Ranking Scale, mRS) at hospital discharge was dichotomized into good outcome (mRS 0-3) and poor outcome (mRS 4-6). Predictive performance of a logistic regression model based on clinical data and conventional imaging markers (Model 1) was compared to a random forest machine learning approach utilizing high-end image features (Model 2). A third model was set up to integrate both conventional markers and high-end image features (Model 3). All models were tested in a 5-fold cross-validation approach with separate training and validation data sets.
Ergebnisse Model 1 achieved a receiver operating characteristic (ROC) area under the curve of 0.77 (95% CI [0.72; 0.82]) vs. 0.81 (95% CI [0.77; 0.86]) in Model 2 and 3. Furthermore, AI-based Models 2 and 3 showed significantly higher sensitivities (80% vs. 66%) compared to conventional assessments at 70% specificity cut-off points. Overall, results of Model 3 were similar to results Model 2.
Schlußfolgerungen Compared to conventional markers in a logistic regression model, AI-based evaluation of high-end image features provided significantly higher discriminatory power in predicting functional outcome. High-end imaging markers captured predictive information of conventional image assessments; an additional integration of conventional markers did not improve results.
#