The widely-available tools can be easily implemented and adapted to different languages and systems, the researchers say.
Machine-learning models that take into account both clinical and imaging features of patients with aneurysmal subarachnoid hemorrhage may help identify those at high risk for developing delayed cerebral ischemia, according to a new study.
“Delayed cerebral ischemia (DCI) is a severe complication in patients with aneurysmal subarachnoid hemorrhage,” lead author Lucas Alexandre Ramos, MSc (Amsterdam University Medical Centers, Amsterdam, the Netherlands), told Neurovascular Exchange in an email.
“The selection of patients with a high risk of developing DCI may improve patient outcome as well as reduce the costs related to futile intensive-care monitoring,” he continued. “Even though many prediction models have been described in the literature, most of them have shown very low prediction accuracy.
“Previously, only a few studies have explored machine learning and deep learning methods with a combination of clinical and image data,” Ramos said, adding that “these studies focused on different medical application (classification of Alzheimer’s disease, for example) and other machine learning methods.”
For this study, published online recently in the Journal of NeuroInterventional Surgery. Ramos and colleagues analyzed clinical and baseline CT imaging data from 317 patients with aneurysmal subarachnoid hemorrhage using three different approaches in order to predict DCI. First, they used logistic-regression models to assess the prognostic value of known predictors. Second, they created machine-learning models using all clinical variables. Third, they extracted image features from the CT images using an autoencoder and combined these with clinical data to create additional machine-learning models.
Based on an area under the curve (AUC) analysis, machine-learning models using clinical features were significantly better at predicting DCI than logistic-regression models using known predictors. Aneurysm width and height were included in many of these models. The most effective strategy of all was machine learning using both clinical and imaging features.
Accuracy of Models Used to Predict DCI
|
AUC (95% CI) |
Logistic Regression Model |
0.63 (0.62-0.63) |
Machine Learning with Clinical Data |
0.68 (0.65-0.69) |
Machine Learning with Clinical and Imaging Data |
0.74 (0.72-0.75) |
Ramos pointed out that previous research demonstrates “with early detection, timely intervention can be offered to reverse DCI as soon as possible before it progresses to infarction.” Thus, machine-learning tools could be employed to identify high-risk patients who require close monitoring in an intensive care unit, he suggested.
Potential for Clinical Application in Different Environments
The study authors utilized “popular and widely available open-source machine learning tools” for this study, Ramos explained. “These tools are available in multiple programming languages and for various operational systems. This means that they can be easily implemented and adapted to different environments. Also, to enable reproducibility, our code is publicly available at GitHub.”
Combining clinical and image data also stands out as a “major contribution” of the study, he noted, adding, “Since our deep learning approach does not require any sort of annotation, our models are less prone to bias from labels. The interpretation variables in prediction models using the LIME (local interpretable model-agnostic explanations) method provides a novel understanding of the prediction models and may improve clinical decision-making in the near future.”
Ramos emphasized, however, that the data set used was fairly small and that the findings still require external validation from different sites to ensure that they are generalizable.
Source:
Ramos LA, van der Steen WE, Sales Barros R, et al. Machine learning improves prediction of delayed cerebral ischemia in patients with subarachnoid hemorrhage. J NeuroInterv Surg. 2018;Epub ahead of print.
Disclosures:
Ramos reports no relevant conflicts of interest.