Tabea Kossen, BCCN Berlin / TU Berlin

Manifold Learning for Individual Assessment of Infarct Progression in Acute Ischemic Stroke

Background: To provide good patient care, an efficient treatment stratification for acute ischemic stroke patients is necessary. Nowadays only patients within a restricted time-window of 4.5 hours after symptom's onset are eligible for treatment. Statistically, beyond that time window, risks exceed the benefits for the patients. However, this approach excludes patients that would benefit from treatment beyond the time-window as well as patients with an unknown symptom's onset. Above all, the current stratification paradigm does not allow for an individual stratification into treatment. In the present study the objective was to establish the ground to shift from the time-window approach to a tissue-based approach as represented by acute-imaging. Using manifold learning techniques, we integrated multi-modal magnetic resonance imaging (MRI) data to allow an individual assessment of stroke progression. Ultimately, we aimed to provide a data-driven tool for estimating the further infarct progression in a personalized approach.

Methods: In a retrospective study 199 acute ischemic stroke patients were analyzed. Three different manifold-learning models were applied to MR imaging-based features for the purpose of identifying a new representation of the data. 1) the compact model - a model applying compressed features representation, 2) the elaborated model - a model applying an extensive feature extraction represented in an elaborated feature vector and 3) the elaborated AD model - an extension of the elaborated model using alternating diffusion (AD) technique. In the second part of this study, voxels were labeled as part of the infarcted tissue or the tissue at risk. Using the new metric of the best performing model, the labels were propagated and utilized to estimate whether the infarcted brain tissue will further expand or not. Hierarchical coupled geometry analysis (HCGA) was used to improve the applied metric. The performance of manifold learning using diffusion maps as well as HCGA was compared to a reference-model based on the raw data without manifold learning application. In order to account for the treatment effect, the analysis was applied to the full cohort as well as separately to two patient groups according to the treatment status.

Results: We found that the elaborated model yielded the best performance with an error rate of 0.051 compared to the compact model and the elaborated AD model with error rates of 0.076 and 0.105 respectively.
In the labels-propagation, the HCGA technique demonstrated the best performance (AUC=0.580) compared with the diffusion maps model (AUC=0.557) and the reference model (AUC=0.553). The division of patients into treated (134 patients) and non-treated (65 patients) yielded a similar pattern. No apparent difference between the groups could be observed.

Conclusions: In this study we showed for the first time the possible application of unsupervised learning algorithms in acute stroke imaging. Our results corroborate that unsupervised learning may assist to reveal underlying information as was demonstrated by the improved performance using manifold learning techniques comparing to the reference model. Of the set of techniques that were investigated in this study, the HCGA method seems to be the most promising. This study provides a proof of concept of unsupervised machine learning to bring additional value in acute stroke imaging. We suggest to further pursue an individualized tissue-based stratification approach for acute stroke patients using integrative unsupervised and supervised methods to yield a maximal gain from both methods and reach a clinical performance standard applicable for clinical settings.

Additional Information

MSc defence in the International Master program Computational Neuroscience.

Organized by

Klaus Obermayer / Robert Martin