Improving stroke recovery prediction by estimating subject information from population disconnectome

1449 

Abstract Submission 

Patrik Bey¹, Kiret Dhindsa², Michael Schirner², Jan Feldheim³, Marlene Bönstrup⁴, Robert Schulz³, Bastian Cheng³, Götz Thomalla³, Christian Gerloff³

¹Berlin Institute of Health at Charité, Berlin, Berlin, ²Berlin Institute of Health, Berlin, Berlin, ³Universitätsklinikum Hamburg-Eppendorf, Hamburg, Hamburg, ⁴Universitätsklinikum Leipzig, Leipzig, Sachsen

Patrik Bey  
Berlin Institute of Health at Charité
Berlin, Berlin

Kiret Dhindsa  
Berlin Institute of Health
Berlin, Berlin

Michael Schirner  
Berlin Institute of Health
Berlin, Berlin

Jan Feldheim  
Universitätsklinikum Hamburg-Eppendorf
Hamburg, Hamburg

Marlene Bönstrup  
Universitätsklinikum Leipzig
Leipzig, Sachsen

Robert Schulz  
Universitätsklinikum Hamburg-Eppendorf
Hamburg, Hamburg

Bastian Cheng  
Universitätsklinikum Hamburg-Eppendorf
Hamburg, Hamburg

Götz Thomalla  
Universitätsklinikum Hamburg-Eppendorf
Hamburg, Hamburg

Christian Gerloff  
Universitätsklinikum Hamburg-Eppendorf
Hamburg, Hamburg

The detrimental impact of stroke on motor function depends on various factors, from lesion topology and location to alterations in both functional and structural connectivity1,2. Reliable biomarkers for predicting recovery remain an ongoing challenge, particularly during the acute phase after stroke3. Population-based disconnectomes capture lesion impact on connectivity, relate to brain function4 and have been shown to enable prediction of clinical scores with moderate accuracy1,5–7. Here we provide proof-of-concept enhancing this approach. We approximated patient information from disconnectomes , constructed from healthy controls, using a regression model and consequently improved machine learning classification of recovery without the need for individual disconnectomes. This is of particular interest as the clinical context of stroke severely restricts data availability, reducing the capacity to create individual disconnectomes. We show that our framework can use limited data to predict metrics for new patients, recovering some classification accuracy driven by individual variability, thus salvaging direct clinical applicability.

We used previously processed8 MRI data from 20 stroke patients9 and corresponding clinical scores including Barthel, ARAT, MRS and NIHSS. Severity of motor deficit was defined via an aggregate of all scores, computed as the ratio of maximum possible severity. Patient data was acquired 3-5 days (acute) and 85-95 days (subacute) post stroke onset. All patients were assigned a recovery class (good, medium, bad) according to longitudinal motor recovery (Figure 1 (A)). Individual lesion disconnectomes were created by extracting white matter streamlines intersecting the lesion volume. Each subject's brain parcellation was then used to create the disconnectome . We repeated this approach using data of 15 healthy controls from the same cohort9, after transferring each patient's lesion mask into control subject brain. The resulting disconnectomes were averaged to create a population-based disconnectome for each patient (Figure 1 (B)). A total of eight network measures were computed describing complementary properties of the disconnectome: degree, strength, centrality, assortativity, clustering coefficient, density as well as skewness and kurtosis of the degree distribution10. Those measures were used as input to a non-linear support vector machine (SVM) to perform three-way classification using leave-one-out cross validation (LOOCV). A linear regression was fitted to map the population disconnectome to the individual network metrics of each patient. The fitted model was then used to estimate network measures for previously unseen data. Finally, the predicted metrics were used as input to the SVM classification framework to investigate differential information based on these estimates of subject specificity.

Both disconnectomes and corresponding network measures show distinct patterns and distributions between individual and population-based data (Figure 1 (B, D)). The classification showed a strong performance using individual disconnectome information (F1 score: 0.75) (Figure 1 (C)). Population-based information resulted in a reduced performance (F1 score: 0.55). Using the regression-based estimates as input features resulted in an increased performance (F1 score: 0.65) while retaining population-based disconnectomes as the basis.

We first showed that individual disconnectomes provide a higher level of information with regards to prediction of motor-deficit recovery compared to population-based approaches. Critically we could further show that by enhancing population-based disconnectome information, improvement of recovery prediction was significant. Future research can expand upon this work by integrating such estimates of subject variability in recovery predictions and in whole brain modelling for stroke, e.g. using TheVirtualBrain, investigating recovery mechanisms.

Classification and Predictive Modeling ¹

Diffusion MRI Modeling and Analysis

Methods Development

White Matter Anatomy, Fiber Pathways and Connectivity ²

Cerebrovascular Disease

Machine Learning

MRI

Other - Stroke