Prediction of Clinical Outcomes in Acute Ischaemic Stroke Patients: A Comparative Study

 What the fuck good does prediction of non-recovery do? Do you tell your patients you have nothing that will get them 100% recovered? Or will you be like my doctor and know nothing and do nothing about stroke?

Prediction of Clinical Outcomes in Acute Ischaemic Stroke Patients: A Comparative Study

Deepthi Rajashekar^1,2,3*, Michael D. Hill^2,3,4,5,6, Andrew M. Demchuk^2,5, Mayank Goyal^2,5, Jens Fiehler⁷ and Nils D. Forkert^2,3,4,8

• ¹Biomedical Engineering Graduate Program, University of Calgary, Calgary, AB, Canada
• ²Depertment of Radiology, Cumming School of Medicine, University of Calgary, Calgary, AB, Canada
• ³Hotchkiss Brain Institute, University of Calgary, Calgary, AB, Canada
• ⁴Department of Clinical Neurosciences, Cumming School of Medicine, University of Calgary, Calgary, AB, Canada
• ⁵Department of Medicine, Cumming School of Medicine, University of Calgary, Calgary, AB, Canada
• ⁶Department of Community Health Sciences, University of Calgary, Calgary, AB, Canada
• ⁷Department of Diagnostic and Interventional Neuroradiology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany
• ⁸Alberta Children's Hospital Research Institute, University of Calgary, Calgary, AB, Canada

Background: Clinical stroke rehabilitation decision making relies on multi-modal data, including imaging and other clinical assessments. However, most previously described methods for predicting long-term stroke outcomes do not make use of the full multi-modal data available. The aim of this work was to develop and evaluate the benefit of nested regression models that utilise clinical assessments as well as image-based biomarkers to model 30-day NIHSS.

Method: 221 subjects were pooled from two prospective trials with follow-up MRI or CT scans, and NIHSS assessed at baseline, as well as 48-hours and 30 days after symptom onset. Three prediction models for 30-day NIHSS were developed using a support vector regression model: one clinical model based on modifiable and non-modifiable risk factors (M_CLINICAL) and two nested regression models that aggregate clinical and image-based features that differed with respect to the method used for selection of important brain regions for the modelling task. The first model used the widely accepted RreliefF (M_RELIEF) machine learning method for this purpose, while the second model employed a lesion-symptom mapping technique (M_LSM) often used in neuroscience to investigate structure-function relationships and identify eloquent regions in the brain.

Results: The two nested models achieved a similar performance while considerably outperforming the clinical model. However, M_RELIEF required fewer brain regions and achieved a lower mean absolute error than M_LSM while being less computationally expensive.

Conclusion: Aggregating clinical and imaging information leads to considerably better outcome prediction models. While lesion-symptom mapping is a useful tool to investigate structure-function relationships of the brain, it does not lead to better outcome predictions compared to a simple data-driven feature selection approach, which is less computationally expensive and easier to implement.

Introduction

The prognosis of clinical and functional outcome in acute ischemic stroke patients is typically made based on multi-modal information such as demographic, clinical, laboratory, and radiological data. Theoretically, machine learning models can identify patterns in high-dimensional data that can be used to make data-driven and reproducible stroke outcome predictions in new patients and support patient management. However, despite the ability to integrate multimodal information, recent machine learning models have mostly utilized clinical data or image-based biomarkers alone (1) to predict stroke outcome. So far, the benefit of using true multi-modal data for stroke outcome prediction has not been investigated comprehensively. One of the few multi-modal predictive models of stroke outcome is described by Brugnara et al. (2). However, clinical assessments at various timepoints are used as input features without addressing the issue of feature collinearity. Furthermore, previous studies often predict the stroke outcome in a binary classification scheme (good vs. bad), which ignores the incremental, yet relevant non-linear differences in stroke severity scores.

Integration of image-based biomarkers for stroke outcome prediction is more complex than using other clinical assessments (in most cases), but has the potential to add considerable predictive power. A key aspect to consider within this context is the selection of regions-of-interest (ROIs) in the brain that are critically associated with the clinical deficit of interest since non-informative and redundant feature can downgrade the prediction accuracy considerably (3). Lesion-symptom mapping (LSM) (4) is able to identify brain regions that are important for a clinical outcome score of interest but has been used rarely for selection of brain regions for stroke outcome prediction (5). The more common ROI selection approach is to use classical feature selection methods during the training process. However, these two general approaches have never been compared to date with respect to stroke outcome prediction.

The aim of this work is to compare different setups of nested machine learning models using clinical information only and a combination of clinical and radiological features selected using lesion-symptom mapping and classical feature selection methods to predict the 30-days NIH stroke scale (NIHSS).

More at link.