Automatic MRI-based Prognostic Prediction After Acute Ischemic Brain Stroke

1381 

Abstract Submission 

Noam Rotenberg¹, Brian Caffo¹, Andreia Faria¹

¹Johns Hopkins University, Baltimore, MD

Noam Rotenberg  
Johns Hopkins University
Baltimore, MD

Brian Caffo  
Johns Hopkins University
Baltimore, MD

Andreia Faria  
Johns Hopkins University
Baltimore, MD

Stroke is a leading cause of long-term disability. In addition to the obvious benefits to patients and families, early prognosis prediction can assist in triage, precision care, and provide benchmarks for clinical interventions. Previous studies have tried to predict mRS-90, a common measure of outcomes 90 days after stroke, using clinical and imaging features [1-3], but they encountered challenges related to their limited and homogeneous samples, and subjectivity of the lesion evaluation. We developed a comprehensive model to predict mRS-90 in patients with acute ischemic stroke, based on automatically extracted features of the acute injury. This model adds up to the "Acute stroke Detection and Segmentation" tool, ADS [1], a public and user-friendly toolbox, accessible to non-image experts clinical researchers, providing objective quantification of acute ischemic strokes in real time.

The sample includes 1154 patients with MRIs with evidence of ischemic stroke in the diffusion weighted images (DWI), and recorded mRS-90, a subset of our public dataset [4,5]. Each patient's imaging data was represented by a quantitative feature vector (QFV) that reflects the ratio of injury in brain structures [6]: basal ganglia, deep white matter (WM), internal capsule, cerebellum, insula, brainstem, thalamus, and occipital, parietal, temporal, and frontal lobes, plus injury volume and hemisphere(s) affected. A variety of machine learning models (Table 1), were trained (n = 800), tuned, and tested (n = 354). In order to make our results both interpretable and comparable with previous studies we evaluate efficiency of our models in two paradigms:
1. mRS per se: We computed: a) Flexible mRS: accuracy, where predicted mRS ±1 compared to the ground truth mRS is considered true positive; and b) 3-level mRS: accuracy, where predicted mRS and true mRS are mapped into three groups: mRS 0-2 (not affected or slightly disabled), 3-4 (moderately disabled), 5-6 (severely disabled or dead).
2. Binary statistics: predicted and true mRS are mapped into two groups: mRS-90 <=3.5 and > 3.5 (good and bad prognosis, respectively). AUROC, F1, sensitivity, and specificity were calculated. This approach makes our results comparable with those of previous studies, which commonly use a binary paradigm, and enable us to see whether the prediction efficiency differs for patients mildly or severely affected.
Feature Analysis was performed with gradient boosting forest feature importance in three models, including 1) all patients; 2) patients with small strokes (below-median (5.8 mL) injury volume); and 3) patients with large strokes, which are known to exhibit more variable prognosis. Stratifying by lesion volume aims to reduce the effect of total lesion volume and to reveal the regions that mostly influences prognosis.

Overall, all models had similar accuracy, both for the "per se" mRS and the binary prognostic prediction, and all had higher specificity than sensitivity (table 1). These results rival with those reported in previous studies for binary prognostic classification [1-3]. This is encouraging since the models were trained and tested in large and independent samples of real clinical data. The feature analysis (figure 1) shows that stroke volume was the most important feature in the model trained with all the patients, as expected. In small strokes, as the importance of the total injury volume diminishes, the importance of lesions in regions that are crucial for vital functions (e.g., brainstem) or in association areas (frontal lobe, WM) increases. Within the large strokes, the injury of regions within the posterior circulation (occipital and temporal lobe, thalamus, and cerebellum) lead to predict worse prognosis.

Our results show that prognosis can be reasonably predicted by lesion location automatically calculated from brain MRIs. This process enables lesion-functional modeling and accessible prognostic prediction for patient stratification and personalized care.

Classification and Predictive Modeling ¹

Databasing and Data Sharing ²

Computational Neuroscience

Data analysis

DISORDERS

Machine Learning

Open Data

Open-Source Software

Structures