Prediction of early neurological deterioration in acute minor ischemic stroke by machine learning algorithms.

OBJECTIVES

A significant proportion of patients with acute minor stroke have unfavorable functional outcome due to early neurological deterioration (END). The purpose of this study was to evaluate the applicability of machine learning algorithms to predict END in patients with acute minor stroke.

PATIENTS AND

METHODS

We collected clinical and neuroimaging information from patients with acute minor stroke with NIHSS score of ≤ 3. Early neurological deterioration was defined as any worsening of NIHSS score within 3 days after admission. Unfavorable functional outcome was defined as a modified Rankin Scale score of ≥ 2. We also compared clinical and neuroimaging information between patients with and without END. Four machine learning algorithms, i.e., Boosted trees, Bootstrap decision forest, Deep neural network, and Logistic Regression, were selected and trained by our dataset to predict early neurological deterioration

RESULTS

A total of 739 patients were included in this study. 78 patients (10.6%) experienced END. Among 78 patients with END, 61 (78.2%) had unfavorable functional outcome at 90 days after stroke onset. On multivariate analysis, the initial NIHSS score (P = 0.003), hemorrhagic transformation (P = 0.010), and stenosis (P = 0.014) or occlusion (P = 0.004) of a relevant artery were independently associated with END. Of the four machine learning algorithms, Boosted trees, Deep neural network, and Logistic Regression can be used to predict END in patients with acute minor stroke (Boosted trees: accuracy = 0.966, F1 score = 0.8 and area under the curve = 0.934, Deep neural network :0.966, 0.8, and 0. 904, and Logistic Regression : 0.966, 0.8, and 0.885).

CONCLUSIONS

This study suggests that machine learning algorithms that integrate clinical and neuroimaging information can be used to predict END in patients with acute minor stroke. Further studies based on larger, multicenter datasets are needed to predict END accurately for designing treatment strategies and obtaining favorable functional outcome.