Machine learning-based prediction of clinical outcomes after first-ever ischemic stroke

Abstract

Background: Accurate prediction of clinical outcomes in individual patients following acute stroke is vital for healthcare providers to optimize treatment strategies and plan further patient care. Here, we use advanced machine learning (ML) techniques to systematically compare the prediction of functional recovery, cognitive function, depression, and mortality of first-ever ischemic stroke patients and to identify the leading prognostic factors.

Methods: We predicted clinical outcomes for 307 patients (151 females, 156 males; 68 ± 14 years) from the PROSpective Cohort with Incident Stroke Berlin study using 43 baseline features. Outcomes included modified Rankin Scale (mRS), Barthel Index (BI), Mini-Mental State Examination (MMSE), Modified Telephone Interview for Cognitive Status (TICS-M), Center for Epidemiologic Studies Depression Scale (CES-D) and survival. The ML models included a Support Vector Machine with a linear kernel and a radial basis function kernel as well as a Gradient Boosting Classifier based on repeated 5-fold nested cross-validation. The leading prognostic features were identified using Shapley additive explanations.

Results: The ML models achieved significant prediction performance for mRS at patient discharge and after 1 year, BI and MMSE at patient discharge, TICS-M after 1 and 3 years and CES-D after 1 year. Additionally, we showed that National Institutes of Health Stroke Scale (NIHSS) was the top predictor for most functional recovery outcomes as well as education for cognitive function and depression.

Conclusion: Our machine learning analysis successfully demonstrated the ability to predict clinical outcomes after first-ever ischemic stroke and identified the leading prognostic factors that contribute to this prediction.

Keywords: cognitive impairment; functional outcome; machine learning; mortality; outcome prediction; post-stroke depression; stroke.

Figure 1

Flowchart depicting patient selection process. PROSCIS, PROSpective Cohort with Incident Stroke; MRI, Magnetic resonance imaging.

Figure 2

Process flow of input data, machine learning analysis and outcome prediction. mRS, modified Rankin Scale; BI, Barthel Index; MMSE, Mini-Mental State Examination; TICS-M, Modified Telephone Interview for Cognitive Status; CES-D, Epidemiologic Studies Depression Scale; SVM-lin, Support Vector Machine with linear kernel; SVM-rbf, Support Vector Machine with radial basis function kernel; GB, Gradient Bossting Classifier; MRI, Magnetic resonance imaging.

Figure 3

Prediction performance in balanced accuracy (BA) for all outcomes, time points and input subdomains. In (A) all input parameters were considered while (B–E) show the results of the (B) demographic, (C) clinical, (D) serological and (E) MRI input subdomain. Results for BI after 1 year were unreliable due to the extreme class imbalance in the dataset (see Table 2). mRS, modified Rankin Scale; BI, Barthel Index; MMSE, Mini-Mental State Examination; TICS-M, Modified Telephone Interview for Cognitive Status; CES-D, Epidemiologic Studies Depression Scale; SVM-lin, Support Vector Machine with linear kernel; SVM-rbf, Support Vector Machine with radial basis function kernel; GB, Gradient Bossting Classifier; MRI, Magnetic resonance imaging.

Figure 4

Decision-making process by the Gradient Boosting Classifier for the modified Rankin Scale (mRS) at patient discharge on the level of individual patients depicted via Shapley values. The relative importance of an input variable can be quantified by its Shapley value and represented by the length of a bar. In this example, features in red counted toward a good outcome while blue features signified poor outcome for mRS at patient discharge. In (A) a patient with a mRS score of 1 point was correctly classified as having a good outcome with variables such as low National Institutes of Health Stroke Scale (NIHSS), high-sensitivity C-reactive protein (hsCRP), cholesterol and acute infarct volume in Diffusion-weighted imaging (DWI) outweighing a high Wahlund Score. In (B) a patient with a mRS score of 4 points was correctly predicted as having poor outcome due to high NIHSS, hsCRP and cholesterol whilst offsetting a low acute infarct volume in DWI. In both instances the decision was made by considering the total impact of all features.