# Machine learning-based feature selection and classification for cerebral infarction screening: an experimental study

## TL;DR

This study uses machine learning to analyze speech and cognitive features for detecting and classifying cerebral infarction, showing promising accuracy for early diagnosis.

## Contribution

A novel machine learning framework called CIS is introduced for cerebral infarction screening and subtype classification using speech and cognitive features.

## Key findings

- The CIS framework using XGBoost achieved 88.89% accuracy in distinguishing cerebral infarction from healthy controls.
- The system achieved 77.78% accuracy in classifying lacunar, non-lacunar infarction, and healthy controls.
- Feature selection via RFECV improved classification performance by identifying the most relevant features.

## Abstract

Cerebral infarction screening (CIS) is critical for timely intervention and improved patient outcomes. We investigate the application of machine learning techniques for feature selection and classification of speech and cognitive function assessments to enhance cerebral infarction screening. We analyze a dataset containing 117 patients (95 patients were diagnosed with cerebral infarction, and 54 were identified as lacunar cerebral infarction of them) comprising speech and cognitive function features from patients with lacunar and non-lacunar cerebral infarction, as well as healthy controls. In this article, we present a framework called CIS which comprises a cerebral infarction screening model to identify cerebral infarction from populations and a diagnostic model to classify lacunar infarction, non-lacunar infarction, and healthy controls. Feature selection method, Recursive Feature Elimination with Cross-Validation (RFECV), is employed to identify the most relevant features. Various classifiers, such as support vector machine, K-nearest neighbor, decision tree, random forest, logistic regression, and eXtreme gradient boosting (XGBoost), were evaluated for their performance in binary and ternary classification tasks. The CIS based on XGBoost classifier achieved the highest accuracy of 88.89% in the binary classification task (i.e., distinguishing cerebral infarction from healthy controls) and 77.78% in the ternary classification task (i.e., distinguishing lacunar infarction, non-lacunar infarction, and healthy controls). The selected features significantly contributed to the classification performance, highlighting their potential in differentiating cerebral infarction subtypes. We develop a comprehensive system to effectively assess cerebral infarction subtypes. This study demonstrates the efficacy of machine learning methods in cerebral infarction screening through the analysis of speech and cognitive function features. These findings suggest that incorporating these techniques into clinical practice could improve early detection and diagnosis of cerebral infarction. Further research with larger and more diverse datasets is warranted to validate and extend these results.