Predicting functional outcomes after a stroke event by clinical text notes: A comparative study of traditional...

Area of Science:

• Neurology
• Medical Informatics
• Machine Learning

Background:

• Accurate prediction of functional outcomes post-acute ischemic stroke is vital for healthcare resource allocation.
• Text mining models for stroke outcome prediction often focus on traditional machine learning, with limited deep learning comparisons.

Purpose of the Study:

• To compare traditional machine learning and deep learning methods for predicting functional outcomes after acute ischemic stroke.
• To evaluate different feature representation techniques for stroke outcome prediction models.

Main Methods:

• Clinical text notes from a Taiwanese hospital were used to develop predictive models.
• Four feature representation techniques were evaluated: Bag-of-Words (BOW), Term Frequency-Inverse Document Frequency (TF-IDF), ELMo, and BERT.
• Four predictive models were tested: k-nearest neighbor (KNN), Support Vector Machine (SVM), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM).

Main Results:

• The Support Vector Machine (SVM) classifier with Bag-of-Words (BOW) features achieved the best performance.
• Feature fusion strategies, such as BOW + TF-IDF and BOW + BERT, also showed strong predictive capabilities.
• The BOW + TF-IDF combination with SVM demonstrated the lowest Type I error, minimizing misclassification of patients with poor outcomes.

Conclusions:

• Traditional machine learning methods outperformed deep learning models in this stroke outcome prediction study.
• The combination of BOW + TF-IDF features with SVM provided the most accurate predictions and the lowest risk of false positives.
• This approach offers a robust method for optimizing stroke patient management and resource allocation.