Comparisons of different classification algorithms while using text mining to screen psychiatric inpatients with...

Area of Science:

• Computational linguistics
• Psychiatric informatics
• Machine learning in healthcare

Background:

• Accurate identification of suicidal behaviors in psychiatric inpatients is crucial for timely intervention and suicide prevention.
• Traditional screening methods can be time-consuming and resource-intensive.
• Electronic Medical Records (EMR) offer a rich source of data for developing automated screening tools.

Purpose of the Study:

• To evaluate the performance of text mining methods for screening suicidal behaviors based on psychiatric inpatients' chief complaints.
• To compare the effectiveness of different text mining algorithms and weighting factors in classifying suicidal behaviors.

Main Methods:

• A dataset of 3600 psychiatric inpatients (1800 with and 1800 without suicidal behaviors) was analyzed.
• Text mining techniques were applied to chief complaint data to identify patterns associated with suicidal behaviors.
• Six algorithms and two term weighting factors were tested with varying training set sizes, with performance assessed using precision, recall, F1-value, and accuracy.

Main Results:

• Common terms in suicidal inpatients' chief complaints included "suicide", "notion", and "suspicion".
• Algorithms achieved stable performance with training set sizes around 1000, reaching accuracy values over 0.95.
• Support Vector Machine (SVM), Random Forest, and AdaBoost, weighted by TF, demonstrated superior generalization ability with high F1 values (e.g., 0.9889 for SVM).

Conclusions:

• Text mining is a feasible approach for filtering psychiatric inpatients with suicidal behaviors using limited, representative terms.
• SVM, Random Forest, and AdaBoost algorithms weighted by TF are highly effective for this classification task.
• This automated approach can significantly save time and resources in identifying high-risk patients, enhancing suicide prevention efforts.