Automatic classification of RDoC positive valence severity with a neural network

Area of Science:

• Computational psychiatry
• Natural Language Processing (NLP)
• Machine Learning (ML)

Background:

• Accurate assessment of Positive Valence symptom severity is crucial for effective psychiatric treatment.
• Initial psychiatric evaluations contain valuable information for severity assessment.

Purpose of the Study:

• To develop a machine learning (ML)-based system for classifying Positive Valence symptom severity.
• To utilize text classification techniques on initial psychiatric evaluations.

Main Methods:

• Treated symptom severity classification as a text classification problem.
• Experimented with logistic regression, gradient boosted trees, and neural networks.
• Employed regularization and mutual information for feature selection; utilized a neural network with three hidden layers.

Main Results:

• The best performing ML system achieved a performance score of 77.86% (based on Mean Absolute Error).
• Error analysis indicated that most misclassifications occurred between adjacent severity categories.
• The system demonstrated effectiveness with a modest amount of training data (600 documents).

Conclusions:

• Machine learning and feature selection can effectively classify Positive Valence symptom severity.
• Increasing annotated data is expected to improve classification accuracy.
• Further enhancements may be achieved with additional features or a larger training corpus.