Automatic recognition of symptom severity from psychiatric evaluation records

Area of Science:

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

Background:

• Accurate assessment of psychiatric symptom severity is crucial for effective treatment.
• Automating severity scoring from clinical notes can improve efficiency and consistency.
• Existing methods may not fully capture the nuances of patient narratives.

Purpose of the Study:

• To develop and evaluate a novel NLP and ML framework for automatic psychiatric symptom severity recognition.
• To compare different machine learning approaches for inferring latent symptom severity.
• To map inferred severity scores to discrete clinical levels.

Main Methods:

• Utilized NLP to extract features from psychiatric evaluation records, focusing on discourse, medical concepts, and negation.
• Employed a two-step ML framework: latent continuous severity score inference and mapping to discrete levels.
• Evaluated three inference methods (pointwise regression, pairwise classification, hybrid) and a cascading SVM tree for classification.

Main Results:

• The hybrid approach combining pointwise and pairwise methods achieved the second-highest performance in the CEGS/N-GRID challenge (84.093% normalized MAE).
• Pairwise information processing yielded more accurate severity scores than pointwise regression.
• A cascading SVM tree classifier outperformed traditional SVMs for discrete severity level determination.

Conclusions:

• The proposed NLP and ML framework effectively automates psychiatric symptom severity assessment.
• Hybrid and pairwise approaches show significant promise for improving accuracy in severity scoring.
• Cascading SVMs offer an advantage for classifying discrete severity levels in clinical evaluations.