Identifying psychosis episodes in psychiatric admission notes via rule-based methods, machine learning, and...

Area of Science:

• Computational psychiatry
• Medical informatics
• Natural Language Processing

Background:

• Accurate psychosis diagnosis is vital but challenging due to symptom variability and underreporting.
• Existing methods using structured Electronic Health Record (EHR) data are often insufficient for psychosis identification.
• Stigma and diminished insight further complicate early and accurate diagnosis of psychotic episodes.

Purpose of the Study:

• To evaluate the effectiveness of Natural Language Processing (NLP) algorithms for identifying psychosis in psychiatric admission notes.
• To compare rule-based, machine learning, and pre-trained language models for psychosis detection.
• To assess the impact of keyword pre-selection on NLP model performance in EHR data.

Main Methods:

• Analysis of 4629 psychiatric admission notes (2005-2019) from patients aged 16-35.
• Application of rule-based algorithms, machine learning (XGBoost with TF-IDF), and pre-trained language models (BlueBERT).
• Evaluation of keyword pre-selection strategies to refine notes before model training and analysis.

Main Results:

• The XGBoost classifier with TF-IDF features and keyword pre-selection achieved the highest F1 score of 0.8881.
• BlueBERT showed comparable performance with an F1 score of 0.8841.
• Both NLP models significantly outperformed traditional ICD code-based detection (F1 score 0.7608), with keyword pre-selection enhancing performance.

Conclusions:

• NLP techniques, particularly with keyword pre-selection, offer a powerful approach to improve psychosis detection in clinical notes.
• Machine learning and pre-trained language models demonstrate superior performance compared to traditional methods for psychosis identification in EHRs.
• This study provides a foundation for leveraging NLP to enhance psychosis diagnosis and patient care within electronic health records.