Predicting post-discharge self-harm incidents using disease comorbidity networks: A retrospective machine learning...

Area of Science:

• Medical Informatics
• Public Health
• Machine Learning

Background:

• Self-harm is a significant public health concern, with early risk identification being crucial for prevention.
• The co-occurrence of multiple diseases (comorbidity) is linked to an increased risk of self-harm.
• Existing prediction models often do not fully leverage complex disease relationships.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework for predicting individual self-harm risk.
• To improve the accuracy of self-harm prediction by incorporating disease comorbidity patterns.
• To identify key diagnoses and comorbidity patterns associated with self-harm risk.

Main Methods:

• A deep learning framework was developed utilizing a comorbidity network constructed from historical diagnoses.
• A novel patient embedding method, Dx2Vec (Diagnoses to Vector), was proposed to represent comorbidity and temporal admission patterns.
• The model was trained and tested on a large dataset of inpatient records from Hong Kong public hospitals.

Main Results:

• The Dx2Vec-based model demonstrated superior performance compared to baseline models, achieving a C-statistic of 0.89.
• The model achieved high precision (0.54 positive, 0.98 negative) and recall (0.72 positive, 0.96 negative) in identifying patients at risk of self-harm within 12 months.
• The framework successfully identified predictive diagnoses and pairwise comorbid diagnoses.

Conclusions:

• Integrating disease comorbidity networks significantly enhances self-harm prediction.
• The findings underscore the importance of considering comorbidity patterns in self-harm screening and prevention strategies.
• The developed framework shows potential for implementation in clinical settings for effective self-harm risk assessment.