Predicting suicidal ideation from depression screening data: A network-augmented machine learning approach.

Area of Science:

• Computational psychiatry
• Network science
• Machine learning in healthcare

Background:

• Suicidal ideation assessment often relies on single self-report items.
• Routine depression symptom data can be leveraged for improved screening.
• A novel model combining machine learning and network analytics was developed.

Purpose of the Study:

• To infer an auxiliary signal for suicidal ideation from routine depressive symptom data.
• To evaluate the performance of machine learning models augmented with network features.
• To compare the novel model against traditional screening methods.

Main Methods:

• Utilized data from the National Health and Nutrition Examination Survey (N=44,922).
• Developed three predictive models: PHQ-8 total score, PHQ-8 items, and items plus network features (centrality, edges, density).
• Employed 10-fold cross-validation and precision-recall area under the curve (PR AUC) for evaluation, with external validation on five independent datasets.

Main Results:

• Item-level models outperformed the PHQ-8 total-score baseline.
• Network-augmented XGBoost demonstrated the strongest performance, achieving a PR AUC of 0.37.
• The model met prespecified screening criteria (recall ≥0.80, specificity ≥0.50) and highlighted key symptom interrelations.

Conclusions:

• Integrating symptom-network features with machine learning enhances interpretability and discrimination.
• The optimized model surpasses the PHQ-8 total-score baseline and meets pragmatic screening criteria.
• This approach is suitable for first-line case finding of suicidal ideation.