Reading focus
Syllabus: Why Are Suicidal Thoughts and Actions Difficult to Explain and Predict?
Follow one performance claim from the reported metric through validation and risk-of-bias appraisal.
Why you’re reading it. Apply the earlier evaluation questions to a larger body of model research.
A question to consider. A model ranks outcomes well in a new dataset. What calibration and impact evidence would you still request before accepting a claim of practical benefit?
Why this reading
This review turns model evaluation into an evidence audit: what was developed, what was tested elsewhere, and which performance claims survive scrutiny? It finds promising examples alongside pervasive methodological weaknesses.
Before you read
External validation tests a model in new data. Overfitting means learning sample-specific patterns that may not generalize. Here, self-harm includes acts regardless of intent; ideation alone is outside the review’s outcomes.
Read for
- Compare discrimination, calibration and risk-of-bias findings instead of choosing one headline metric.
- Track the search windows: development studies through 2021, with later external validations of those models searched in 2024.
Consider afterwards
- What remains unknown when a model performs well in another dataset but has not been evaluated for its effect on care?
Editorial context
Published in 2025, this is not a review of every model developed through 2025. Most appraisal was conducted by one reviewer, with a subset checked independently. Two authors disclose involvement in model studies, including OxMIS. Adequate validation performance does not itself demonstrate beneficial implementation.