[Help Needed] Suicide Risk Detection from Long Clinical Notes (Few-shot + ClinicBERT approaches struggling)

Models
3

Fix:

Use a hierarchical model: first summarize or extract key sentences from each note (using a smaller model or rules), then classify the summary.

Use oversampling or data augmentation for the minority class.

Try sentence-transformer embeddings + classical classifier (SVM, XGBoost).

Ensemble sliding window outputs (not just majority vote—try mean/max prob).

If using LLMs, prompt for “warning signs” or “suicide risk factors” and classify based on presence.

Solution provided by Triskel Data Deterministic AI.

1 Like