NYU Langone Health

Approximately 15% of NYU Langone patients are discharged to skilled nursing facilities, but identifying these patients early is difficult. Lengthy, unstructured physician admission notes were too long for AI models to process directly, and late identification leads to stressful situations where patients are medically ready for discharge but lack a safe care destination.

Researchers developed a two-step AI pipeline: a generative AI model first reads each full admission note and extracts seven key risk factors (e.g., living situation, ability to perform daily tasks) into a condensed 'AI Risk Snapshot' that is 94% shorter than the original note. A second AI component then uses that snapshot to predict whether the patient will require skilled nursing facility placement at discharge.

The model achieved 88% accuracy in predicting skilled nursing facility need at discharge. When nurse case managers independently reviewed the AI-generated summaries without seeing the model's prediction, their assessments strongly aligned with the AI risk scores. A high-risk score from the model made it 13.5 times more likely that a nurse would independently flag the patient as needing skilled nursing care. The study was published in the Nature-family journal npj Health Systems.

• Summarizing complex clinical notes before prediction—rather than feeding raw text directly—significantly improves both accuracy and feasibility for AI discharge planning models.
• Human-AI alignment validation (having clinicians review summaries blind to predictions) is an effective method to establish clinical trustworthiness before real-world deployment.
• Targeting a small set of structured risk factors (7 in this case) can compress notes by 94% while retaining predictive signal, enabling models with context-length constraints to function.