New AI model helps oncologists predict 90-day mortality in metastatic breast cancer

Oncologists often overestimate how much time their patients have — and that gap has real consequences. A new prognostic model developed by UNC Lineberger researchers aims to close it by using readily available electronic health record (EHR) data to flag metastatic breast cancer patients at high risk of dying within 90 days.

The model was trained on data from 9,270 patients and achieved 72%–83% accuracy in internal testing, with external validation yielding an AUC of 0.68. Rather than delivering a definitive prognosis, the tool is designed to nudge oncologists toward earlier conversations about palliative care, patient values, and goals of care — before a crisis forces an abrupt transition.

By the Numbers:

• 72%–83% — model accuracy for predicting 90-day mortality in internal testing
• 0.81 AUC in internal validation; 0.68 AUC in external validation
• 9,270 patients used to train the model (CancerLinQ Discovery database)
• 474 patients used for external validation (UNC EHR database)
• ~50% of advanced cancer patients receive aggressive end-of-life care despite evidence supporting palliative approaches

Why it matters: Nearly half of advanced cancer patients receive aggressive care at end of life, while palliative care — shown to improve both quality of life and longevity — remains underutilized. This model could serve as a scalable clinical decision-support tool to shift that dynamic, with researchers already planning real-world implementation studies.