EOLCI overestimates one-year mortality

The End-of-Life Care Index ranked risk reasonably but overestimated one-year mortality.
*Retrospective prognostic validation study; Level 2 (OCEBM).

Citation

Kohn R, Courtright KR, Grau-Sepulveda M, et al. Accuracy and Equity of the End-of-Life Care Index in Predicting 1-Year Mortality. JAMA Network Open. 2026;9(9):e2633038. doi:10.1001/jamanetworkopen.2026.33038

Background

Hospitals increasingly use electronic record tools to identify seriously ill patients who may benefit from palliative care. This study tested whether a widely available commercial one-year mortality tool performs accurately and fairly in real hospital populations.

Patients

Adults hospitalized at least 36 hours in 39 hospitals in two United States health systems during 2022. Exclusions included psychiatry, rehabilitation, hospice, obstetrics, and neonatology admissions.

Intervention

End-of-Life Care Index prediction score using electronic health record data.

Control

Observed one-year vital status in the health record.

Outcome

One-year mortality prediction performance, including scaled Brier score (-1 to 1) and C statistic (0 to 1).

Follow-up Period

One year after 36 hours of hospitalization.

Results

Health system Encounters One-year deaths Overall performance Risk ranking
Trinity Health 154,063 10.3% SBS -0.01 (95% CI -0.03 to 0.01) C statistic 0.76 (95% CI 0.76 to 0.77)
Kaiser Permanente Southern California 133,043 17.9% SBS 0.18 (95% CI 0.16 to 0.20) C statistic 0.81 (95% CI 0.80 to 0.81)

SBS: scaled Brier score; higher is better, 0 means no improvement over a simple baseline model.

Calibration was poor in both systems: predicted risk was generally higher than observed mortality, especially at higher scores. At the 70% risk threshold, positive predictive value was 0.45 at Trinity Health and 0.62 at Kaiser Permanente Southern California (CIs not reported). Performance was similar across most demographic groups, but worse among the oldest patients and some diagnosis groups.

Limitations

Deaths outside the health record may have been missed, making overprediction appear worse. Race, ethnicity, and diagnosis data may be inaccurate. The model was not recalibrated locally, and results may differ in other hospitals.

Funding

Patient-Centered Outcomes Research Institute; no funder role reported.

Clinical Application

Use this tool only after local validation; it may flag high-risk patients but should not guide prognosis discussions alone.