Combining EHR, claims and outside death data captures most cancer deaths in real-world database | AMCP Nexus 2026

Electronic health records (EHRs) alone missed more than 4 in 10 deaths among patients with advanced solid tumors in the Tempus AI real-world database. Combining those records with an outside mortality data source and claims data raised the capture rate to 86.6%, according to a poster presented at AMCP Nexus 2026 held this week in San Diego.

Researchers at Tempus, a Chicago-based precision medicine and data company, checked the combined measure against the National Death Index (NDI). Maintained by the CDC's National Center for Health Statistics, the NDI is considered the national gold standard for death data. It includes all U.S. deaths since 1979, but records become available for matching only after a lag, according to the Office of the Assistant Secretary for Planning and Evaluation (ASPE) at HHS.

Joshuah Kapilivsky, B.S., data scientist with Tempus, was the lead author, and Emilie Scherrer, M.Sc., senior director and head of strategy at Tempus, presented the poster.

Gaps in EHR death data

Real-world data from EHRs are important for evaluating treatment effectiveness and overall survival in cancer, the authors wrote. However, EHR data have known gaps that can be reduced by combining them with other data sources.

Regulators have flagged these same gaps. In fact, in the final guidance issued in July 2024 on EHR and claims data, the FDA noted that death and cause of death are generally not included in these data. The agency advised drug sponsors to carefully document the quality of their mortality data in study protocols.

Other real-world data companies have taken a similar approach.

Flatiron Health has described a composite mortality variable built from EHR data, commercial death data and the Social Security Death Index and benchmarked against the NDI. A 2019 analysis by Roche and Flatiron researchers found that missing deaths had minimal impact on survival analyses when the completeness of mortality data was high.

Each data source added deaths

The Tempus team drew on deidentified clinical and molecular data from 17,597 patients diagnosed with advanced solid tumors between 2016 and 2020. Cancer types included non-small cell lung, breast, ovarian, pancreatic, prostate and colorectal cancer. Records were linked to a third-party mortality data aggregator, Veritas, and to a claims dataset using deidentified patient tokens. The analysis was limited to the 4,795 patients with ongoing direct EHR integrations, so the results describe that subset rather than the full cohort.

On its own, each data source had moderate sensitivity, meaning the share of NDI-confirmed deaths it captured, and high specificity. EHR data alone caught 56.7% of deaths. Adding the third-party mortality data raised that to 84.3%. Adding claims on top brought the final figure to 86.6%.

The combined measure's specificity, or how often patients who were alive were correctly classified as alive, was 96.2%. When the composite variable recorded a death, the NDI confirmed it 96.7% of the time. When it showed a patient as alive, it was correct 84.7% of the time.

Dates were close, but results varied by group

It was revealed that the composite variable's date of death matched the NDI exactly 74.6% of the time. Agreement rose to 91.8% within 15 days and 96.1% within 30 days. The authors noted that those windows reflect the date precision needed for real-world overall survival analyses in cancer. Death dates taken directly from the EHR had the highest exact agreement, at 81.1%, but were similar to the composite at the 30-day window (96.4%).

Sensitivity was at least 84% for every cancer type, ranging from 84.1% in breast cancer to 90.2% in a group that combined ovarian, fallopian tube and primary peritoneal cancer. Results were nearly identical for men (86.7%) and women (86.5%). Sensitivity was highest among patients 75 and older, at 91.2%, compared with 84.4% to 86.3% in younger age groups.

Researchers also found that performance was less even by race. Sensitivity was 88.3% among White patients, 87.6% among patients of other races and 83.7% among Black patients. It dropped to 75.9% among Asian patients, though that group was small, at 184 patients. Other research has found that death-record matching can be less reliable for some groups.

A 2023 study in BMJ Health & Care Informatics of more than 1 million Kaiser Permanente Mid-Atlantic States members found that Asian and Pacific Islander and Hispanic members were underrepresented among NDI-matched deaths. Those members were also more likely to be missing a Social Security number, which the matching process relies on.

Patients diagnosed with metastatic disease had slightly lower sensitivity (86.5%) than those diagnosed before their cancer had spread (89.9%). Most of the cohort, 4,440 patients, was diagnosed with metastatic disease.

In a supplementary analysis, researchers treated patients without a recorded death as followed only through their last known date alive, which the authors said better reflects how survival data are used in real-world studies. Under that approach, sensitivity was 95.3% at six months, 95.5% at 12 months and 97.3% at 24 months, with specificity of 98.9% or higher.

The authors said this suggests patients with missed deaths are properly handled rather than misclassified as alive.

The researchers concluded that the composite variable is a reliable endpoint for real-world evidence analyses. Tempus sponsored the study, and all of the authors are affiliated with the company.