A scalable electronic medical record–based risk score for predicting heart failure in cancer survivors: external validation of CHERISH
C Soh, M N Gynnild, H Dalen, T MarwickAbstract
Background
Early identification of cancer survivors at risk of heart failure (HF) is essential to enable preventive intervention. Existing tools such as the HFA-ICOS (Heart Failure Association and the International Cardio-Oncology Society) risk score incorporate clinical examination, cardiac biomarkers and imaging, limiting its scalability in large survivorship populations. The CHERISH risk score was developed as an electronic medical record–based tool using diagnosis-level variables to enable rapid and universally applicable HF risk stratification.
Purpose
To externally validate the CHERISH risk score and compare its predictive performance with HFA-ICOS in a population-based cohort of cancer survivors.
Methods
Our study included cancer survivors aged >39 years, enrolled in a population-based study (2006-2008) and linked to the National Cancer Registry. Individuals with prevalent HF at baseline were excluded. Participants were followed through 2023 via national registry linkage to national registries for incident HF. Model performance was evaluated by discrimination (C-statistics) and calibration (calibration plots and observed-to-expected ratios). For 10-year HF prediction, sensitivity and specificity were assessed using thresholds that maximized specificity while maintaining sensitivity ≥80%.
Results
A total of 2,400 cancer survivors (mean age 67.9±11.4 years, 55% female) were included, with the most common cancer types being breast (20%) and prostate cancer (20%). Median time since cancer diagnosis was 8 years (IQR 3-15). Over a median follow-up of 14.5 years (IQR 7.0–15.9), 348 (15%) participants developed HF and 929 (39%) died without HF. The CHERISH risk score was assessable in all participants (median 10-year predicted risk, 1.7%; IQR, 1.1-3.8) and demonstrated moderate discrimination for 10-year HF (C-statistic 0.728; 95% CI 0.702-0.753). Calibration analysis suggested systematic underestimation of risk (E/O ratio 0.49). HFA-ICOS, calculated using clinical variables only as cardiac biomarkers and imaging were unavailable, classified 628 (26%), 872 (36%) and 900 (38%) participants as low, medium and high risk respectively. HF incidence increased stepwise across categories (3%, 13% and 24%). Discrimination was similar to CHERISH (C-statistic 0.729; 95% CI 0.708-0.750). At a sensitivity threshold ≥80%, CHERISH achieved higher specificity than HFA-ICOS (48.4% [sensitivity 81.7%] vs 28.5% [sensitivity 96.9%]).
Conclusions
In this external validation cohort, CHERISH demonstrated moderate discrimination comparable to HFA-ICOS while being universally applicable using diagnosis-level data alone. Its scalability and improved specificity at clinically relevant sensitivity thresholds support its potential role as a pragmatic triage tool in cardio-oncology care pathways.Calibration of 10-year HF risk using CHERISHCHERISH and HFA-ICOS in predicting HF