External validation of models predicting heart failure in haematological malignancies
C Gomes, K Dziopa, J Geels, A F Schmidt, F W Asselbergs, M LinschotenAbstract
Background/Introduction
Patients with haematological malignancies are at increased risk of heart failure and other cardiovascular events.1 Although several cardiovascular risk prediction models have been developed, patients with haematological malignancies are underrepresented in model development and existing models lack robust external validation.2
Purpose
To externally validate cardiovascular risk prediction models developed in cancer populations across eleven haematological malignancies.
Methods
Patients diagnosed with one of the eleven most common haematological malignancies between 2014 and 2023 were identified from the Netherlands Cancer Registry. Patients who received cancer treatment and had complete treatment information were included and linked to nationwide hospital, claims, ambulatory medication, and mortality databases. Previously identified2 cardiovascular risk prediction models designed to predict cancer therapy-related cardiac dysfunction, either as a primary outcome or as part of a composite cardiovascular outcome, were considered. Model performance was assessed in terms of discrimination using Uno’s C-index at 1-, 3-, and 5 years after diagnosis, and by group-level calibration assessment where feasible. Predictor importance was estimated using permutation importance.
Results
Seven out of 63 models were retained and externally validated in 77,908 patients (largest group: diffuse large B-cell lymphoma (n=12,822); smallest: chronic myeloid leukemia (n=1,937)). For the model with the highest overall discrimination (Abdel-Qadir 2019), 3-year C-indices ranged from 0.59 (95% confidence interval (CI) 0.44-0.73) in acute myeloid leukemia to 0.90 (95% CI 0.84-0.94) in Hodgkin lymphoma. This model, originally developed in breast cancer patients, underestimated risk on calibration assessment. Formal calibration assessment of other models was limited by incomplete reporting of regression equations. Across models, discrimination was consistently lower in malignancies with poorer prognosis compared with those with more favourable prognosis. Age was the most important predictor across malignancies, followed by prior heart failure and hypertension.
Conclusion(s)
Marked heterogeneity in model performance suggests that malignancy-specific recalibration or model updating, or development of malignancy-specific prediction models, may be required to obtain reliable risk estimates. Group-level calibration could only be assessed for one model due to poor reporting in the development studies.