A deep learning model to predict competing cancer and cardiac risks after anthracyline exposure for early breast cancer

Research output: Contribution to conferenceAbstractpeer-review

Abstract / Description of output

Background: Clinical trials have demonstrated that anthracycline chemotherapy for the adjuvant treatment of early breast cancer (EBC) reduces breast cancer mortality but increases cardiac risk. Attempts to quantify this risk in routine care have been limited by short follow up and inability to adjust for confounding factors and competing risks. The aim of this study was to implement a deep learning framework to quantify excess cardiac risk from anthracycline chemotherapy in real-world care.
Methods: Patients treated surgically for stage I-III invasive breast cancer between 2000 & 2016 were identified from in the Scottish Cancer Registry. Information on treatment and clinical outcomes was captured by linkage to the Scottish Morbidity Record and a regional audit database. The primary outcome was a composite of cardiac diagnosis or cardiac death. The cause-specific cumulative incidence function was used to calculate pseudo survival probabilities for the primary outcome, and the competing risks of death from breast cancer and death from other causes. A deep learning framework was constructed to predict patient survival probabilities and competing risk types at discrete time points, given the pseudo values and patient covariants including: age, deprivation (SIMD), co-morbidity (Charlson), year of diagnosis, side of radiotherapy, cancer stage, grade, ER & HER 2 status.
Results: 4080 EBC patients were identified, 1658 received an anthracycline-based chemotherapy, 297 received non-anthracycline chemotherapy & 2125 received no chemotherapy. At a median follow up of 8.2 years, 448 cardiac events & 559 breast cancer deaths occurred. Age & Charlson score were associated with an increased cardiac risk. Stage & grade were statistically associated with breast cancer death. After hyper-parameter tuning, the deep learning model predicted cardiac events at 8 years with high confidence (F1-score=0.89), and survival probabilities comparable to the more traditional Fine & Gray model; C-index 0.66, [95% Cl 0.62, 0.70] vs. 0.65, [95% CI 0.61- 0.69]).
Conclusions: Taking into account competing risks, there was no statistically increased rate of cardiac events in women treated with anthracycline compared with non-anthracycline chemotherapy or no chemotherapy. The comparable results found with traditional methods in this study is consequence of the reliance on base-line covariants. Further research will explore alternative methods such as a neural network model and time varying covariates. Real world evidence appears reassuring for women treated with anthracyclines for EBC.
Original languageEnglish
Publication statusPublished - 13 Sept 2022
EventESMO Congress 2022 - France, Paris, France
Duration: 9 Sept 202213 Sept 2022


ConferenceESMO Congress 2022


Dive into the research topics of 'A deep learning model to predict competing cancer and cardiac risks after anthracyline exposure for early breast cancer'. Together they form a unique fingerprint.

Cite this