TY - JOUR
T1 - Comparing survival of older ovarian cancer patients treated with neoadjuvant chemotherapy versus primary cytoreductive surgery
T2 - Reducing bias through machine learning
AU - Huang, Yongmei
AU - Rauh-Hain, J. Alejandro
AU - McCoy, Thomas H.
AU - Hou, June Y.
AU - Hillyer, Grace
AU - Ferris, Jennifer S.
AU - Hershman, Dawn
AU - Wright, Jason D.
AU - Melamed, Alexander
N1 - Publisher Copyright:
© 2024 Elsevier Inc.
PY - 2024/7
Y1 - 2024/7
N2 - Objective: To develop and evaluate a multidimensional comorbidity index (MCI) that identifies ovarian cancer patients at risk of early mortality more accurately than the Charlson Comorbidity Index (CCI) for use in health services research. Methods: We utilized SEER-Medicare data to identify patients with stage IIIC and IV ovarian cancer, diagnosed in 2010–2015. We employed partial least squares regression, a supervised machine learning algorithm, to develop the MCI by extracting latent factors that optimally captured the variation in health insurance claims made in the year preceding cancer diagnosis, and 1-year mortality. We assessed the discrimination and calibration of the MCI for 1-year mortality and compared its performance to the commonly-used CCI. Finally, we evaluated the MCI's ability to reduce confounding in the association of neoadjuvant chemotherapy (NACT) and all-cause mortality. Results: We included 4723 patients in the development cohort and 933 in the validation cohort. The MCI demonstrated good discrimination for 1-year mortality (c-index: 0.75, 95% CI: 0.72–0.79), while the CCI had poor discrimination (c-index: 0.59, 95% CI: 0.56–0.63). Calibration plots showed better agreement between predicted and observed 1-year mortality risk for the MCI compared with CCI. When comparing all-cause mortality between NACT with primary cytoreductive surgery, NACT was associated with a higher hazard of death (HR: 1.13, 95% CI: 1.04–1.23) after controlling for tumor characteristics, demographic factors, and the CCI. However, when controlling for the MCI instead of the CCI, there was no longer a significant difference (HR: 1.05, 95% CI: 0.96–1.14). Conclusions: The MCI outperformed the conventional CCI in predicting 1-year mortality, and reducing confounding due to differences in baseline health status in comparative effectiveness analysis of NACT versus primary surgery.
AB - Objective: To develop and evaluate a multidimensional comorbidity index (MCI) that identifies ovarian cancer patients at risk of early mortality more accurately than the Charlson Comorbidity Index (CCI) for use in health services research. Methods: We utilized SEER-Medicare data to identify patients with stage IIIC and IV ovarian cancer, diagnosed in 2010–2015. We employed partial least squares regression, a supervised machine learning algorithm, to develop the MCI by extracting latent factors that optimally captured the variation in health insurance claims made in the year preceding cancer diagnosis, and 1-year mortality. We assessed the discrimination and calibration of the MCI for 1-year mortality and compared its performance to the commonly-used CCI. Finally, we evaluated the MCI's ability to reduce confounding in the association of neoadjuvant chemotherapy (NACT) and all-cause mortality. Results: We included 4723 patients in the development cohort and 933 in the validation cohort. The MCI demonstrated good discrimination for 1-year mortality (c-index: 0.75, 95% CI: 0.72–0.79), while the CCI had poor discrimination (c-index: 0.59, 95% CI: 0.56–0.63). Calibration plots showed better agreement between predicted and observed 1-year mortality risk for the MCI compared with CCI. When comparing all-cause mortality between NACT with primary cytoreductive surgery, NACT was associated with a higher hazard of death (HR: 1.13, 95% CI: 1.04–1.23) after controlling for tumor characteristics, demographic factors, and the CCI. However, when controlling for the MCI instead of the CCI, there was no longer a significant difference (HR: 1.05, 95% CI: 0.96–1.14). Conclusions: The MCI outperformed the conventional CCI in predicting 1-year mortality, and reducing confounding due to differences in baseline health status in comparative effectiveness analysis of NACT versus primary surgery.
KW - All-cause mortality
KW - Machine learning
KW - Multidimensional comorbidity index
KW - Neoadjuvant chemotherapy
KW - Primary cytoreductive surgery
UR - http://www.scopus.com/inward/record.url?scp=85189031110&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85189031110&partnerID=8YFLogxK
U2 - 10.1016/j.ygyno.2024.03.016
DO - 10.1016/j.ygyno.2024.03.016
M3 - Article
C2 - 38554626
AN - SCOPUS:85189031110
SN - 0090-8258
VL - 186
SP - 9
EP - 16
JO - Gynecologic oncology
JF - Gynecologic oncology
ER -