Zijing Yang1, Yawen Hou2, Jingjing Lyu1, Di Liu3, Zheng Chen4. 1. Department of Biostatistics, School of Public Health (Guangdong Provincial Key Laboratory of Tropical Disease Research), Southern Medical University, Guangzhou, P.R. China. 2. Department of Statistics, Jinan University, Guangzhou, P.R. China. 3. Department of Oncology, Second Affiliated Hospital of Xi'an Jiaotong University, Xi'an, P.R. China. 4. Department of Biostatistics, School of Public Health (Guangdong Provincial Key Laboratory of Tropical Disease Research), Southern Medical University, Guangzhou, P.R. China. Electronic address: zheng-chen@hotmail.com.
Abstract
PURPOSE: Providing up-to-date information on patient prognosis is important in determining the optimal treatment strategies. The currently available prediction models, such as the Cox model, are limited to making predictions from baseline and do not consider the time-varying effects of covariates. METHODS: A total of 1501 cervical cancer patients from the Surveillance, Epidemiology, and End Results (SEER) database were included. We introduced three landmark dynamic prediction models (models 1-3) that explore the dynamic effects of prognostic factors to obtain 5-year dynamic survival rate predictions at different prediction times. The performances of these models were evaluated by Harrell's C-index and the Brier score using cross-validation. RESULTS: Some variables did not meet the proportional hazards assumption, indicating that the constant hazard ratios were unreliable. Model 3, which showed the best performance for prediction, was selected as the final model. Significant time-varying effects were observed for age at diagnosis, International Federation of Gynecology and Obstetrics (FIGO) stage, lymph node metastasis, and histological subtypes. Three patients were as examples used to illustrate how the predicted probabilities change at different prediction times during follow-up. CONCLUSIONS: Model 3 can effectively incorporate covariates with time-varying effects and update the probability of surviving an additional 5 years at different prediction times. The use of the landmark approach may provide evidence for clinical decision making by updating personalized information for patients.
PURPOSE: Providing up-to-date information on patient prognosis is important in determining the optimal treatment strategies. The currently available prediction models, such as the Cox model, are limited to making predictions from baseline and do not consider the time-varying effects of covariates. METHODS: A total of 1501 cervical cancerpatients from the Surveillance, Epidemiology, and End Results (SEER) database were included. We introduced three landmark dynamic prediction models (models 1-3) that explore the dynamic effects of prognostic factors to obtain 5-year dynamic survival rate predictions at different prediction times. The performances of these models were evaluated by Harrell's C-index and the Brier score using cross-validation. RESULTS: Some variables did not meet the proportional hazards assumption, indicating that the constant hazard ratios were unreliable. Model 3, which showed the best performance for prediction, was selected as the final model. Significant time-varying effects were observed for age at diagnosis, International Federation of Gynecology and Obstetrics (FIGO) stage, lymph node metastasis, and histological subtypes. Three patients were as examples used to illustrate how the predicted probabilities change at different prediction times during follow-up. CONCLUSIONS: Model 3 can effectively incorporate covariates with time-varying effects and update the probability of surviving an additional 5 years at different prediction times. The use of the landmark approach may provide evidence for clinical decision making by updating personalized information for patients.
Authors: David Cibula; Lukáš Dostálek; Jiri Jarkovsky; Constantijne H Mom; Aldo Lopez; Henrik Falconer; Anna Fagotti; Ali Ayhan; Sarah H Kim; David Isla Ortiz; Jaroslav Klat; Andreas Obermair; Fabio Landoni; Juliana Rodriguez; Ranjit Manchanda; Jan Kosťun; Ricardo Dos Reis; Mehmet M Meydanli; Diego Odetto; Rene Laky; Ignacio Zapardiel; Vit Weinberger; Klára Benešová; Martina Borčinová; Darwin Pari; Sahar Salehi; Nicolò Bizzarri; Huseyin Akilli; Nadeem R Abu-Rustum; Rosa A Salcedo-Hernández; Veronika Javůrková; Jiří Sláma; Luc R C W van Lonkhuijzen Journal: Eur J Cancer Date: 2021-10-16 Impact factor: 10.002