Purpose: Post-operative prognostic tools allow for improved prediction of future recurrence risk, patient counselling, assessment of eligibility for adjuvant treatments, and ensure appropriate follow-up surveillance. The purpose of this analysis is to validate existing prognostic models for patients with kidney cancer.
Materials and methods: The Canadian Kidney Cancer Information System (CKCis) is a prospective cohort of patients managed at 14 institutions since January 1, 2011, to present. CKCis was used to assess 15 predictive models for kidney cancer recurrence, 6 for cancer specific mortality, and 4 for all-cause mortality in patients with a solitary, non-metastatic kidney tumor treated with surgery (partial or radical nephrectomy). Discrimination was measured using c-statistics, 5-year calibration plots for calibration, and decision curve analysis at 5-years post-surgery for net-benefit when considering adjuvant therapy.
Results: 7,174 patients were included. For kidney cancer recurrence, c-statistics ranged from 0.62 to 0.83, depending on whether the model was derived, and applied, to all patients without further stratification, specific risk groups, or to specific histological subtypes. Cancer specific mortality models had c-statistics ranging from 0.60 to 0.89 and all-cause mortality models from 0.60 to 0.73. Using decision curve analysis in clear-cell patients, the best models for choosing adjuvant therapy to prevent recurrence and cancer-related death were the Mayo Clinic prediction models.
Conclusions: Model performance varied considerably with some suitable for clinical use. If using prediction models to select adjuvant therapy, the Mayo Clinic models were best when applied to a large contemporary cohort of Canadian patients.
Take home message: The performance of kidney cancer predictive models varied greatly when applied to a large contemporary cohort. We identified the most accurate models to use when counseling patients about prognosis and adjuvant therapy.