Machine learning algorithms for estimating 10-year survival in patients with bone metastases from prostate cancer: towards a disease-specific survival estimation tool

This article was originally published here

BMC Cancer. 2022 Apr 30;22(1):476. doi: 10.1186/s12885-022-09491-7.

ABSTRACT

BACKGROUND: Prognostic indicators, treatments, and survival estimates vary by cancer type. Therefore, disease-specific models are needed to estimate patient survival. Our primary objective was to develop models to estimate survival time after treatment for skeletal-related events (SREs) (symptomatic bone metastases, including impending or actual pathologic fractures) in men with metastatic bone disease due to to prostate cancer. Such disease-specific models could be added to the clinical decision support tool PATHFx, which is freely available worldwide. Our secondary objective was to determine disease-specific factors that should be included in an international cancer registry.

METHODS: We analyzed the records of 438 men with metastatic prostate cancer who underwent SRE requiring treatment with radiation therapy or surgery from 1989 to 2017. We developed and validated 6 survival models at 1, 2, 3, 4, 5 and 10 years after treatment. Model performance was assessed using calibration analysis, Brier scores, area under the receiver operator characteristic curve (AUC), and curve analysis of the receiver. decision to determine the clinical utility of the models. We characterized the magnitude and direction of the model’s features.

RESULTS: Models exhibited calibration, accuracy (Brier scores 0.73) acceptable. Decision curve analysis determined that all 6 models were suitable for clinical use. The order of importance of features was distinct for each model. In all models, 3 factors were positively associated with survival time: younger age at diagnosis of metastasis, proximal prostate-specific antigen (PSA)

CONCLUSIONS: We have developed models that estimate survival time in patients with metastatic bone disease due to prostate cancer. These models require external validation but should in the meantime be included in the PATHFx tool. PSA and APV data must be registered in an international cancer registry.

PMID:35490227 | DO I:10.1186/s12885-022-09491-7

Sharon D. Cole