A survey of clinical prediction tools in colorectal and lung cancers and melanoma.
1592 Background: Clinical prediction in cancer depends on a myriad of prognostic factors, and relies on sound methodology for model building and validation. Increased understanding of complex tumour biology allows for simultaneous consideration of biological markers and standard clinical and pathological factors for prediction. We evaluated published studies supporting existing prediction tools in three cancers. Methods: Scientific literature and online resources were searched for clinical prediction tools for survival in three cancers: colorectal, lung, and melanoma. A priori criteria determined by the Molecular Modellers Working Group of the AJCC were evaluated and included: defined patient population, consideration of standard prognostic variables, model development approaches, validation strategies, performance metrics, presentation form of prediction tool, and intended clinical use. Results: Seventy-eight tools intended for prediction of survival were identified for the three cancers: 41 in colorectal, 23 in lung, and 14 in melanoma. Clinical presentations varied within each: 23 of the colorectal cancer tools focused on advanced disease with liver metastases and the remaining varied by stage; 16 lung cancer tools focused on NSCLC and 7 on SCLC. Even in narrowly defined situations, there was no consensus on key variables; for example, no variables were common to all 8 prediction tools for metastatic lung cancer. Variable definitions were missing or vague and the form of the model was often not provided, hampering independent validation and usability. Only 32/78 tools were supported by appropriate internal validity statistics and 21/78 with external validation. Often the development of risk scores did not create groups for whom treatment decisions would be similar. Conclusions: The quality of the literature supporting clinical prediction tools is variable, and the accuracy and utility of many existing tools is undetermined. Methodological guidelines for prediction tool development and validation should be adopted and adhered to. Studies developing and validating clinical prediction tools in cancer must be reported in complete and transparent fashion to facilitate proper interpretation and judgment of utility.