Statistical characterization of a good biomarker in oncology

The  aim of this article is to review and illustrate the attributes that analyze the performance of a predictive model, suchas discrimination, calibration and clinical utility.

To illustrate a biomarkervalidation process, we analyzed 216 patientsrecruited in the Miguel Servet University Hospital Zaragoza, Spain. The outcome of the study was clinicallysignificant prostate cancer (Gleason ≥ 7). A newbiomarker was built using logistic regression modelfrom age, prostate-specific antigen, prostate volumeand digital rectal exam variables. To analyze the discriminationability, the receiver operating characteristiccurve, its area under the curve (AUC), and Youdenindex were estimated. In addition, the calibration wasanalyzed through calibration curve, intercept and slope;and the clinical utility was studied by means of decisionand clinical utility curves.

The discrimination ability was good:AUC 0.790 (0.127-0.853 95% C.I.), Youden index cutoffpoint 0.431 (specificity 0.811, sensitivity 0.697).The Intercept was 0 and Slope 1 showing a perfect calibration.Decision curve showed good net benefit in athreshold probability range 25%-80%. Clinical utilitycurve showed that for a 18% cutoff point, a minimum4.5% of CsPCa patients are wrongly classified belowthe cutoff point, saving 18.5% biopsies.

A complete validation process isnecessary to analyze the performance of a biomarkerin oncology, based on their discrimination ability, theconcordance between predicted and actual occurrenceof the outcome, and its applicability in clinical practice.