Study of Parametric ROC Curve-fitting in Mathematical Model for Predicting the Probability of Malignant or Benign in Patients with Solitary Pulmonary Nodules

杨德松,李运,姜冠潮,王俊
2012-01-01
Abstract:Objective Based on the mathematical model for predicting the probability of malignant or benign in patients with solitary pulmonary nodules(SPN),the probability of malignant in 145 patients were calculated.The Receiver Operation Characteristic(ROC) curve was plotted by nonparametric method,and then a smooth curve was generated by parametric curve-fitting method to study how to determinate the Optimal Operating Point(OOP) and doubtable value interval.Methods A retrospective cohort study in our institution included 145 patients(98 malignant and 47 benign) with definite pathological diagnosis of SPN from Oct 2009 to Aug 2011.After calculating the probability of malignance of each patient,a ROC curve was graphed by nonparametric method.Another smooth curve(y=f(x)) was drawn through parametric curve-fitting method using software named CurveExpert1.3.The OOP was then determined by solving the maximum of the functionU=Y-X=f(x)-x,which meant by Youden index(sensitivity+specificity-1).The curvature radius equation was obtained followed by its first-order derivative.Afterwards,by analyzing the tendency of the curvature,the doubtable value interval between two points with stron-gest curvature variation were exactly established.Finally,the value of this method were verified through square test of the proportions of malignant-or-benign among the 3 groups divided by the 2 points.Results The area under the ROC cure(AUC) was 0.874±0.035,indicating the diagnostic value of this mathematical model for SPN.By analyzing the smooth ROC curve,the OOP with the value of 57.65% was combined with preferable sensitivity and specificity.The doubtable value interval was 51.15%~60.67% around the OPP.The differences of the malignant probability among 3 groups were statistically significant,unveiling the correlation between results from this prediction model and pathology.Conclusion Evaluation of predicting model using ROC curve was of great clinical value.By parametric curve-fitting,this simple but effective diagnostic method could determine the OOP and doubtable value interval more accurately.
What problem does this paper attempt to address?