Generalized methodology for radiomic feature selection and modelling in predicting clinical outcomes

Background Quantitative radiomic features of medical images could provide clinical significance in assisting decision making, but the existing feature selection and modelling methods are usually parameter dependent. We aim to develop and validate a generalized radiomic method applicable to a variety of clinical outcomes. Methods and materials A generalized methodology for radiomic feature selection and modelling ("GRFM" for short), including a two-step feature selection and logistic regression, was performed for correlating with clinical outcomes. The two-step feature selection consists of Pearson correlation analysis followed by sequential forward floating selection algorithm to identify robust feature subset. We also applied an adaptive searching strategy to systematically determine globally optimal parameters, rather than relying on preset parameters. The area under the receiver operating characteristic curve (AUC) was used to evaluate the performance of three outcomes: lymph node metastasis of gastric cancer (GC), the 5-year survival status of high-grade osteosarcoma (HOS), and pathological grade of pancreatic neuroendocrine tumors (pNETs). Results The optimal Pearson thresholds were 0.85, 0.80 and 0.75, and the optimal feature numbers were 11, 14 and 8 in GC, HOS and pNETs respectively. AUC values of three predictive models combined with the corresponding parameters were 0.9017 vs. 0.9026, 0.7652 vs. 0.7113, and 0.8438 vs. 0.8212 at the training and validation cohorts, showing higher generality than other methods and better performance than other classifiers. Conclusion The proposed method was helpful in predicting different clinical outcomes, which has potential as a general and noninvasive prediction tool to guide clinical decision-making of various patients.