AI-Radiomics Can Improve Inclusion Criteria and Clinical Trial Performance.

Purpose: Success of clinical trials increasingly relies on effective selection of the target patient populations. We hypothesize that computational analysis of pre-accrual imaging data can be used for patient enrichment to better identify patients who can potentially benefit from investigational agents. Methods: This was tested retrospectively in soft-tissue sarcoma (STS) patients accrued into a randomized clinical trial (SARC021) that evaluated the efficacy of evofosfamide (Evo), a hypoxia activated prodrug, in combination with doxorubicin (Dox). Notably, SARC021 failed to meet its overall survival (OS) objective. We tested whether a radiomic biomarker-driven inclusion/exclusion criterion could have been used to improve the difference between the two arms (Evo + Dox vs. Dox) of the study. 164 radiomics features were extracted from 296 SARC021 patients with lung metastases, divided into training and test sets. Results: A single radiomics feature, Short Run Emphasis (SRE), was representative of a group of correlated features that were the most informative. The SRE feature value was combined into a model along with histological classification and smoking history. This model as able to identify an enriched subset (52%) of patients who had a significantly longer OS in Evo + Dox vs. Dox groups [ p = 0.036, Hazard Ratio (HR) = 0.64 (0.42-0.97)]. Applying the same model and threshold value in an independent test set confirmed the significant survival difference [ p = 0.016, HR = 0.42 (0.20-0.85)]. Notably, this model was best at identifying exclusion criteria for patients most likely to benefit from doxorubicin alone. Conclusions: The study presents a first of its kind clinical-radiomic approach for patient enrichment in clinical trials. We show that, had an appropriate model been used for selective patient inclusion, SARC021 trial could have met its primary survival objective for patients with metastatic STS.