5P - Application of a novel machine learning model in predicting survival in adrenocortical cancer (ACC)

Background

Machine Learning has the potential to revolutionize cancer care, however, its application is lacking in rare cancers such as ACC. ACC has a dismal prognosis and in need of effective therapies and prognostic tools. Here we applied a novel AI-assisted technology utilizing ¹⁸F-FDG PET scans using a scoring system to characterize metabolic signatures that correlates with survival.

Methods

69 patients with at least two ¹⁸F-FDG PET scans were analyzed, 67% female and 33% male, median age is 50 (80% metastatic [59% with lung metastasis], 16% localized and 4% NED on follow-up). All lesions were contoured for both the baseline and follow-up ¹⁸F-FDG PET scans. Lesions were quantified and matched between timepoints using TRAQinform IQ technology (AIQ Solutions) and TRAQinform Profile was calculated. Imaging features were extracted from each patient, including basic features (SUV_max, SUV_mean, total lesion glycolysis, and number of lesions at baseline), basic response features (baseline, follow up, and change in basic features), location features (number of lesions in different organs) and heterogeneity (intrapatient heterogeneity of disease and response) features. Univariate predictive power of overall survival prediction of each measure was determined using Cox regression models. The TRAQinform Profile was calculated to predict overall survival using 3-fold cross-validation of a random survival forest. The model performance was evaluated using the c-index.

Results

The overall disease burden at the baseline has the strongest predictor to overall survival (c-index = 0.68), followed by disease burden at follow-up (0.65), number of lesions in the lungs at follow-up (0.62), and number of lesions with increasing disease burden (0.62). TRAQinform Profile was able to predict the responder’s vs suboptimal responders to the standard of care treatment (c-index = 0.76).

Conclusions

Machine Learning algorithms are rapidly evolving and new tools using AI are being added to the repertoire of cancer management. Here we present an AI-assisted technology that can help predict the prognosis of ACC patients disease based on the analysis of the lesions present in the ¹⁸F-FDG-PET. This can be translated to tailoring the treatment options to a more personalized approach.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

National Cancer Institute.

Disclosure

O. Lokre, R. Munian-Govindan, T. Perk: Financial Interests, Full or part-time Employment: AIQ Solutions, Inc. All other authors have declared no conflicts of interest.