ESTRO 2024 - Abstract Book

S4423

Physics - Machine learning models and clinical applications

ESTRO 2024

456

Digital Poster

Simulating impact of targeted ablation of progressive disease in mPC patients with machine learning

Timothy G Perk 1 , Glenn Liu 2,1 , Robert Jeraj 3,1

1 AIQ Solutions, Research and Development, Madison, USA. 2 University of Wisconsin, Medicine, Madison, USA. 3 University of Wisconsin, Medical Physics, Madison, USA

Purpose/Objective:

Treatment response heterogeneity is common in oncology patients with multiple lesions. In previous studies, it has been shown that patients identified to have failed on systemic treatments, have only a few progressing lesions that are driving this treatment response. This investigation simulated the impact of targeted ablation of resistant lesions in metastatic prostate cancer (mPC) patients using artificial intelligence models.

Material/Methods:

18F-NaF PET/CT imaging, at baseline and early follow-up, and progression-free survival (PFS) data from 118 patients with mPC from four separate prospective clinical trials were gathered retrospectively. TRAQinform IQ technology (AIQ Solutions) was used to identify, delineate, and track individual regions of interest (ROI) from baseline to follow-up. Each ROI was categorized as new, increasing, stable, decreasing, or disappeared. Eighty-four imaging features were extracted from each patient, including baseline, follow-up, response, patient-level (no inter-ROI comparison), and intra-patient heterogeneity (comparisons between ROIs). The TRAQinform Profile in the form of a random survival forest model was trained using leave-one-out. For each patient, the largest five increasing or new ROIs were simulated as lesions that had been treated with radiation therapy. This was achieved through removal of the lesions, simulating ablation on the second scan, and recomputing all imaging features post simulated ablation. The TRAQinform Profile was used to generate a risk score and to predict PFS and odds of progression by one year for both scenarios, original patient and following simulated radiation therapy. The risk score was evaluated using the c-index and the impact of the ablation was assessed using the change in predicted PFS and odds of PFS over a year.

Results:

TRAQinform IQ identified 99/118 (84%) patients had at least one ROI identified as new or increasing, and 50/118 (42%) had at least 5 new or increasing ROIs. TRAQinform Profile score was a strong predictor of PFS with a c-index of 0.83 across all patients. 63/118 (53%) patients were identified to have an improvement in predicted PFS (mean: 143 days, range: 2-728 days). This resulted in an increase of the likelihood patients would reach one year before failing on treatment (mean: 14%, range: -2 to 54%).

Conclusion: