Machine Learning Models of Early Longitudinal Toxicity Trajectories Predict Cetuximab Concentration and Metastatic Colorectal Cancer Survival in the Canadian Cancer Trials Group/AGITG CO.17/20 Trials.

Abstract

Purpose: Cetuximab (CET), targeting the epidermal growth factor receptor, is a systemic treatment option for patients with colorectal cancer. One known predictive factor for CET efficacy is the presence of CET-related rash; other putative toxicity factors include fatigue and nausea. Analysis of early CET-associated toxicities may reveal patient subpopulations that clinically benefit from long-term CET treatment.

Methods: We analyzed data from CO.20 (ClinicalTrials.gov identifier: NCT00640471) trial arms, CET + brivanib alaninate (BRIV) (n = 376) and CET + placebo (n = 374), and CO.17 (ClinicalTrials.gov identifier: NCT00079066) trial arms, CET (+best supportive care [BSC]; n = 287) and BSC only (n = 285). Patients were clustered into subpopulations using KmL3D, a machine learning method, to analyze 14 joint longitudinal toxicity trajectories from weeks 0 to 8 of treatment. Landmark survival analyses were performed from 8 weeks after treatment initiation. Regression analyses assessed the relationship between subpopulations and plasma CET concentrations. Three supervised machine learning models were developed to assign patients in the CO.20-CET trial arm into subpopulations, which were then validated using CO.20-CET-BRIV and CO.17-CET trial arm data.

Results: Joint longitudinal toxicity clustering revealed dichotomous high- and low-toxicity clusters, with all CET-containing arms showing consistent toxicity trajectories and characteristics. High-toxicity clusters were associated with male predilection, fewer metastatic sites, fewer colon-only primaries, and higher body mass indices. In CO.20 trial samples, higher toxicity clusters were associated with improved overall survival and progression-free survival outcomes (adjusted hazard ratios ranging from 2.21 to 4.36) and higher CET concentrations (P = .003). The random forest predictive model performed the best, with an AUC of 0.981 (0.963-0.999).

Conclusion: We used an innovative machine learning approach to analyze longitudinal joint drug toxicities, demonstrating their role in predicting patient outcomes through a putative pharmacokinetic mechanism.