Machine Learning Application in a Phase I Clinical Trial Allows for the Identification of Clinical-Biomolecular Markers Significantly Associated with Toxicity

Machine learning (ML) algorithms have been used to forecast clinical outcomes or drug adverse effects by analyzing different data sets such as electronic health records, diagnostic data, and molecular data. However, ML implementation in phase I clinical trial is still an unexplored strategy that implies challenges such as the selection of the best development strategy when dealing with limited sample size. In the attempt to better define prechemotherapy baseline clinical and biomolecular predictors of drug toxicity, we trained and compared five ML algorithms starting from clinical, blood biochemistry, and genotype data derived from a previous phase Ib study aimed to define the maximum tolerated dose of irinotecan (FOLFIRI (folinic acid, fluorouracil, and irinotecan) plus bevacizumab regimen) in patients with metastatic colorectal cancer. During cross-validation the Random Forest algorithm achieved the best performance with a mean Matthews correlation coefficient of 0.549 and a mean accuracy of 80.4%; the best predictors of dose-limiting toxicity at baseline were hemoglobin, serum glutamic oxaloacetic transaminase (SGOT), and albumin. The feasibility of a prediction model prototype was in principle assessed using the two distinct dose escalation cohorts, where in the validation cohort the model scored a Matthews correlation coefficient of 0.59 and an accuracy of 82.0%. Moreover, we found a strong relationship between SGOT and irinotecan pharmacokinetics, suggesting its role as surrogates' estimators of the irinotecan metabolism equilibrium. In conclusion, the potential application of ML techniques to phase I study could provide clinicians with early prediction tools useful both to ameliorate the management of clinical trials and to make more adequate treatment decisions.