Machine-Learning Exploration of Exposure-Effect Relationships of Cisplatin in Head and Neck Cancer Patients

Abstract

Background: Cisplatin is a pivotal drug in the treatment of head and neck cancer, and personalized dosage should help the preservation of an optimal toxicity–efficacy ratio. Methods: We analyzed the exposure-effect relationships of 80 adult patients with head and neck cancers and treated with standard Cisplatin-based regimen administered as three-hour infusion. Individual pharmacokinetics (PK) parameters of Cisplatin were identified using a Bayesian approach. Nephrotoxicity and ototoxicity were considered as typical Cisplatin-related toxicities according to Common Terminology Criteria for Adverse Events (CTCAE) standards. Efficacy was evaluated based upon Response Evaluation Criteria in Solid Tumors (RECIST) criteria. Up to nine different machine-learning algorithms were tested to decipher the exposure-effect relationships with Cisplatin. Results: The generalized linear model was the best algorithm with an accuracy of 0.71, a recall of 0.55 and a precision of 0.75. Among the various metrics for exposure (i.e., maximal concentration (Cmax), area-under-the-curve (AUC), trough levels), Cmax, comprising a range between 2.4 and 4.1 µg/mL, was the best one to be considered. When comparing a consequent, model-informed dosage with the standard dosage in 20 new patients, our strategy would have led to a reduced dosage in patients who would eventually prove to have severe toxicities while increasing dosage in patients with progressive disease. Conclusion: Determining a target Cmax could pave the way for PK-guided precision dosage with Cisplatin given as three-hour infusion.

Keywords

• cisplatin
• machine-learning
• Bayesian
• pharmacokinetics
• exposure-effect relationships