Researchers from MIT Media Lab have developed a new machine learning approach for reducing the toxicity of chemotherapy and radiotherapy doses for glioblastoma without hindering their effectiveness.

Presenting their findings in a paper this month at the 2018 Machine Learning for Healthcare conference at Stanford University, the authors elaborated on their AI models, designed to calibrate the lowest possible potency and frequency of dose to potentially minimize side effects and still reduce tumor size to a degree comparable to that of traditional methods.

“Individual drug dosages, to the extent that they can vary between patients, are often calculated as a fixed function of patient body mass or body surface area, though there is reason to believe the latter overestimates the drug's efficacy,” Dr. Pratik Shah, a principal investigator at the Media Lab, told HCB News. “The AI models described in this study were able to successfully reduce tumor sizes in patients during simulated trials despite administering lower frequency and concentrations of drugs.”

Traditional clinical trials focus primarily on the efficacy and effectiveness of drugs and dosing, and often involve the uniform administration of the largest dose concentration and frequency based on predetermined safety calculated from preclinical data.

Utilizing reinforced learning, a technique inspired by behavioral psychology, the model determines an optimal treatment plan by learning to choose behaviors that lead to a desired outcome.

The method comprises artificially intelligent "agents" that complete "actions" in an unpredictable, complex environment to fulfill a desired objective, rewarding or punishing itself depending on whether the action positively contributes to the outcome. It then adjusts its performance based on each result.

The model applies this "self-learning" AI approach to treatment regimens at work, iteratively adjusting doses while evaluating traditional administered regimens. It first decides whether to initiate or withhold dose. If dose is administered, the model then determines if the entire dose or just a portion should be delivered.

With each action, it pings another clinical model to determine if its choice has in any way helped shrink the mean tumor diameter. For a successful dose task, the model assigns itself a positive numerical value, such as +1, and for each unsuccessful one, a negative value, such as -1, until it finds a maximum outcome score for a given task.

Rad Oncology Homepage