AI Can Make Sure Cancer Patients Get Just Enough (but Not Too Much) Treatment
QUALITY OF LIFE. Patients with glioblastoma, a malignant tumor in the brain or spinal cord, typically live no more than five years after receiving their diagnosis. And those five years can be painful — in an effort to minimize the tumor, doctors often prescribe a combination of radiation therapy and drugs that can cause debilitating side effects for patients.
Now, researchers from MIT Media Lab have developed artificial intelligence (AI) that can determine the minimum drug doses needed to effectively shrink glioblastoma patients’ tumors. They plan to present their research at Stanford University’s 2018 Machine Learning for Healthcare conference.
CARROT AND STICK. To create an AI that could determine the best dosing regimen for glioblastoma patients, the MIT researchers turned to a training technique known as reinforcement learning (RL).
First, they created a testing group of 50 simulated glioblastoma patients based on a large dataset of those that had previously undergone treatment for their disease. Then they asked their AI to recommend doses of several drugs typically used to treat glioblastoma [oftemozolomide (TMZ) and a combination of procarbazine, lomustine, and vincristine (PVC)] for each patient at regular intervals (either weeks or months).
After the AI prescribed a dose, it would check a computer model capable of predicting how likely a dose is to shrink a tumor. When the AI prescribed a tumor-shrinking dosage, it received a reward. However, if the AI simply prescribed the maximum dose all the time, it received a penalty.
According to the researchers, this need to strike a balance between a goal and the consequences of an action — in this case, tumor reduction and patient quality of life respectively — is unique in the field of RL. Other RL models simply work toward a goal; for example, DeepMind’s AlphaZero simply has to focus on winning a game.
“If all we want to do is reduce the mean tumor diameter, and let it take whatever actions it wants, it will administer drugs irresponsibly,” principal investigator Pratik Shah told MIT News. “Instead, we said, ‘We need to reduce the harmful actions it takes to get to that outcome.’”
GETTING PERSONAL. The AI conducted about 20,000 test runs for each simulated patient to complete its training. Next, the researchers tested the AI on a group of 50 new simulated patients and found it could decrease both the doses and their frequency while still reducing tumor size. It could also take into account information specific to each patient, such as their tumor size, medical history, and biomarkers.
“We said [to the model], ‘Do you have to administer the same dose for all the patients? And it said, ‘No. I can give a quarter dose to this person, half to this person, and maybe we skip a dose for this person,’” said Shah. “That was the most exciting part of this work, where we are able to generate precision medicine-based treatments by conducting one-person trials using unorthodox machine-learning architectures.”
The AI will still need to undergo further testing and vetting by the Food and Drug Administration (FDA) before doctors could put it into practice. But if it passes those tests, it could eventually help people with glioblastoma attack their brain tumors without causing them more pain in the process.