Using machine learning techniques, researchers from the Massachusetts Institute of Technology (MIT) Media Lab, US, have created an AI model that can learn from patient data to determine the fewest doses needed to optimise cancer treatment, while reducing toxicity.
For patients diagnosed with glioblastoma – an extremely aggressive type of brain tumour – a course of chemotherapy, radiotherapy, and medication is likely to be recommended by doctors. But unpredictable side effects from maximum safe dosage, including memory and sleep problems, hair loss, and nausea, can make treatment an unpleasant experience.
Using a method called reinforced learning – inspired by behavioural psychology – the model designs treatment cycles based on the lowest potency needed to shrink a tumour. During the trial, the AI model chose to reduce the dosage by a quarter or half – even skipping doses all together in some cases.
‘We kept the goal, where we have to help patients by reducing tumour sizes but, at the same time, we want to make sure the quality of life – the dosing toxicity – doesn’t lead to overwhelming sickness and harmful side effects,’ said Pratik Shah, Principal Investigator on the team at the MIT Media Lab.
The method rewards or penalises certain actions depending on how they help to achieve the set outcome, which in this case is to reduce tumour size at the same rate as traditional regimens whilst reducing toxicity. Rewards are given a positive value and penalties a negative, for example +1 or -1, which vary depending on the action taken by the model.
‘If all we want is to reduce the mean tumour diameter, and let it take whatever actions it wants, it will administer drugs irresponsibly,’ said Shah. ‘Instead, we said, “We need to reduce the harmful actions it takes to get to that outcome.”’
At each planned dosing interval, the model decides to give or withhold a dose. If the former, it can choose to give the whole dose or just a portion of it. After each interval, the model learns from its feedback system and chooses the smallest dose for the maximum effect and least toxicity.
‘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 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 trails using unorthodox machine-learning architectures.’
The paper will be presented at the Machine Learning for Healthcare conference on 17-18 August 2018 at Stanford University, US.
By Georgina Hines