New model predicts patients at risk of developing Alzheimer disease up to 2 years in the future.
Oggi Rudovic, PhD
A new machine learning diagnostic tool could help predict cognition test scores up to 2 years in the future for patients at risk of developing Alzheimer disease (AD).
A team from the Massachusetts Institute of Technology (MIT) has developed new model-based technology to help clinicians better predict if patients with known AD risks will experience a clinically significant cognitive decline.
Oggi Rudovic, PhD, a researcher in the MIT Media Lab, said the tool could help select at-risk patients for clinical trials, leading to better drug development.
"Being able to accurately predict future cognitive changes can reduce the number of visits the participant has to make, which can be expensive and time-consuming,” Rudovic said in a statement. “Apart from helping develop a useful drug, the goal is to help reduce the costs of clinical trials to make them more affordable and done on larger scales."
The MIT researchers first trained a model on the Alzheimer's Disease Neuroimaging Initiative dataset, which included clinically significant cognitive test scores and other biometric data from 1,700 AD patients collected between biannual doctor’s visits, as well as healthy individuals.
The model learns patterns from the data that help it predict how patients will score on cognitive tests taken between doctor’s visits.
Testing the new model on 100 participants who made more than 10 visits and had less than 85% missing data in a sub-cohort study, investigators found that accurate predictions can be made looking ahead 6-24 months in the future.
The team also trained a population model powered by a “nonparametric” probability framework called Gaussian Processes to deal with gaps in missing data from patients.
Then a second model is personalized for each new patient, updating score predictions based on newly recorded data, such as data collected during recent doctor’s visits.
For each patient, a metalearning scheme that automatically chooses the optimal model type is unleashed to provide the most accurate assessment of each participant.
“We couldn't find a single model or fixed combination of models that could give us the best prediction," Rudovic said. "So, we wanted to learn how to learn with this metalearning scheme. It's like a model on top of a model that acts as a selector, trained using metaknowledge to decide which model is better to deploy."
The struggle to both understand AD and develop drugs to treat and prevent some of the symptoms has been evident for the last 20 years, despite significant funding commitments.
There have been 146 unsuccessful attempts to develop drugs that treat or prevent AD from 1998 and 2017, with only 4 new medications approved to treat symptoms, according to a 2018 report from the Pharmaceutical Research and Manufacturers of America. There are currently more than 90 drug candidates in development.
Recruiting trial volunteers to test new AD drugs remains a challenge, with recent studies suggesting that it’s better to use patients who are in the early stages of the disease without symptoms present.
The MIT team is expected to present their new machine-learning model at the Machine Learning for Healthcare conference from August 8-10 at the University of Michigan in Ann Arbor. They also are exploring partnerships with pharmaceutical firms to implement the new model into clinical trials.
Last month, the Cleveland Clinic released their fourth annual “Alzheimer's disease drug development pipeline: 2019,” giving an overview of AD clinical trials in the US.
The investigators, led by Jeffrey Cummings, MD, ScD, director emeritus of Cleveland Clinic Lou Ruvo Center for Brain Health, identified all pharmacologic Alzheimer trials currently in development from Clinicaltrials.gov and found 132 agents currently in 156 clinical trials—28 of which are in 42 phase 3 trials; 74 in 83 phase 2 trials; and 30 in 31 phase 1 trials.