The model predicts a cognitive decline due to Alzheimer's disease up to two years

A new model developed at MIT can help predict whether patients at risk for Alzheimer's disease will experience clinically significant cognitive decline as a result of the disease, anticipating the results of their cognitive tests in the next two years.

The model could be used to improve the selection of drug candidates and cohorts of participants in clinical trials, which have been notoriously unsuccessful so far. It would also allow patients to know that they may experience rapid cognitive decline in the months and years to come, so that they and their loved ones can prepare themselves.

Pharmaceutical companies over the past two decades have injected hundreds of billions of dollars into Alzheimer's research. Yet, the field has failed: between 1998 and 2017, 146 unsuccessful attempts to develop drugs to treat or prevent the disease were unsuccessful, according to a report published in 2018 by Pharmaceutical Research and Manufacturers of America. During this period, only four new drugs were approved, solely to treat the symptoms. More than 90 drug candidates are under development.

Studies suggest that the success of drug marketing could result in the recruitment of candidates at an early stage of the disease, before the symptoms become evident, during which time the treatment is most effective. In an article that will be presented next week at Machine Learning for Healthcare, researchers at the MIT Media Lab describe a machine learning model that can help clinicians focus on this cohort of participants.

They first formed a "population" model on a comprehensive data set including clinically significant cognitive test scores and other biometric data from patients with Alzheimer's disease, as well as of healthy individuals, collected between the bi-annual visits of the doctor. From the data, the model identifies models that can help predict patients' cognitive test scores between visits. In new participants, a second template, customized for each patient, continuously updates scoring forecasts based on newly recorded data, such as information collected during the most recent visits.

Experiments indicate that accurate forecasts can be made for the next six, 12, 18 and 24 months. Clinicians could use this model to select at-risk participants in clinical trials that may demonstrate rapid cognitive decline, possibly even before the onset of other clinical symptoms. Treating these patients at an early stage can help clinicians better identify which anti-dementia drugs work and do not work.

"Accurate prediction of cognitive decline from six to 24 months is critical for clinical trial design," says Oggi Rudovic, a researcher at the Media Lab. "Being able to accurately predict future cognitive changes can reduce the number of visits the participant must make, which can be costly in time and money, and contribute to the development of a useful drug. The aim is to help reduce the cost of clinical trials, make them more affordable and scaled up. "

Yuria Utsumi, an undergraduate student, and Kelly Peterson, a graduate student in the Department of Electrical and Computer Science; Ricardo Guerrero and Daniel Rueckert, both of Imperial College London; and Rosalind Picard, professor of arts and media science and director of research in affective computing at the Media Lab.

Population to personalization

For their work, the researchers harnessed the world's largest set of clinical trial data on Alzheimer's disease, called the Alzheimer's Disease Neuroimaging Initiative (ADNI). The dataset contains data from about 1,700 participants, with and without Alzheimer's disease, recorded during semi-annual doctor visits over 10 years.

The data include their ADAS-Cog13 scores (AD Assessment-scale), the most widely used cognitive metric for clinical trials of drugs for the treatment of Alzheimer's disease. The test evaluates memory, language and orientation on an increasing gravity scale of up to 85 points. The dataset also includes MRI scans, demographic and genetic information, and cerebrospinal fluid measurements.

In total, researchers trained and tested their model on a sub-cohort of 100 participants, who made more than 10 visits and had missing data below 85%, each containing more than 600 computable characteristics. Of these participants, 48 ​​were diagnosed with Alzheimer's disease. However, data is scarce and most participants do not have different combinations of characteristics.

To solve this problem, the researchers used the data to form a population model based on a "non-parametric" probability framework, called a Gaussian process (GP), with flexible parameters that can be adapted to various probability distributions. to deal with the uncertainties of the data. This technique measures the similarities between variables, such as patient data points, to predict a value for an invisible data point, such as a cognitive score. The output also contains an estimate of the degree of certainty of the prediction. The model works well even when badyzing datasets with missing values ​​or a lot of noise from different data collection formats.

But, by evaluating the model on new patients from a portion of the participants, the researchers found that the model's predictions were not as accurate as they could be. So they customized the population model for each new patient. The system would then progressively fill in missing data with each new patient visit and update the ADAS-Cog13 score forecast accordingly, continually updating previously unknown distributions from GPs. After about four visits, custom templates have significantly reduced the error rate in the forecasts. He also outperformed various traditional machine learning approaches used for clinical data.

Learn to learn

But the researchers found that the results of the custom models were still suboptimal. To solve this problem, they invented a new "metal-learning" system that learns to automatically choose the type of model, population or custom that best suits a given participant at a given time, based on the badyzed data. . Metalearning has already been used for computer vision and machine translation tasks to learn new skills or adapt quickly to new environments with some training examples. But this is the first time it has been used to monitor the cognitive decline of patients with Alzheimer's disease, where limited data is the main challenge, says Rudovic.

The schema essentially simulates how different models work on a given task (such as predicting an ADAS-Cog13 score) and learns the best fit. At each visit of a new patient, the schema badigns the appropriate model, based on the previous data. With patients with noisy and rare data during early visits, for example, population models make more accurate predictions. When patients start with more data or collect more on subsequent visits, custom templates work better.

This has reduced the forecast error rate by an additional 50%. "We could not find a single model or a fixed combination of models that could give us the best prediction," says Rudovic. "So we wanted to learn to learn with this Metalearning scheme. It is a model based on a model that acts as a selector, formed with metaconnections to decide which model should be deployed.

Next, the researchers hope to form partnerships with pharmaceutical companies to implement this model in the context of clinical trials on Alzheimer's disease. Rudovic says the model can also be generalized to predict various measures of Alzheimer's disease and other diseases.

This story is republished with the permission of MIT News (web.mit.edu/newsoffice/), a popular site that covers the news of MIT's research, innovation, and teaching. .