Artificial neural networks have long been used to understand how brain processes are affected by drugs or environmental toxins, but they’re not yet ready for the job of predicting whether or not a patient will develop a particular brain disorder.
Now a team of researchers has developed a computer algorithm that can do just that.
The new system uses deep learning to build and analyze large datasets of data in order to predict a patient’s risk of developing a specific brain disorder, such as Parkinson’s or Alzheimer’s.
The researchers describe their work in the journal Science Advances.
The research team, led by MIT neuroscientist James Corbett, and graduate student Jie Zhu, used a new method called deep convolutional neural networks to build the system.
These neural networks are a class of deep learning algorithms that are widely used in computer vision and machine learning, but are not used for much else.
In a way, deep convolutions are a little like a neural network in the traditional sense: Each neuron is an image that is fed into a pipeline of images and the resulting output is an output image.
Corbett and Zhu used this pipeline to build an algorithm that could predict a person’s risk for Parkinson’s disease by analyzing their images and their neural connections.
In this case, the model is modeled after a human brain that has the exact same characteristics as the brain of a patient with Parkinson’s.
Corbet and Zhu say that the model uses deep convolving neural networks in order do this: they use a technique called convolution to train a neural net to model the image and neural connections of the brain in real time.
The network then builds a model that computes the probability of a specific image being the same as the image that the person is thinking about.
In other words, the network learns the likelihood of the person thinking about that particular image, based on the information that is being fed into the network.
“What this means is that the network doesn’t need to know anything about the brain to predict what the person might be thinking about,” Corbett told Phys.org.
“The model is built in real-time, and the model does it all in realtime.”
Using the same method, Corbett’s team built a model to build for the first time a model of a human with Alzheimer’s disease.
This model can be used to predict whether or that person will develop Alzheimer’s, a disease that is often found in older people.
“We’re not going to be able to tell if a person is going to develop Alzheimer and that they’re going to go on to develop Parkinson’s,” Corbet said.
“But we can tell if they are going to have a brain lesion.”
This type of model is a very powerful tool for studying disease and understanding how different genetic mutations affect the brain, said Zhu, who worked on the research with Corbett.
“It’s an important step forward in the field because the models that are used in the medical field, they’re very complicated and they’re often built on a single-dimensional, finite-dimensional model.”
The work also shows that deep convolved neural networks can be more efficient than traditional neural networks.
“These systems can perform better than traditional network architectures,” Zhu said.
The deep convolve networks also work at a much higher rate than a simple neural network, which means they can be trained on much larger datasets.
The researchers say that this new method could be useful in the future to develop more accurate models of the disease.
“What is important for us right now is to use this system to train models for Parkinson, Alzheimer’s and other neurodegenerative diseases,” Zhu added.
The work was funded by the National Institutes of Health.