Predicting Behavior and Health Of Individuals Using Machine Learning: Why Do Brain Models Fail?

Scientists have used machine learning to understand how the brain creates complex human characteristics. This has revealed patterns of brain activity that are associated with actions like working memory, traits like impulsivity, as well as conditions such depression. These models can be used by scientists to predict people’s behavior, and even their health.

It only works if models are representative of everyone. Past research has proven otherwise. There are some people who don’t fit into every model.

Yale University researchers have examined the causes of these failures and how to correct them in a recent study published in Nature.

Abigail Greene (M.D.-Ph.D. student, Yale School of Medicine) is the research lead. She believes that models should be adaptable to each individual in order to be of the greatest benefit.

Greene and her colleagues are looking at two ways that models might be able to provide more accurate psychiatric categorization. A diagnosis of schizophrenia can cover a broad range of symptoms that may vary from one person to the next. If researchers have a better understanding of the neuro underpinnings of schizophrenia and its symptoms, they may be able to classify people more accurately.

Second, impulsivity is a characteristic of many conditions. Understanding the neural basis for impulsivity can help doctors tackle this symptom more effectively, regardless if the medical diagnosis is correct.

Greene stated that both advancements would have consequences for treatment responses.

She said that models must be universally applicable to everyone first.

Greene and her colleagues developed models that used patterns of brain activity to predict the scores of people on cognitive tests. The models predicted the scores of most people when tested. However, some models were wrong. They incorrectly predicted that people would score low when they actually scored high.

The research team looked into the failures of models to correctly categorize people.

The next step was to determine if similar misclassifications could possibly be explained by brain differences. There were no consistent differences. They found that misclassifications were influenced by sociodemographic factors such as age, education, and clinical factors such as the severity of symptoms.

Greene explained that the models were actually reflecting complex profiles — a mix of cognitive abilities and other sociodemographic and clinical factors.

One example is that models in the study linked higher cognitive scores with more education. Individuals with lower education scores who score well were not included in the model’s profile, and they were often incorrectly predicted to be low scorers.

The model was not able to access sociodemographic data, adding to the complexity.

Greene explained that the cognitive test score is influenced by sociodemographic variables. The results can be affected by biases in the way cognitive tests are administered, scored, and interpreted. Bias can also be an issue in other areas; research has shown that biases in input data can affect models used in criminal justice or health care.

Greene stated that test scores are composites of cognitive ability and other factors. The model predicts the composite. Researchers need to be more aware of what is being measured and therefore what a model can predict.

The authors of the study offer several suggestions to address this problem. Scientists should use strategies to minimize bias and maximize validity during the study design phase. Researchers should use statistical methods to correct stereotypical profiles after collecting data.

These measures will result in models that are more representative of the cognitive construct being studied, researchers claim. However, it is impossible to eliminate bias completely so it is important to acknowledge this when interpreting the model output. For some measures, more than one model may be necessary.

Todd Constable, a Yale School of Medicine professor of radiology/biomedical imaging and the senior author of this study, stated that “there’s going to come a time when you just need different models” “One model does not fit all.”

What is Machine Learning?

Machine learning has been around since the beginning. Arthur Samuel, a computer scientist at IBM who pioneered AI and computer gaming, coined the term “machine-learning”. The program learned more from playing, and used algorithms to predict future events.

Machine learning is a discipline that studies the analysis and construction of algorithms that can learn from data and make predictions.

ML is a valuable tool because it solves problems faster than the human brain can. Machines can learn to recognize patterns and relationships in input data and automate repetitive tasks by using large amounts of computational power.

Data is key: Machine learning algorithms are crucial to the success of machine learning. Machine learning algorithms use sample data (or “training data”) to build mathematical models that can make predictions and decisions. They are not programmed to do this. This information can be used by businesses to improve their decision-making, optimize efficiency, and capture actionable data on a large scale.

AI is the Goal: ML is the foundation of AI systems that automate business processes and solve data-based problems autonomously. It allows companies to augment or replace certain human abilities. Chatbots, self-driving cars, and speech recognition are some of the common machine learning applications that you might find in real life.

Machine Learning | Deep Learning | Neural Networks

Deep learning and machine learning are often interchangeable, so it is worth paying attention to the differences between them. Deep learning, machine learning, and neural network are all sub-fields within artificial intelligence.

Deep learning and machine learning are different in the way each algorithm learns. Deep machine learning, also called supervised learning, can use labeled data, also known to be supervised learning, but it does not necessarily need a labeled data set. Deep learning can read unstructured data (e.g. text or images) and can determine the features that distinguish different types of data. This reduces the need for human intervention and allows for larger data sets. Deep learning can be described as “scalable machine-learning” according to Lex Fridman.

Machine learning that is “non-deep” or classical, requires more human intervention. The set of features that are used to distinguish between inputs and outputs is determined by human experts. This usually requires more structured data in order to learn.

If the output of any node is above the threshold value, that node activates and sends data to the next layer in the network. Deep learning algorithms or deep neural networks can be defined as neural networks that have more than three layers. This would include the input and output. A basic neural network is one that has only three layers.

Deep learning and neural networks have been credited with speeding up progress in areas like computer vision, natural speech processing, and speech recognition.

Also Read: Quantum Cryptography: Hacking Impossible

4 thoughts on “Predicting Behavior and Health Of Individuals Using Machine Learning: Why Do Brain Models Fail?”

Leave a Comment