YuLin Zhen, Photography Editor

This August, a group of Yale researchers received a $20.6 million federal grant to conduct long-term research on mental illness. 

This expansive study will bring together researchers in psychiatry, biostatistics, neuroscience and more to build a predictive machine-learning model that can enhance patient diagnoses. This project was born from the National Institute of Mental Health’s new initiative, called Individually Measured Phenotypes to Advance Computational Translation in Mental Health, or IMPACT-MH. 

According to Christopher Pittenger, a professor of psychiatry and one of the project leaders, there is unexplored overlap between symptoms of different mental illness diagnoses. Patients are diagnosed based on lists of symptoms instead of identifying underlying causes.

“This IMPACT program is an effort to take a big step back and say, let’s not be guided by [current] diagnoses,” Pittenger said. “Let’s not just study people with depression or OCD or PTSD or addiction. Let’s study them all and measure as much as we can and see if we can develop new ways to think about mental illness.”  

The Yale IMPACT-MH program aims to recruit 2,400 patients over five years, divided into three waves. Each wave will follow 600 participants with a psychiatric diagnosis and 200 participants without a diagnosis for two years, tracking as many variables as possible.

The project has three main goals: finding measures that predict mental illness progression, identifying mental illness patterns over time and looking at how symptoms cluster and overlap to refine the current mental illness diagnosis system. 

“Depression is kind of like pain,” Pittenger said. “If a doctor said you have pain, that would be a fairly useless diagnosis because that pain could be because of a cut versus nerve damage versus a headache. Those are totally different kinds of pain that require totally different kinds of treatment.”

The researchers are utilizing a combination of assessments that are cheap and easy to administer, Pittinger said, helping psychiatrists avoid more expensive examinations.  

According to Sarah Yip, project lead and professor of psychiatry, the project will combine traditional clinical measures like diagnostic interviews and self-report scales with new behavioral tasks. These new tasks are designed to track subtler changes in the patient’s mental state. For example, researchers may task patients with choosing between risky or safe monetary gambles to assess their perception of risk and reward.

Researchers will analyze patients’ spoken narrative data using artificial intelligence models specifically designed to  identify trends in thought patterns and self-image, according to Yip. 

“An individual isn’t just their depression score,” Yip said. “Instead of this focus on single variables in isolation, by focusing on interactions between multiple variables that we think are highly relevant, the hope is that we can get a more nuanced understanding of the individual as their symptoms are changing over time.”

Beyond behavioral markers, Yip told the News, the researchers also plan to collect data from patients’ electronic health records to better understand social determinants of health. For example, they’ll aggregate information about a patient’s available social support, insurance status and changes in their medication to have a more comprehensive view of how external factors may be contributing to their mental state. 

With each wave of patients, the researchers plan to refine the behavioral markers they are testing for. As they analyze the data, they hope to identify what measures strongly relate to the patient’s mental state and future trajectory. 

Part of the project involves producing a machine learning model that can analyze this deluge of data, according to Yize Zhao, a Yale professor of biostatistics. Researchers will feed the data to the AI model, allowing them to analyze trends such as when patients feel better or worse. 

“Based on their performance on those tasks, we can give them a [computational] fingerprint,” Zhao said. “And the hope is eventually we can find we can eventually give a prediction of their future behavior.”

Additionally, the researchers hope the machine learning model will be able to group sets of patients together that — based on the data collected — may respond favorably to a specific treatment.

According to Yip, the project is a first step towards a broader goal to make psychiatric treatment predictive instead of reactive. 

“The ideal goal would be that we end up with actionable sets of measures that can be used to more accurately predict how an individual is going to be doing in terms of their symptoms over time,” Yip told the News. “Whether or not it’s realistic to expect that within the context of this first study, I don’t know if that’s realistic.”

Godfrey Pearlson, professor of psychiatry and neuroscience, is also leading the project.