A research team at Kennesaw State University, one of Georgia's largest universities, is developing new systems and software to help Cobb County first responders appropriately manage mental health 911 calls.

This technology brings hope that Cobb first responders can use the system and software throughout Georgia and outside state lines.

According to a KSU press release, "The idea began when Cobb Police and Fire recognized the need for help dealing with the overwhelming number of 911 calls they received about people in emotional distress."


The initiation

About six years ago, Cobb Fire Department contacted KSU about the many phone calls the department received. These calls were not typical fire emergencies or medical emergencies but behavior health calls, according to KSU's Monica Nandan.

"That means people have stopped taking their meds or were having an episode for some reason that could be behavioral health or mental health-related," she told GPB. "[First responders] weren't sure because they're not trained for that. They are trained for EMT, CPR and things like that."

Nandan is a social work and human services professor and director of strategic syndicates and social impact at the Wellstar College of Health and Human Services. She serves as the bridge between the university, her college particularly, and the community needs and community partners.

As a result of talking to Cobb police and fire, KSU wanted to know how many 911 calls CCPD received related to mental health.

"They said about 2,500 give or take, in a year," Nandan said. "Just one Cobb [County] call center gets 800,000 calls annually. That's your call volume from one call center, and police and fire almost respond to every call."

Because police don't write an incident report for every call they receive, Cobb police could only provide KSU with an estimated number of mental health calls.

The up-and-coming research team at KSU was interested in looking at CCPD's data to see the volume of calls they mentioned. 

After two years, CCPD was able to grant the research team the data from their 2019 calls, and they began working.


The research

The KSU research team consisted of Nandan, providing her expertise in social work; Dominic Thomas, associate professor of information systems; and Md Abdullah Al Hafiz Khan, assistant professor of computer science.

After the team received the 2019 data from CCPD consisting of about 50,000 case reports, they realized the data needed a lot of work.

"The data is not clean; it's very corrupt, meaning it's in a format that we can't use," Nandan said.

Thomas and his team spent nearly eight months cleaning the data. Once they cleaned, they began writing the algorithm to look for mental health-related words.

This project aimed for the computer to show how many cases have mental health terms. The team intends to look for those words and create software to load onto the police dashboard where police can type a word in the software, accurately telling them how many cases have it.

The software can answer questions after it detects the number of behavioral health cases. Those questions help break the data down into different categories.

"So you can say, it looks like we've got about 8,000 cases in the dataset listed here in terms of the time of day they get reported," Thomas said. "For example, we can see some distribution and that they peak on Friday evening and then Sunday is low."

He explained that the software is intended as to become a tool a team can use to figure out staffing needs.

"It has implications for who you need to hire, and then we can search different cases," he said. "For example, in one search, the most common types are runaway juveniles. Now we understand how many mental health-related runaway juvenile cases we have per year. We can go back to case management and say, 'How many cases can one [social worker] handle in one year?'"

On the computer science end, Hafiz Khan has had to develop artificial intelligence technology and apply natural language processing to build the system.

"We work with natural language processing, which means when we have some physician report, like when we visit a doctor, and they write text reports, they may suggest medication," he said. "We can apply this here to recognize information from the text report with the drug name, medication name, or whether the person has some specific disease."

During the development of the algorithm, Hafiz Khan and his team focused on programming the technology to detect things automatically. But since humans develop algorithms, it will not be "100% perfect," he says.

"They make mistakes, and to resolve that, we plan to involve humans when they do," he said. "AI will not correctly recognize things, like the drug name. Humans can say, 'AI is missing the drug name, and this is why,' so it can automatically update itself next time."

The team continuously worked on building on the 2019 data until they realized they needed more than one year's worth. They wanted to legitimize that the work they were doing was correct.

Cobb County police recently handed the team the 2020, 2021, and 2022 case report data.

"Now we've got new cases that we can run our algorithm on to make sure it works nicely," Nandan said.


Grad students get involved

KSU graduate students have also been a part of the process. From categorizing mental health-related words within police reports to training the computer to recognize such words, their involvement has been essential to developing the software.

"We used this project not only to serve the community — that is goal No. 1 — but as a teaching tool for students in the MSW (Master of Social Work) program, information systems program, and computer science," Nandan said.

Social work graduate students worked with Nandan to highlight keywords often used by police in case reports indicating a behavioral health issue.

"We categorize everything we could to see if it was related to mental health and would note that," MSW student Miyanna Clements-Williamson said. "Then we would see if we noticed the same pattern, because we look at things through a clinical lens. We do a lot of studying for mental health and different diagnoses, so [we can] see [things] that the average person wouldn't."

Clements-Williamson and Moe Winograd, an MSW and business organization student, noticed that the cases were very complex, requiring them to use context in finding those warning signs.

"The largest thing that we both saw was that there weren't many cases that were strictly dealing with specific things like suicide or mental health disorders," Winograd said. "But we immediately knew, 'This is a huge red flag; we got to mark this down.' Despite our biases, we agreed and knew we needed to focus on this."

Both were shocked at how "intense" some police reports were when examined. It gave them a reality of what police have to deal with "day to day."

"The least we could do is give them a little reprieve on the back end," Winograd said. "There are certain cases that are very intense, and we have to know what our boundaries are. Sometimes we had to leave the room, catch our breath, and know our limits."

The team used the highlighted words to help train the computer program to get smarter at recognizing mental and behavioral health terms in cases. According to Nandan, police typically won't use the same language as clinicians but instead "laypersons language."

The team used a human-computer interface, a practice of humans training the computer, which then puts out results for the human to check and see if the results are correct.

Experts in data information systems, including Dominic Thomas and Martin Brown, a fourth-year Ph.D. data science and analytics student, worked with the human-computer interface. 

"I take the police report data that they've given us access to, and I try to teach a machine learning model to learn the thought processes of the social workers," Brown said.

Brown said that after the social work researchers hard-copied and highlighted indicators of mental health-related topics, he taught that information to a computer model or algorithm.

“I’m just trying to give a model information so that it can give police a probability of how extremely related to behavioral health issues is the case that they're working on, so they could send the relevant social workers or parties to help with the case,” he said.

The computer science portion of building the software required the help of computer science students like Abm Adnan Azmee, working on his Ph.D. in computer science, and Gabriel Gillott, a computer science undergraduate student.

Azmee said he works with developing a natural language procession model that can determine potential mental or behavioral-related issues in cases.

"In these cases, there are many linguistic cues which we, as humans, can figure out 'This might be a mental health case, or this might be a behavioral case,'" he said. "When you are talking about one or two cases, we can read it and determine it, but we're talking about hundreds and thousands of cases. It's impossible to go through them manually and do it, so we need to build a model for that."

Throughout working together on this project, the graduate students have learned from each other's areas of expertise in cases where they would typically never cross paths.

“They’re learning to work with each other, understand each other’s language and build sound systems for the police,” Nandan said.


More to do

The team continues to ensure the algorithm is at its highest accuracy with as much data as possible. Within the following year, they will continue to improve the algorithm's accuracy.

"Our goal is to develop a sophisticated AI algorithm to detect with a higher accuracy and precision at more than 95%," Khan said. "We are confident enough with the system, so it's hard to say how long it would take, but maybe in the next year or so, we will at least have something more than a 90% accuracy."

On accuracy, Nandan also said: "We want to get a more accurate percentage now that we have access to more data, so we can take it to the county government and county commissioners and say, 'We need to be more innovative in how we respond to 20-30% of cases that don't need a cop.'"

The KSU research team said the need for this software has been long overdue for years, but they hope their work can benefit first responders and the community.

"This technology, it's so ahead of its time, but I wish it had been implemented years ago," Moe Winograd said. "Just because it's something that we desperately need, and we've needed it for the last 10 years."

He adds, "Hopefully, we can make some change."