By Deborah Borfitz 

February 22, 2023 | Research scientists at Indiana University–Purdue University Indianapolis (IUPUI) and Regenstrief Institute have developed a “passive digital marker” (PDM), powered by machine learning, to help identify people at risk for Alzheimer’s disease and related dementias (ADRD) at an earlier stage in primary care settings. Application of the PDM will now be tested in the real world in a pair of pragmatic randomized controlled trials, the first of which is already underway in Indianapolis, according to Malaz Boustani, M.D., a geriatrician, neuroscientist, and implementation scientist at the Regenstrief Institute and the Indiana University School of Medicine. 

The algorithm itself was validated several years ago in a study that published in Artificial Intelligence in Medicine (DOI: 10.1016/j.artmed.2019.101771), where it was shown to achieve 80% and 77% accuracy, respectively, for one- and three-year prediction horizons based on routine care data found in the electronic health record (EHR). What now must be established is whether it can change the behavior of clinicians in the wild, Boustani says. 

To move the needle on ADRD detection using the marker, patients at risk must first get that message from their physician who might be too busy or distracted to have the conversation or perhaps aren’t convinced there’s a problem, he continues. Patients, when told they’re at heightened risk, must also respond by following their physician’s advice to go to a memory clinic for diagnostic confirmation. 

As with mammography screening, the PDM is of no value without that next step, notes Boustani. That requires changing the behavior of doctors, so they effectively present the data to patients and follow through on the alerts. 

The goal here is to lower the number of unrecognized cognitively impaired people in the U.S. healthcare system, he says. The troubling reality is that 50% to 80% of dementia cases go unnoticed—and an even higher proportion if patients living with mild cognitive impairment are included. The Digital Detection of Dementia (D Cubed) study protocol, as described in a recently published article in Trials (DOI: 10.1186/s13063-022-06809-5), will evaluate use of the PDM alone and in concert with a patient-reported outcomes survey compared to usual care in affecting change. 

Lack of early recognition of ADRD has multiple contributing factors, among them the limited time primary care clinicians have with patients and the stigma of dementia, Boustani says. Memory care also involves a lot of non-medical costs that the Centers for Medicare & Medicaid Services (CMS) and most private insurance companies do not currently cover. 

Twin Trials 

Two identical and independently powered trials comprise the D Cubed study, which is being funded by a five-year grant from the National Institutes of Health’s National Institute on Aging. Only the enrolled population will look different, says Boustani. “We are trying to mimic the way FDA [Food and Drug Administration] approves medication,” namely through the “reproducibility principle.”   

The trial has been enrolling patients since last July who are being seen in primary care clinics at federally qualified health center (FQHC) sites in Indianapolis affiliated with Eskenazi Health in partnership with the IU School of Medicine, he says. Participants are expected to be predominantly people who are Black, socially frail, reside in urban areas, and insured through the federal Medicare and state Medicaid programs.  

The second trial is scheduled to launch next month at primary care clinics affiliated with the University of Miami, continues Boustani. Here, the participants are likely to include mostly Hispanics and people living in rural areas. 

Both trials will run for two years and are anticipated to have a total enrollment of between 3,500 and 4,000 individuals across the two sites. Eligibility criteria include being at least 65 years old, having had at least one visit to a primary care practice within the past year, and the ability to communicate in English or Spanish. EHR data from at least the past three years must also be available. Individuals residing in a nursing home or with serious mental illness are not eligible.  

At each location, patients at three of nine clinics will be randomized to usual care, three to just the PDM, and three to the PDM plus the patient-reported outcomes survey, he explains. The survey, called the Quick Dementia Rating System (QDRS), was developed by cognitive neurologist James E. Galvin, M.D, a professor of neurology at the University of Miami’s Miller School of Medicine. The tool is designed to capture real-world patient functioning, cognition, mood, and behavior that is not routinely captured in the EHR. 

Data Inputs 

Three primary buckets of data from the EHR feed the PDM, including structured diagnostic and prescription information as well as medical notes where natural language processing (NLP) is deployed to convert text to a standardized format decipherable by machine learning, says Boustani. These inputs include mentions of memory issues, notations of vascular concerns, and comorbid conditions potentially linked to dementia. 

Development of the algorithm is credited to Zina Ben Miled, Ph.D., IUPUI associate professor of electrical and computer engineering in the Purdue School of Engineering & Technology and co-principal investigator with Boustani on the D Cubed study. 

When working in the EHR, physicians are known to do what Boustani terms “data collection below the neck,” such as documenting information about patients’ heart and other organs. Data about their mood and cognition is not captured in a structured way, but it is sometimes captured in their note-taking, and tapping those clues therefore improves the algorithm’s performance. 

But even that “might not be enough,” he adds, which is why the QDRS is being included in the D Cubed study. “It might substantially improve the performance of the algorithm, and potentially the response of patients and doctors to the algorithm.” 

An estimated 6% of patients being seen in primary care settings are impacted by dementia but only 2% of them are recognized as such, reports Boustani. The researchers anticipate their PDM, possibly in tandem with the QDRS, will double the number of cases identified. 

The PDM uses a neural network method that teaches computers to process data in a way that is inspired by the human brain, Boustani says, and was trained and tested using EHR data of incident ADRD cases and non-ADRD controls. In its current version, it is being used solely to classify risk—i.e., if a person is or isn’t cognitively impaired and at heightened chance of having Alzheimer’s disease or another related dementia over the next one to three years.  

Behavioral ‘Nudges’ 

For the D Cubed study, the output of the PDM is being displayed directly to physicians as behavioral “nudges” within the constraints of the existing EHR (in this case Epic) as well as indirectly to patients through Epic’s MyChart patient web portal, says Boustani. In other contexts, the PDM would work just as well with any other EHR, including the systems offered by Cerner and Allscripts. 

Earlier last year, Geisinger announced it was testing the algorithm’s effectiveness. Its ability to predict dementias there was similar, if not slightly better, than what was previously achieved at IUPUI. Kaiser Permanente is also conducting a pragmatic trial looking at the potential of readily available EHR data elements to identify patients at high risk for dementias, including Alzheimer’s. Boustani is a consultant on this trial, and separately working on a systematic review of all existing passive digital markers that he hopes will be accepted for publication soon.  

Using the PDM as an alert for screening purposes is inherently inexpensive because it identifies people who may be at risk for Alzheimer’s disease without the need for invasive and costly tests—or even having humans in the loop, says Boustani. However, Indiana University is now commercializing the PDM to the start-up company DigiCARE Realized, cofounded by Boustani, which has the expertise to convert the open-source algorithm into a market-ready product.  

Collaborative Care 

The diagnostic assessment offered by Eskenazi Health is like that available at the University of Miami, both of which have specialized Alzheimer’s disease centers, Boustani points out. They follow gold-standard practices for conducting a diagnostic evaluation for memory problems.  

But the way dementia gets managed post-diagnosis differs considerably between the two locations, he adds. At Eskenazi Health, “usual care” is an evidence-based collaborative dementia care model where patients are assigned a care coordinator assistant who works with them and their family members to manage behavioral symptoms, reduce caregiver stress and inappropriate hospitalizations, and keep people living at home longer, lowering the overall costs to them and to the healthcare system.  

IU is working with the AARP and Alzheimer’s Association to ensure the type of dementia care provided at Eskenazi is available to dementia care patients nationwide, says Boustani. The holdup is reimbursement, since the collaborative care model is not currently reimbursed by CMS. 

As a non-profit FQHC, Eskenazi Health is eligible for CMS support and grants from other government agencies as well as the private sector to offer collaborative care, he continues. CMS has recently signaled it would consider experimenting with an alternative payment model that would cover expenses associated with delivering Eskenazi-style dementia care elsewhere.