A new study, published by researchers from the Regenstrief Institute, Indiana University, and Purdue University, details their low-cost, scalable approach to identifying individuals at risk of developing dementia at a preliminary stage. Although there is no cure, there are many common risk factors that if targeted, can potentially lower the odds of having dementia or slow down the rate of decline in cognitive function.
“It’s important to detect the risk of dementia to provide appropriate care, management, and planning,” said study senior author Malaz Boustani, M.D., MPH, Regenstrief Institute and IU School of Medicine. He also mentioned their efforts to develop a solution to identify potential dementia cases before it’s too late for them. “We do this by using existing information—passive data— already available in a patient’s medical notes for what we call zero-minute assessment, which is less than a dollar.” It is used to develop an individualized dementia risk prediction or provide evidence of mild cognitive impairment.” The study, ‘Dementia Risk Prediction Using Decision-Focused Content Selection from Medical Notes,’ was published in Computers in Biology and Medicine.
Tread uses machine learning to narrow down a subset of phrases or sentences from the patient’s electronic health record as written by the doctor, nurse, social worker, or other provider within a defined observation period relevant to the target outcome. An EHR’s medical notes are narratives regarding the health of the patient written in free text format.
The text information selected from the medical notes to predict dementia risk might be clinician comments, patient remarks, blood pressure or cholesterol values over time, observations of mental status by a family member, or medication history, including prescription and over-the-counter drugs and “natural” remedies and supplements. It helps the patient, the family, and health care providers to access resources, like support groups and Centers for Medicare and Medicaid GUIDE model programs to help people stay in their homes longer.
This also could prompt clinicians to deprescribe medications commonly consumed by older people that are known to adversely impact the brain and spur a dialogue with the patient about similar over-the-counter drugs. Knowing dementia risk may mean consideration by the physician of newly FDA-approved amyloid-lowering therapies that change the course of Alzheimer’s disease.
“Our approach brings together supervised and unsupervised machine learning to extract sentences from thousands of medical notes for each patient, which are relevant to dementia,” study co-author Zina Ben Miled, Ph.D., M.S., said, a Regenstrief Institute affiliate scientist and a former Purdue University in Indianapolis faculty member. Furthermore, this improves predictive accuracy, while also providing the health provider the ability to quickly confirm cognitive impairment by viewing the specific text that was used to fuel the risk assessment from our language model.”
Since the early 1970s, Regenstrief Institute and Indiana University investigators have pioneered the utility of electronic health records. “Our extensive data capture efforts aim to maximize clinical value, surpassing the current overwhelming reliance on electronic health records in medical care,” says study coauthor Paul Dexter, M.D., Regenstrief Institute and IU School of Medicine.
“This study is an excellent and innovative example of how the clinical value from EHRs can be achieved by applying machine learning methods to identify patients at high risk of dementia down the road.” “Indeed, the early identification of dementia will become increasingly important, as new treatments are developed.”
The ultimate beneficiaries of the use of the new technique are patients, and caregivers but the zero-minute assessment at less than a dollar cost has obvious advantages for primary care clinicians who are overburdened and often lack time and training to conduct specialty cognitive testing.
The cohort initially included 26,236 patients: 4,796 cases and 21,440 controls. To ensure balance, the controls were randomly down-sampled to 4,796, resulting in a total of 9,592 patients evenly split between cases and controls. The mean age of cases was 74.43 years, compared to 72.80 years for controls, with females representing 63% and 64% of the groups, respectively. On average, the medical notes from cases and controls contained over 5,000 tokens and 200 sentences. The proposed method achieved the highest AUC (78.43 ± 0.60) and accuracy (70.64 ± 0.90%) compared to benchmark techniques, demonstrating its effectiveness in identifying dementia risk.
Their 5-year clinical trial of the risk prediction tool they are conducting in Indianapolis and Miami is concluding its final year. They will use lessons learned from this trial to carry forward the application of the dementia risk prediction framework into primary care practices. Future work of the researchers includes a fusion of medical notes with other information in electronic health records as well as environmental data.
Reference: Li S, Dexter P, Ben-Miled Z, Boustani M. Dementia risk prediction using decision-focused content selection from medical notes. Comput Biol Med. 2024;182:109144. doi:10.1016/j.compbiomed.2024.109144
