06 March 2020

Predicting Dementia from Medical Records

Welcome back. Other than occasionally forgetting a name or the right word, I guess I’m doing OK for my age. Alzheimer's disease is the most common cause of dementia among older adults. At least half of those living with the disease and related dementias never receive a diagnosis. Many more live with symptoms for years before being diagnosed.
Neuropsychological testing for dementia or Alzheimer’s disease (from dailycaring.com/diagnosing-alzheimers-or-dementia-neuropsychological-testing/).
As yet, there is no cure, but early identification can provide the opportunity for screening and delaying the disease onset. Although current tests to screen for dementia risk are invasive, time-consuming and expensive, there may be a breakthrough.
Brain imaging, such as by CT, MRI and PET, for studying structural and biochemical changes associated with Alzheimer's disease and related dementias (photo of amyloid PET scan from www.medicalnewstoday.com/articles/324877).
Two recent studies were able to predict dementia without specialized testing. The level of accuracy, especially for predicting dementia within one year of its onset, was suitable for automated pre-screening.

Machine Learning of Electronic Medical Records
The two studies were conducted by largely the same team of researchers affiliated with Indiana University-Purdue University at Indianapolis, Regenstrief Institute, Merck & Co., Albert Einstein College of Medicine and Indiana University School of Medicine. Collaborators from Georgia State University and Solid Research Group LLC participated in one of the studies.

Both studies achieved early identification of Alzheimer's disease and related dementia by applying machine learning, a subset of artificial intelligence, to routine medical care data widely available through electronic medical records.

Electronic medical records offer a rich source of data for machine-learning analysis (photo from nursingeducation.lww.com/blog.entry.html/2017/02/14/virtual_simulationa-lMOf.html).
The analyzed data were from the records of patients with dementia (cases) and patients without dementia (controls) of comparable age, race and sex. Most patients were from multiple institutions, and all of the data were from the Indiana Network for Patient Care, a regional health information exchange.

Modeling Approaches
The studies took different approaches to model development.

To predict dementia one year and three years prior to disease onset, one study relied on training and testing with data sets for time periods covering 1 to 10 years and 3 to 10 years. Training for one-year prediction was based on 1,728 cases and controls, while testing used 431 cases and over 9,800 controls. About half the number of cases and controls were analyzed for the three-year prediction.

The second study used area under the receiver operating characteristics curve to determine the best models for 1 to 10 years, 3 to 10 years, and 5 to 10 years. The derivation sample consisted of 10,504 cases and 39,510 controls; the validation sample included 4,500 cases and 16,952 controls.

Structured and Unstructured Data
The data sets used by each study included structured data--drug prescriptions and diagnoses--as well as unstructured data--medical notes on visits, progress and medications. The medical notes, a sequence of records identifying the patient and date, followed by a list of reports, each in free text, were the best source of predictive features.

Each study tested models that isolated and combined the different data sets, and each obtained the highest accuracy with combined structured and unstructured data. The models were found to be unaffected by biases related to race, sex or the source institution.

Wrap Up
The researchers are continuing to explore development of dementia models with machine learning and electronic medical records data.

If model performance is confirmed to be comparable to that based on specialized medical tests, the approach would be accessible to the general patient population at a reduced cost. Minimally, the modeling could be used to pre-screen for targeted medical tests for at-risk patients.

Once again, stay tuned. And again, thanks for stopping by.

Key facts about Alzheimer’s disease: www.cdc.gov/dotw/alzheimers/
Biomarkers for dementia detection: www.nia.nih.gov/health/biomarkers-dementia-detection-and-research
Studies of predicting dementia from electronic medical records in Artificial Intelligence in Medicine journal and Jour. of American Geriatrics Society:
Article on studies on EurekAlert! website: www.eurekalert.org/pub_releases/2020-02/ri-sdn021020.php

No comments:

Post a Comment