Might AI Help Sharpen Dementia Diagnosis?

When a person arrives at their doctor’s office having trouble thinking clearly or remembering things, one or a combination of several neurodegenerative diseases could be to blame. At this year’s AAIC, held July 12–15 in London, Daisuke Ono, working in Dennis Dickson’s lab at Mayo Clinic in Jacksonville, Florida, presented an AI model that might help point clinicians to the correct diagnosis. Using medical records from people who had undergone autopsy, the model achieved an AUC around 0.95 for amyotrophic lateral sclerosis and multiple system atrophy; for Alzheimer’s, it was 0.83. The AI might eventually help primary care physicians identify people to refer for further testing.

• AI compared medical records with autopsy data to develop prediction models.
• The best-performing model posted an AUC of 0.83 for Alzheimer’s disease
• Scientists are developing an iPad app for use in primary care.

In primary and secondary care, biomarker tests, such as PET scans or immunoassays of blood or CSF, have proved invaluable for helping clinicians distinguish among different causes of cognitive dysfunction, but they can be costly, time-consuming, and are not available in every clinic. Previously, scientists led by Nima Aghaeepour and Thomas Montine at Stanford University in California developed an AI model to predict specific neuropathologies from longitudinal clinical data, such as cognitive test scores and symptoms, collected in research cohorts contributing to the National Alzheimer’s Coordinating Center (NACC) database (Phongpreecha et al., 2023). Ono and colleagues developed a similar model but based on routine medical records.

The records came from people who had died with a neurodegenerative disease confirmed by autopsy. While alive, they had developed cognitive dysfunction—such as memory problems, changes in behavior or personality, and speech and language difficulties. Ono included people whose cognitive issues were their first symptom or who had initially developed other symptoms, such as hallucinations, REM sleep behavior disorder, or motor symptoms, and then developed cognitive problems within three years.

Their final cohort, of 2,785 people, included 870 people who had had Alzheimer’s disease (AD), 538 with progressive supranuclear palsy (PSP), 506 with AD plus Lewy body disease (LBD), 194 with frontotemporal lobar degeneration (FTLD), 178 with corticobasal degeneration (CBD), 164 with LBD, 121 with PSP plus AD, 118 with multiple system atrophy (MSA), and 96 with amyotrophic lateral sclerosis (ALS).

Ono then tweaked ChatGPT-4 to comb through clinical notes in their health records and pick out 197 neurological symptoms. Then, using age at onset of cognitive dysfunction, along with when each symptom first occurred, sex, and family history, Ono trained six machine-learning models—CatBoost, LightGBM, XGBoost, Random Forest, a multilayer perceptron (MLP), and a stacked ensemble—to learn which clinical patterns corresponded to each diagnosis. These models differ in how they learn patterns from data: Some build and combine decision trees, while others use neural networks. To see how each model stacked up, Ono used fivefold cross-validation, training each model on 80 percent of the cases and testing it on the remaining 20 percent until every case had served as a test case. He first did this using clinical data available through one year after the onset of cognitive dysfunction, then repeated the process, adding progressively more data from the following years.

CatBoost performed the best, with an overall AUC of 0.77 across all disease diagnoses using data up to one year after the onset of cognitive symptoms. With two additional years of clinical data, that climbed to 0.82. It found some diseases easier to identify than others. For ALS, MSA, and PSP, AUCs were 0.96, 0.95, and 0.88, respectively. For AD alone, the AUC was 0.83. Mixed pathologies proved more challenging. AUCs for LBD-AD and PSP-AD were 0.74 and 0.73, respectively.

Looking at which features the model weighed most heavily revealed that disorientation and memory loss, as well as the absence of swallowing difficulties, slurred speech, and other parkinsonian features, steered it toward and AD diagnosis. Male sex, tremor, and hallucinations were among the strongest indicators for LBD.

The next step, Ono told the audience, is to develop an iPad-based application where someone concerned about their health, or a caregiver, could enter their clinical symptoms. Their primary care physician would then receive a notification listing the probabilities of different diagnoses and could order follow-up blood tests or PET scans. “I think it’s in the near future,” he said.—George Heaton

Find this and more great content on Alz Forum – https://www.alzforum.org/news/conference-coverage/might-ai-help-sharpen-dementia-diagnosis