AI brain MRI analysis tool aids Alzheimer’s disease prognostication

A brain MRI–based artificial intelligence (AI) tool developed by researchers from the Chinese University of Hong Kong (CUHK) outperforms MRI markers such as hippocampal volume (HV) in predicting syndromal conversion of mild cognitive impairment (MCI) to Alzheimer’s disease (AD) dementia, according to new data presented at Advances in Medicine (AIM) 2023.

 

The tool, based on the AD–Resemblance Atrophy Index (AD-RAI), previously demonstrated higher accuracy than single-region brain volumetrics in AD identification. [J Alzheimers Dis 2021;79:1023-1032] It is commercially available in Hong Kong to assist clinical detection of early AD, generating a report in 10 minutes after brain MRI images are uploaded to a secure cloud server. [https://www.cuhkmc.hk/specialties/neurology/accubrain-ai-brain-image-analysis]

 

“AD-RAI is a machine learning algorithm based on data sets of 180 AD dementia patients and 223 healthy controls. On a scale of 0–1, the threshold for differentiating AD from controls is 0.5, and a score of 1 indicates complete similarity to AD brains,” said Professor Vincent Mok of the Division of Neurology, CUHK.

 

In an external validation cohort comprising 50 AD patients and 50 controls, AD-RAI demonstrated an area under the receiver operating characteristic curve (AUC) of 0.92. “AD-RAI outperformed single-region volumetrics in diagnosing AD dementia,” said Mok.  [J Alzheimers Dis 2021;79:1023-1032]

 

In a more recent study using data of individuals with MCI (n=363) and unimpaired cognition (n=226) from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database, AD-RAI outperformed HV in predicting conversion of MCI to AD dementia on top of clinical features (ie, age, gender, education, and baseline Montreal Cognitive Assessment [MoCA] score). [Alzheimers Dement 2023;doi:10.1002/alz.13083]

 

“[At baseline,] all study participants had cerebrospinal fluid [CSF] analysis of amyloid b and tau to define who had underlying AD,” explained Mok.

 

Among participants with MCI, mean AD-RAI score was 0.7 in those with amyloid b and tau in CSF who converted to dementia (n=85; mean baseline Mini-Mental State Examination [MMSE] score, 27.2; mean baseline MoCA score, 21.5) over 4 years of follow-up, compared with 0.4 in those without amyloid b and tau in CSF who did not convert to dementia (n=278; mean baseline MMSE score, 28.3; mean baseline MoCA score, 23.7) (p<0.001). Of all independent variables, AD-RAI demonstrated the strongest association with conversion to dementia (adjusted odds ratio, 10.016; 95 percent confidence interval, 4.367–22.972; p<0.001) on multivariate logistic regression.

 

“AD-RAI outperformed other MRI-based markers, such as HV, and was almost as good as plasma p-tau₁₈₁ [AUC, 0.779 for AD-RAI, 0.748 for HV, 0.788 for plasma p-tau₁₈₁] in predicting conversion of MCI to AD dementia [on top of clinical features],” highlighted Mok. “Combining [clinical features with] AD-RAI and plasma p-tau₁₈₁ yielded an AUC of 0.828, and adding apolipoprotein E e4 further increased AUC to 0.853.”

 

However, access to brain MRI may be limited by long waiting times in public hospitals and costs in the private sector. “We may order CT to rule out other diseases for public elderly patients with memory problems,” suggested Mok.

 

Retinal photography–based AI model for AD detection

“In the coming years, we may consider using retinal photographs to rule in AD,” said Mok. “Our deep learning model based on retinal photographs from 648 AD patients and 3,240 controls can detect AD dementia with good accuracy, demonstrating AUC of 0.93 in an internal validation cohort and 0.73–0.91 in external validation cohorts.” [Lancet Digit Health 2022;4:e806-e815]

 

“We are evaluating the model for detection of early AD in an ongoing study,” he added.