When people come to their doctor with memory problems, they tend to want to know if Alzheimer's is to blame and, if so, what they can expect. Answering these questions with a degree of certainty is a tall order, and many view amyloid-PET scans and fluid biomarkers as the best options, when available. However, subtle signs of atrophy lurking within routine MRI scans might also foretell AD—and artificial intelligence might help pick them out. On January 12, the FDA approved BrainSee, an AI-driven software platform that uses MRI data, along with scores from routine cognitive tests, to predict the likelihood that a person with mild cognitive impairment will develop AD dementia within five years. Developed by San Francisco-based Darmiyan, BrainSee purportedly detects subtle brain shrinkage patterns that correlate with future AD. The approval follows BrainSee’s 2021 breakthrough designation granted by the FDA.

Nick Fox of University College London, who, among others, has reported that hippocampal atrophy predicts future decline, wondered how much of BrainSee’s prognostic output relies upon this single, routine measure. “It is notable that the FDA approval appears to relate to estimating risk of progression in amnestic MCI (aMCI), since hippocampal atrophy has long been recognized as a predictor of progression from aMCI to AD dementia. This may be the key MRI feature driving prediction inside the ‘black box’ of the AI package,” he wrote.

Fox added that peer-reviewed studies are needed to fully understand the sensitivity and specificity of the approach. One preprint, and findings presented at CTAD conferences in 2020 and 2021, suggest that, among people with amnestic MCI, the algorithm predicts impending AD dementia with an accuracy of 88 to 91 percent (Vejdani et al., 2019; CTAD abstracts, 2020; ROCO4 in CTAD 2021 abstract). 

The approval of this technology is a promising development toward helping patients understand how quickly they might progress in their disease, commented Kate Papp of Brigham and Women’s Hospital in Boston. However, she noted that it is unlikely that a single modality or test will be the ultimate predictor of dementia. “Instead, gathering information from multiple sources—such as blood-based biomarkers, current cognition and function, and genetics—will be necessary,” she wrote. Blood biomarkers—which detect AD pathology festering in the brain with an accuracy of 90 percent or more—are also vying for regulatory approval (Aug 2022 conference news; Dec 2022 conference news; Dec 2023 news).

Years before brain atrophy due to Alzheimer’s disease becomes blatantly obvious on an MRI scan, neurons start to die off in specific regions of the brain. Researchers have long proposed that early in the disease process, subtle atrophy patterns would predict the impending onslaught of AD or other neurodegenerative diseases (for example, de Leon et al., 1993; Fox et al., 1996; Dec 2011 news). However, detecting these small changes and calculating their predictive value pose computational challenges. These happen to be the bread and butter of AI, which gleans this information through training with terabytes of real world data, and is increasingly being used in all aspects of dementia research (Jul 2023 news).

For its part, BrainSee was trained on brain scans and cognitive scores from thousands of patients, according to Darmiyan. The algorithm uses scores on the mini mental state examination (MMSE) and the Clinical Dementia Rating scale sum of boxes (CDR-SB), as inputs along with the standard whole-brain MRI. It then outputs an AD dementia probability score, ranging from 0 to 100. For example, a score of 70 indicates a 70 percent chance of dementia within five years.

Structural MRI is already a part of the clinical work-up for patients with subjective cognitive complaints. “BrainSee is available to physicians taking care of cognitively impaired adult patients, often in primary care settings,” Darmiyan’s Kaveh Vejdani wrote to Alzforum. “These include neurologists, geriatricians, psychiatrists, family physicians, and internal medicine physicians.” The physician uploads a patient’s MRI scan files, along with their scores on the MMSE and CDR-SB, onto BrainSee’s web portal, which renders an AD dementia probability score on the same day. The software comes with a tutorial and interpretation guide for physicians, as well as a guide for patients and caregivers.

BrainSee costs $1,500, but Vejdani told Alzforum that Darmiyan offers it for $300, pending Medicare coverage. Medicare already covers the cost of the scans themselves, which have an average price tag of $1,000 for those paying out of pocket. Vejdani foresees BrainSee as an initial screening test. Patients with a low probability of dementia can focus on other possible explanations for their cognitive issues, while those with a high probability of AD can seek further testing and treatment.—Jessica Shugart


  1. In a side-by-side comparison between BrainSee and hippocampal volume to predict progression to AD-dementia in 409 aMCI patients, hippocampal volume performed with 65.0 percent sensitivity, 78.2 percent specificity and 71.6 percent balanced accuracy; BrainSee performed with 84.3 percent sensitivity, 81.6 percent specificity and 83.0 percent balanced accuracy.

  2. Perhaps the BrainSee algorithm is detecting thinning of cortical gray matter as in Satizabal et al., 2023. That said, a blood draw is easier for most patients than an MRI, and plasma p-tau217 immunoassay detects AD with 90-95 percent accuracy (Ashton et al., 2024).


    . A novel neuroimaging signature for ADRD risk stratification in the community. Alzheimers Dement. 2023 Dec 26; PubMed.

    . Diagnostic Accuracy of a Plasma Phosphorylated Tau 217 Immunoassay for Alzheimer Disease Pathology. JAMA Neurol. 2024 Jan 22; PubMed.

  3. Sorry to be the skeptic, but 71.6 percent balanced accuracy (hippocampal volume) versus 84.3 percent (BrainSee) does indeed sound like hippocampal volume is a/the primary feature driving prediction (referencing Dr. Fox's comment). I have a hunch that hipp. vol. + MMSE + CDR-SB in a supervised model would get even closer to the full predictive power of BrainSee's model, without all the drawbacks of an unsupervised approach.

    Which is rather the chief concern. It has been well-documented that overtrained AI models fail to reproduce their advertised validation set accuracy (Sohn, 2023), so I would wonder how much of that additional accuracy is overtraining of the test cohort? How much of that is a confounding variable that isn't identifiable? Looking through the bioRxiv article and the abstracts, I don't see a description of the type of machine-learning model used. Have they considered an attention-based ML approach so they can extract soft feature weights? Assumedly, these would be voxels, but mapping them back to their location one might see the hippocampus light up.

    Per Dr. Papp's comment, a heterogeneous disease like AD is not going to be effectively addressed by chasing high classification accuracy in one or two modalities, particularly if the treatment regimen itself is going to be heterogeneous and tailored to the individuals' specific disease manifestation.


    . The reproducibility issues that haunt health-care AI. Nature. 2023 Jan;613(7943):402-403. PubMed.

  4. One of the most pressing concerns for patients and their families is, "What will happen to me?" And right now we as clinicians are not very good at predicting how slowly or quickly individuals will progress in their disease. Patients and families want to know what to expect—how long before I will need extra help? How long before I can't live independently anymore?

    The approval of this technology is a promising development toward providing some of these answers. However, it is unlikely that one modality (e.g., imaging, or imaging+cognition) will be the best predictor. Instead, gathering information from multiple sources, such as blood-based biomarkers, current cognition and function, and genetics will be necessary.

    These tools are also only as good as the data they are trained on, and we know there is considerable heterogeneity in AD and related dementias, so it will be important for clinicians to know the size and nature of the samples that were used to develop these algorithms.

Make a Comment

To make a comment you must login or register.


News Citations

  1. Blood Tests: Charting the Path to Primary Care
  2. Plasma P-tau217 Picks Up Plaques, Tangles, Future Decline
  3. Two New p-Tau217 Blood Tests Join a Crowded Field
  4. Is Structural MRI Closing In on Preclinical AD Biomarker?
  5. Artificial Intelligence Is Everywhere You Look in Dementia Research

Paper Citations

  1. . A novel technology for objective, accurate and non-invasive early diagnosis and monitoring of Alzheimer’s disease in clinics and clinical trials. 2019 Oct 02 10.1101/790469 (version 1) bioRxiv.
  2. Abstract: Symposia, Conferences, Oral communications: 14th Clinical Trials on Alzheimer's Disease (CTAD) November 9-12, 2021. J Prev Alzheimers Dis. 2021;8(S1):S2-S72. PubMed.
  3. . The radiologic prediction of Alzheimer disease: the atrophic hippocampal formation. AJNR Am J Neuroradiol. 1993;14(4):897-906. PubMed.
  4. . Presymptomatic hippocampal atrophy in Alzheimer's disease. A longitudinal MRI study. Brain. 1996 Dec;119 ( Pt 6):2001-7. PubMed.

External Citations

  1. CTAD abstracts, 2020

Further Reading


  1. . A Transfer Learning Approach for Early Diagnosis of Alzheimer's Disease on MRI Images. Neuroscience. 2021 Apr 15;460:43-52. Epub 2021 Jan 17 PubMed.