How Can AI Techniques Improve How We Diagnose Dementia?

Developing treatments for dementias, such as Alzheimer’s disease, remains one of the most forbidding challenges facing neurology.

But the task of identifying the signs of dementia at an early enough stage for such treatments to have an effect is also an obstacle for the field.

Traditionally, paper-based assessments, such as the Mini-Mental State Exam (MMSE) and the Montreal Cognitive Assessment (MoCA), are used by clinicians to diagnose patients. But London-based company Cognetivity says that its AI-based platform could improve dementia diagnostics. In the wake of a publication detailing Cognetivity’s platform, we spoke with the company’s co-founder and chief science officer, Dr. Seyed Khaligh-Razavi, to find out more.

Ruairi Mackenzie (RM): Your recent study examined the ability of the Integrated Cognitive Assessment (ICA) to assess cognitive impairment in healthy adults and those with mild cognitive impairment (MCI) and mild Alzheimer’s disease. What were the key findings?

Seyed Khaligh-Razavi (SKR): We demonstrated that the ICA’s results (powered by its explainable AI engine) could generalize well from one cohort of individuals to a different cohort with substantial cultural and language differences, without the need to collect new normative data. This makes the test particularly suitable for cognitive screening and monitoring in large and diverse populations. 

RM: Data in the study suggests that the ICA has a comparable sensitivity and lower specificity than the MoCA. What is lacking from currently available diagnostic tests, such as the MoCA, that the ICA could augment?

SKR: There are few key dimensions to compare here: 

a) In terms of time to administer the test: ICA is a self-administered 5-min test, with  automatic result generation). MoCA takes about 15-20 min to administer. It requires the precious time of a trained clinician to administer the test. Then it takes a few more minutes to score the test. And unfortunately, you can face discrepancies in scoring from one clinician to another. 

b) In terms of biases and test accuracy: the conventional pen and paper tests are biased by differences in education, language and culture, as shown in this study. We demonstrate here how the ICA is designed to be free from these biases and even show that it can generalize well to a substantially different population without re-training and still achieve a high accuracy comparable to that of the current standard of care. 

c) Finally, as we show in the paper, thanks to its AI engine, ICA’s performance is further improved over time as more clinical data is added to its training set.

RM: What role does AI play in the function of the ICA?

SKR: In sum, the AI engine has helped the test to be free from typical biases of the conventional tests and generalize across populations and further provides us with a unique opportunity to learn from new data and never stop improving its accuracy. 

RM: What role do you see AI having in the future of dementia diagnostics?

SKR: The ICA in its current form is an aid for diagnosis. Along with our recently announced patent, and in light of disease-modifying therapies (DMT) becoming available for Alzheimer’s, it is foreseeable in the future to have AI algorithms on your phone that predict your personalized risk of developing AD and will be able to stratify patients to help healthcare providers to navigate which DMT is likely to be effective for an individual (e.g. a DMT targeting amyloid-beta , or another one targeting tau accumulation) in the earliest stages of the disease – when a treatment will be effective. 

RM: It is increasingly being recognized that for many putative therapies for AD and other dementias to have maximum effect, they must be administered potentially decades prior to the onset of cognitive symptoms. Does this challenge the therapeutic value of cognitive diagnostics for these conditions in the long run, in comparison to molecular approaches?

SKR: Molecular approaches (such as PET scans) are extremely expensive and not scalable. Even CSF and blood tests are extremely difficult, if not impossible, to use for population-wide screening. Molecular approaches are complementary to a digital biomarker, like ICA. ICA is easily scalable and particularly suited for population-wide screening, then people can be guided to go through a complementary diagnostic pathway, based on the outcome of their ICA results. In simple terms, ICA can be used to screen a large population, find people at high risk and refer them for further testing such as CSF, PET etc. 

Dr. Seyed Khaligh-Razavi was speaking to Ruairi J Mackenzie, Senior Science Writer for Technology Networks.