Macnica uses cutting-edge AI technology to predict memory ability "related to early dementia diagnosis"; paper on deep learning approach using EEG analysis accepted at international conference EMBC2024

Macnica (headquarters: Yokohama, Kanagawa Prefecture; Representative Director and President: Kazumasa Hara; hereinafter referred to as Macnica) announced today that its paper, "Label Distribution Learning for Memory Decline: A Deep Learning Approach Using EEG Analysis," was accepted for presentation at IEEE Engineering in Medicine and Biology Society (EMBC) 2024, a leading international conference in the field of biomedical engineering, held in Orlando, Florida on July 15.

Research Background
In modern society, cognitive decline, such as Alzheimer's disease and mild cognitive impairment (MCI), is becoming a serious social issue as the population ages. Early detection and treatment of mild MCI has been shown to improve dementia symptoms and delay onset, making it extremely important to detect dementia at an early stage in order to prevent an increase in the number of dementia patients. Currently, there are existing methods for dementia testing, such as blood tests, but Macnica 's Brain AI Innovation Lab (BRAIL) ^{* 1} is focusing on the cognitive functions of the brain and verifying new evaluation methods that complement existing evaluation methods. The results of the new evaluation method for memory domain were presented as a poster paper.

Summary of the paper
<Verification details>
Focusing on memory, one of the main causes of cognitive decline, we have constructed a model to predict "memory ability" from EEG measurement results. The accuracy of the model to predict an individual's cognitive ability has been achieved up to 90%.
We conducted an analysis using the following two items of data from the open data of the Max Planck Institute Leipzig's Mind-Brain-Body (hereinafter LEMON) ^*2.

Overview: The instruction pipeline of the proposed model.

[Data used]
① CVLT (California Verbal Learning Task): A method to evaluate short-term and long-term memory by memorizing 16 words and taking a test five times to see how many of them you can remember.
② EEG data: Resting state EEG measurement results

<Analysis method>
A novel framework for predicting memory performance using severity index and expectancy regression approach.
This method utilizes techniques called "Dynamic Graph Convolutional Neural Network (DGCNN) ^*3" and "Label Distribution Learning (LDL)."

Feature points were extracted from LEMON's EEG data using DGCNN, and the CVLT results were standardized using LDL. In addition, a Gaussian function was used to generate a label distribution, and Kullback-Leibler (KL) divergence was applied as a loss function ^*4 to optimize the model. It has been shown that the use of LDL makes it possible to build a model that can interpret detailed factors of dementia, unlike conventional methods.

The paper can be found here.
https://www.macnica.co.jp/business/ai/manufacturers/files/embc24-POC3poster-v4.pdf
https://www.macnica.co.jp/business/ai/manufacturers/files/Poc3_1-Page_abstract_final.pdf

This paper was written by BRAIL (Wei Chen, Aldrin Domer, Kapeleshh KS) and Professor Hong Ji of Xi'an Polytechnic University.

■ Future outlook
At this year's EMBC, there were many research presentations related to MCI dementia. It was reaffirmed that this is recognized as a social issue worldwide, and it is expected that MCI-related research and services will continue to develop in the future.
In BRAIL, the analysis in this paper was limited to memory domains, but LEMON's open data includes data on multiple dementia-related measures (such as reaction speed and speed of completing tasks). In the future, we plan to add items other than memory ability to verify methods of evaluating complex cognitive abilities.
Macnica will continue to develop cutting-edge technologies that contribute to improving people's health and welfare.

It is also possible for companies, medical institutions, and research institutes to conduct research using actual task test data, so for detailed information or inquiries, please see the link below.

*1: BRAIL offers an approach to solving problems by extracting essential data from the brain that people are not aware of through "Brain-AI," a combination of brain and AI. We collaborate with various product partners and academic organizations such as universities and research institutes, both domestically and internationally, and through such partnerships, we share not only ideas and methods for utilizing brain science, but also knowledge based on cutting-edge research with society. We believe that information generated from brain activity will provide new value to society. BRAIL will continue to focus on the new value that brain science can bring, and in addition to accelerating the digitalization of society, we will propose and promote better ways of interacting with people and the digital society based on research and development with our partners.
https://www.macnica.co.jp/aboutus/technology/ai/brail/

*2: Mind-Brain-Body (LEMON) is an open dataset for studying the interactions between the mind, body, and emotions, which involved 153 young people and 74 older people, with MRIs, EEG, cardiovascular measurements, and mental evaluations.

*3: Disordered Graph Convolutional Neural Network (DGCNN) is an AI model that uses a graph neural network to capture the dynamic features of each frequency in the brain.

*4: A loss function is a formula that calculates the difference between the result predicted by a machine learning model and the correct answer.

*Company names and product names mentioned in the text are trademarks or registered trademarks of Macnica and the respective companies.
*Information published in news releases (including product prices, specifications, etc.) is current as of the date of announcement. Please note that it may be subject to change without prior notice.