Oita University and Eisai have developed a machine learning model to predict amyloid beta (Aβ) accumulation in the brain using routine clinical data. The model combines background information such as age, gender, smoking history, and medical history with general blood test results and Mini-Mental State Examination (MMSE) items. The details of the model were published in the online edition of Alzheimer’s Research & Therapy on 21 January 2025.
The model is expected to enable primary care physicians to predict brain Aβ accumulation, a key pathological factor in Alzheimer’s disease (AD), during routine medical examinations. Current methods for detecting Aβ accumulation, such as positron emission tomography (amyloid PET) and cerebrospinal fluid (CSF) testing, present challenges due to high costs and invasiveness. Recent research has focused on blood biomarkers as a more accessible screening method. However, there is limited research on models predicting Aβ accumulation using only routine clinical data.
This study is the first to develop a machine learning model predicting amyloid PET positivity using 34 routinely collected clinical data points. These include background data (age, gender, smoking history, and medical history), general blood test data (kidney function, liver function, thyroid function), and MMSE scores. The model’s performance was evaluated using the Area Under the Curve (AUC) metric. The AUC was 0.70 for the model combining background and blood test data and 0.73 for the model that also incorporated MMSE data.
Anti-Aβ antibody treatments have shown potential benefits when initiated at earlier stages of AD, making early detection of Aβ accumulation significant. This machine learning model provides a prediction method based on clinical data collected during routine medical care. By helping primary care physicians determine the necessity of amyloid PET and CSF tests, the model is expected to facilitate early diagnosis and treatment initiation while reducing economic and physical burdens on patients.
The study used outpatient data from Oita University Hospital (September 2012–November 2017) and data from the prospective cohort study (USUKI STUDY) conducted in Usuki City, Oita Prefecture (October 2015–November 2017). The model was developed using three machine learning techniques: Support Vector Machine, Elastic Net, and L2 regularisation logistic regression. It was trained on 262 individuals aged 65 and older with mild cognitive impairment or normal cognitive function.
The L2 regularisation logistic regression model incorporating participant backgrounds and MMSE scores, as well as the model combining participant backgrounds and general blood test results, both achieved an AUC of 0.70. The model incorporating all three elements—background data, blood test results, and MMSE scores—achieved an AUC of 0.73. Key factors contributing to Aβ accumulation prediction included delayed recall and place orientation among MMSE items, age, thyroid-stimulating hormone levels, and mean corpuscular volume.
The study was conducted by researchers from Oita University and Eisai Co., Ltd., including Noriyuki Kimura, Kotaro Sasaki, Teruaki Masuda, Takuaki Ataka, Mariko Matsumoto, Mika Kitamura, Yosuke Nakamura, and Etsuro Matsubara. The research findings contribute to the ongoing efforts to improve early screening for AD and optimise the use of advanced diagnostic tests.