Wearable Amyloid Monitor

FYI -

Pat & Dennis Bender Early Dementia Diagnosis & Prognosis Fund

Dennis & Pat 07-84 $C:\Users\jdenb\AppData\Local\Microsoft\Windows\INetCacheContent.Word\Dennis.jpg$ $C:\Users\jdenb\AppData\Local\Microsoft\Windows\INetCacheContent.Word\DSCN2587.jpg$

J. Dennis Bender

Office, Home & Cell Phone: 859-391-5226

5726 La Jolla Blvd. – Suite 311

La Jolla, CA 92037-7345

Office - 100 Riverside Pl. - Suite 303

Covington, KY 41011-5711

We support the development of improved diagnostic methods for the early detection and diagnosis of MCI, Alzheimer’s, vascular and other dementias, their likely prognosis, and best treatment options. We focus on the development of Bayesian-based medical-decision-support systems, comparative-effectiveness research, and the better utilization of these for the above. (After incorporating in KY as a 501(c)(3) in 2002, we dissolved that entity for a simplified form of two entirely self-financed, private philanthropies utilizing a Vanguard Charitable Trust for making $100K annual-research-grants for early-dementia-detection and its correct differential-diagnosis and likely-prognosis. They will continue on, after I am long gone, either mentally or physically. Prof. Randall Bateman is the first of our fund’s research advisors, KMK Law is our legal advisor and David J. Bender is my Estate Rep. (See: https://www.alz.org/alzheimers-dementia/research_progress/earlier-diagnosis)

This email address is being protected from spambots. You need JavaScript enabled to view it.

www.JDBender.com – EMS/eVTOL & Educational Experimental Aviation Fund (Vanguard Charitable Trust)

www.JDBender.org – Dementia Diagnosis Fund (Vanguard Charitable Trust)

January 7, 2024

A person standing in a field with a camera and a tripod

Description automatically generated A graph of a positive rate

Description automatically generated with medium confidence

“This research is the world’s first attempt to create a machine-learning-model to predict-amyloid-positivity in subjects by utilizing biological-data collected with a wristband-sensor, including daily-activity, sleep, speech, and heart-rate, and lifestyle-data collected through medical-consultation. . . Development of an inexpensive and easy-to-use screening-method has been sought after to identify those who need amyloid PET or CSF testing.” [That I helped to fund the original development of and now, even though very willing and able to pay for them, I am constantly refused them, unless I participate in a trial, for which the requirements always exclude me! A pox on all their houses!]

I think that this approach has a lot of potential! It is exactly the sort of thing I am most interested in providing for another $100,000 annual research grant in 2024. I only wish that I’d had the excellent set of tools available now days, versus having to create them myself from scratch using just P&G’s Matran, SAS and PL-1 on their old IBM-7080 and later IBM-360, and having to come in at 3:00AM after OSB was run to do my technical-computing on their P&G mainframe, and later having to truck my card-decks over to the Kroger building in Cincinnati to transmit my programs and data up to the then ‘super-computer’ at Control Data’s facility in Minneapolis to do the computational work. Later I finally installed the RUSH time-sharing at P&G to do more of this work inhouse in my office at Corporate headquarters in downtown Cincy. (See my History of Computing at P&G that I presented at their annual international management meeting at the Cincy Zoo that year.)

Congrats to This email address is being protected from spambots. You need JavaScript enabled to view it. & This email address is being protected from spambots. You need JavaScript enabled to view it. for a great piece of research to start off this promising New Year! It’s a new dawn! Be glad you don’t live back in the 1960s & ‘70s computing environment that I had to contend with at P&G! (I had to program Tukey’s then brand-new Fast Fourier Transform algorithm using P&G’s own Matran-assembler-language when that algorithm was first published. Boy did it speed-up that calculation!)

I note that my old-favorite logistic-regression did not all that badly against these much-newer, more-sophisticated, computational approaches easily-available today without one’s having to program much of anything oneself. What a wonderful new day and new world for the younger generation to inherit! All that, along with my now $15M+ research-trust-fund that I’ve set-up to help pay for its further development, after I am long-gone, one way or the other!

My next-step is to see if there is some way to rework this from a Bayesian-based, decision-theory sort of approach, explicitly incorporating both continuous and polychotomous data and missing-data; as was previously done in an earlier study that I reviewed, especially liked and recommended, some time ago. Looking forward to contacting these authors about funding a potential next-step in that direction, if they have any interest in doing so.

Development of Prediction Model for Brain Amyloid-Beta Accumulation Using Wristband-Sensor

For Print（132KB）- December 26, 2023

Oita University and Eisai Co., Ltd. (Eisai) hereby announce the development of the world's first machine-learning-model to predict-amyloid-beta*¹-(Aβ)-accumulation in the brain using a wristband-sensor. This model is expected to enable screening for brain-Aβ-accumulation, which is an important pathological-factor of Alzheimer's-disease*² (AD), simply by collecting biological and lifestyle-data from daily-life.

[The details of this model were published in the online edition of the peer-reviewed medical journal Alzheimer's Research & Therapy on December 12, 2023.]

In AD, which is said to account for over 60% of the causes of dementia, Aβ begins to accumulate in the brain about 20-years before the onset-of-the-disease. This has prompted the development of new therapeutic-drugs-targeting-Aβ, leading to the approval of a humanized-anti-soluble-aggregated-Aβ-monoclonal-antibody in Japan. The key to maximizing-treatment-effects of this medicine is detecting-Aβ-accumulation in the brain of patients with mild-cognitive-impairment before-the-onset-of-symptoms.

Currently, although brain-Aβ-accumulation can be detected by positron-emission-tomography*³ (amyloid-PET) and cerebrospinal-fluid-testing*⁴ (CSF-testing), the number of medical institutes able to perform those tests is limited, and the high-cost and invasiveness of these tests are considered issues. Therefore, development of an inexpensive and easy-to-use screening-method has been sought after to identify those who need amyloid PET or CSF testing.

Although lifestyle-factors, including lack-of-exercise, social-isolation, and sleep-disorders, as well as diseases, including hypertension, diabetes, and cardiovascular-disease are known risk-factors-for-AD, thus far, studies applying machine-learning-models to predict-brain-Aβ-accumulation have used only cognitive-function-tests, blood-tests, and brain-imaging-tests. In contrast, this is the first machine-learning-study to focus on biological-data and lifestyle-data.

This study integrated biological-data collected by wristband-sensors, such as physical-activity, sleep, and heart-rate, lifestyle-data obtained from medical-consultations, such as number of household members, employment-status, frequency of going outdoors, means of transportation, number of days participating in community activity, and subjects’ background, such as age, education-history, history-of-alcohol-intake, and medical-history (hypertension, stroke, diabetes, heart-disease, thyroid-disease) to create a machine-learning-model to predict individuals who are likely to test-positive by brain-amyloid-PET, and evaluated the model’s-performance. [My first research-grant for $50,000 went to the U. of Pittsburgh in 2003 to help fund the original development of PiB for just this purpose.]

The result of this study showed that the Area-Under-the-Curve (AUC), an evaluation-index of the prediction-model consisting of biological-data, lifestyle-data and subjects’-background, was 0.79, leading to the decision that the model has sufficient capability for screening. [Sufficient, but not great, so room for improvement by adding the other potential biomarker factors, such as blood-plasma, that I have been investigating for the past two-decades now.]

This machine-learning-model is able to predict-brain-Aβ-accumulation using readily-available non-invasive-variables. As a result, this model appears widely-applicable as a pre-screening for people living in areas with little access to amyloid-PET and CSF-testing, and to reduce the financial and physical burden to patients, as well as the costs-of-clinical-studies.

Glossary of Terms

*¹ Amyloid-beta: A protein viewed as a cause of Alzheimer's-disease, which accumulates in the brain for about 20 years prior to the onset of the disease and forms senile plaques

*² Alzheimer's-disease: the most common cause of dementia, and its pathological characteristics include senile-plaques, neurofibrillary-tangles, and neuronal-cell-death

*³ Amyloid-PET: a brain-imaging-test visualizing Aβ-accumulation in the brain [that I helped to fund the early development of at the U. of Pittsburgh when everyone scoffed at it as not being worth working on.]

*⁴ Cerebrospinal-fluid-testing: A test analyzing cerebrospinal-fluid for Aβ42, phosphorylated-tau, and total-tau as biomarkers of Alzheimer's-disease

Background and Outline of Research As Japan has become a super-aging-society with the rise in the number of dementia-patients over-the-age-of-65, the development of new therapeutic agents for AD, the most common cause of dementia, is an urgent issue. Lifestyle-habits such as lack-of-exercise, social-isolation, and sleep-disorders, as well as diseases such as hypertension, diabetes, and cardiovascular-disease, are known-risk-factors-for-AD. [Need to also mention the very-important, inexpensive and easily-measured APOEe4 genetic-risk-factor.]

The development of drugs targeting-Aβ has seen progress in recent years, and this year, an anti-Aβ-protofibril-antibody was approved in Japan. To maximize the effects of this medicine, it is essential to identify individuals with pre-symptomatic mild-cognitive-impairment (MCI) who are likely to have elevated-brain-amyloid-accumulation. To date, although reports on machine-learning-models for predicting-brain-Aβ-accumulation by cognitive-function-tests, blood-tests, and brain-imaging-tests have been available, no research has focused on biological-data or lifestyle-data. This research is the world’s first attempt to create a machine-learning-model to predict-amyloid-positivity in subjects by utilizing biological-data collected with a wristband-sensor, including daily-activity, sleep, speech, and heart-rate, and lifestyle-data collected through medical-consultation.

Results and Significance of Research, Future Development This research utilized data from a prospective cohort study on the elderly without dementia aged 65 and older living in Usuki City, Oita Prefecture, conducted between August 2015 and September 2019. 122 individuals (54 men, 68 women, median-age 76-years) with mild-cognitive-impairment or subjective-memory-impairment wore a wristband-sensor for approximately 7-days every-3-months. [I would have fit their then criteria, even now with Biden and I both being at age-81 and me now formally diagnosed with amnestic-MCI-of-mixed-etiology to qualify for that subgroup.]

The study also collected lifestyle-data through medical-consultation, and subjects underwent regular-amyloid-PET (once-a-year) examinations over the course of 3-years. The research evaluated a predictive-model created with 3 machine-learning technologies, support-vector-machine, Elastic-Net, and logistic-regression, to integrate biological-data collected by wristband-sensors, such as physical-activity, sleep, and heartbeat, and lifestyle-data obtained from medical-consultations, such as living-with-household-members, employment-status, frequency-of-going-outdoors, means-of-transportation, number-of-days-participating-in-community-activity, as well as the subjects’-background, such as age, education-history, history-of-alcohol-use, and medical-history (hypertension, stroke, diabetes, heart-disease, thyroid-disease).

For instance, while the AUC of a predictive-model created with only biological-data-alone collected by the wristband-sensors using Elastic-Net was 0.70, the AUC of a predictive-model created with additional lifestyle-data and patient-background was 0.79, exhibiting improved performance.

This research is the world’s first attempt to create a machine-learning-model to predict-brain-Aβ-accumulation by utilizing biological-data collected with a wristband-sensor, including daily-activity, sleep, speech, and heart-rate, and lifestyle-data collected through medical-consultation, as well as subjects’-background.

Furthermore, using state-of-the-art algorithms to identify multiple-factors which contribute to predicting-Aβ-accumulation, 22-common-factors were identified that were common across the 3-learning-machine-technologies. Specifically identified were physical-activity, sleep, heart-rate, amount-of-conversation, age, length-of-education, living-with-or-without-children, means-of-transportation, presence-of-accompanying-person-for-hospital-visits, communication-frequencies, and number-of-outings.

Academic Paper - Title: Predicting Positron-emission-tomography Brain Amyloid-Positivity Using Interpretable Machine-Learning-Models With Wearable-Sensor Data and Lifestyle-Factors

Authors: Noriyuki Kimura (Department of Neurology, Faculty of Medicine, Oita University) 1,2, Tomoki Aota (Eisai Co., Ltd.) 1, Yasuhiro Aso(Oita Prefectural Hospital), Kenichi Yabuuchi (Department of Neurology, Faculty of Medicine, Oita University), Kotaro Sasaki (Eisai Co., Ltd.), Teruaki Masuda (Department of Neurology, Faculty of Medicine, Oita University), Atsuko Eguchi(Department of Neurology, Faculty of Medicine, Oita University), Yoshitaka Maeda (Eisai Co., Ltd.), Ken Aoshima (Eisai Co., Ltd./University of Tsukuba) 2, Etsuro Matsubara (Department of Neurology, Faculty of Medicine, Oita University)

1. These authors contributed equally to the manuscript. 2. Corresponding author.

Publisher: Alzheimer's Research & Therapy

Please direct any interview requests or inquiries to the contact information provided below

For further information or any inquiries regarding this study

Noriyuki Kimura, Associate Professor, Department of Neurology, Faculty of Medicine, Oita University

TEL: +81-(0)97-586-5814, FAX: +81-(0)97-586-6502

Email：This email address is being protected from spambots. You need JavaScript enabled to view it.

Research - Open access - Published: 12 December 2023

Predicting Positron-emission-tomography Brain Amyloid-Positivity Using Interpretable Machine-Learning-Models with Wearable-Sensor Data and Lifestyle-Factors

Alzheimer's Research & Therapy Volume 15, Article number: 212 (2023) Cite this article

Abstract

Background Developing a screening method for identifying individuals at higher risk of elevated brain amyloid burden is important to reduce costs and burden to patients in clinical trials on Alzheimer’s disease or the clinical setting. We developed machine-learning-models using objectively measured lifestyle-factors to predict elevated brain amyloid burden on positron-emission-tomography.

Methods Our prospective cohort study of non-demented, community-dwelling older-adults aged ≥ 65-years was conducted from August 2015 to September 2019 in Usuki, Oita Prefecture, Japan. 122 individuals with mild-cognitive-impairment or subjective-memory-complaints (54 men and 68 women, median-age: 75.5-years) wore wearable-sensors and completed self-reported-questionnaires, cognitive-test, and positron-emission-tomography imaging at baseline.

Moreover, 99 individuals in the second-year and 61 individuals in the third-year were followed-up. In total, 282 eligible records with valid wearable-sensors, cognitive-test results, and amyloid-imaging and data on demographic-characteristics, living-environments, and health-behaviors were used in the machine-learning-models. Amyloid-positivity was defined as a standardized-uptake-value-ratio of ≥ 1.4. Models were constructed using kernel-support-vector-machine, Elastic-Net, and [our good old favorite] logistic-regression for predicting-amyloid-positivity. The mean-score among 10-times, 5-fold, cross-validation-repeats was utilized for evaluation. [Isn’t it great to now have all the inexpensive computing resources that I never had at P&G, Nielsen, Milward-Brown, MSA and Marketing Analytics, prior to my early retirement following my dear wife Pat’s untimely passing from breast-cancer at only age-58.]

Results In Elastic-Net, the mean-area-under-the-receiver-operating-characteristic-curve of the model using objectively-measured lifestyle-factors alone was 0.70, whereas that of the models using wearable-sensors in combination with demographic-characteristics and health and life environment questionnaires was 0.79. Moreover, 22-variables were common to all machine-learning-models.

Conclusion Our machine-learning-models are useful for predicting-elevated-brain-amyloid-burden using readily-available and noninvasive variables without the need to visit a hospital.

Keywords (Note this my today’s trial of my new approach to mostly only hyphenating MeSH and keywords in the following text, versus my practice over the past many years utilizing my own custom-search-code based on it for my own document search purposes across the now 9,000+ such annotated research files now housed in my two research-topic databases. Now trying to use that PubMed-MeSH approach combined with keywords instead of my old search-system.)

Trial Registration This prospective study was conducted in accordance with the Declaration of Helsinki and was approved by the local ethics committee of Oita University Hospital (UMIN000017442). A written informed consent was obtained from all participants. This research was performed based on the Strengthening the Reporting of Observational Studies in Epidemiology reporting guideline.

Background Dementia is a growing public health issue in an aging society with an increasing life expectancy that poses a serious social and economic impact in patients and caregivers. Alzheimer’s disease (AD) is a major cause of dementia in people aged over 65 years. Recently, an anti-amyloid β (Aβ) antibody has been approved by the Food and Drug Administration as a new disease-modifying therapy for AD. This agent reduced brain amyloid deposition and the rate of cognitive decline in the mild-cognitive-impairment (MCI) or mild dementia stage [1]. Therefore, more precise, and early identification of patients with AD is crucial to enhance the benefit from disease-modifying therapy.

The neuropathologic hallmarks of AD are the extracellular aggregation of Aβ plaque, the presence of neurofibrillary tangles, and neuronal degeneration [2, 3]. Aβ accumulation in the brain is a key pathological characteristic and the initial event in the pathological process of AD [4, 5]. Therefore, Aβ is the major target of disease-modifying therapy in clinical trials on AD [6, 7]. However, almost all agents were not effective due to delayed intervention or selection of inappropriate participants without amyloid pathology [7, 8]. ¹¹C-Pittsburgh Compound B (PiB) [that I helped to fund the early development of] positron-emission-tomography (PET) and measurement of cerebrospinal fluid (CSF) Aβ levels are established biomarkers of amyloid pathology, and they can predict the incidence of AD before the onset of dementia [2, 3].

Nevertheless, their use in clinical and research setting are limited due to their high cost or the need for invasive lumbar puncture. Moreover, the low prevalence of amyloid positive MCI in multicenter research (46.6%) or population-based study (22%) may cause higher cost and burden to the patients due to a large number of PiB-PET scan or CSF analysis in clinical trials on AD or clinical setting after the availability of disease modifying therapy [9, 10].

Therefore, it is challenging to develop a cost-effective and noninvasive method for detecting amyloid pathology prior to PiB-PET or CSF analysis. Although blood-based biomarkers are a promising solution for predicting individuals with amyloid pathology in the memory clinics [11,12,13,14], they are not suitable for population screening due to the need to visit a hospital and draw blood.

In this study, we developed and validated predictive-models using 3 machine-learning techniques by integrating easily available and noninvasive lifestyle variables to detect individuals with elevated brain amyloid deposition. Moreover, overall daily physical-activity and sleep patterns was objectively and continuously collected using wearable-sensors without recall-bias [15, 16].

These predictive-models can be useful for the population-wide screening or primary care setting to identify the patients eligible for PiB-PET or CSF analysis. Almost all previous studies have developed machine-learning-models for classifying individuals with healthy cognition, MCI, and AD, or for predicting the incidence-of-dementia [17, 18]. Meanwhile, a few studies have applied machine-learning techniques in neuropsychological tests, neuroimaging, and blood-based biomarker analysis combined with demographic characteristics and APOE-genotype for predicting brain amyloid-positivity [19,20,21,22,23,24,25,26].

To the best of our knowledge, few studies have focused on lifestyle-factors as predictive variables for identifying individuals at high risk for brain amyloid deposition [27]. There is growing evidence obtained from cohort studies showing that physical inactivity, social isolation, sleep disturbance, depression, and vascular risk factors are important predictors of late-life cognitive impairment [28,29,30]. Similarly, our prospective cohort study showed an association between objectively measured modifiable lifestyle-factors using wearable-sensor with cognitive function, brain amyloid deposition, or cortical glucose metabolism in community-dwelling older adults [31, 32].

These results lead us to the hypothesis that machine-learning-models using objectively-measured lifestyle-factors may predict elevated brain amyloid deposition measured via PiB-PET. Therefore, the current study aimed to predict brain-amyloid-positivity with 3 machine-learning-models using wearable-sensor data alone or in combination with demographic characteristics and living environment and health behaviors.

Our models are different from previous machine-learning-models as they used lifestyle-factors, which can be assessed using wearable-sensors and questionnaires without visiting a hospital. Further, this study developed and validated predictive-models using 3 machine-learning techniques by integrating easily available and noninvasive lifestyle variables to detect individuals with elevated brain amyloid deposition. These models can be implemented widely for the population-based prescreening to detect amyloid-positivity and can reduce unnecessary invasive lumbar-puncture and PET-scans, leading to successful clinical trials on AD and maximize the therapeutic effect of disease-modifying therapy.

Materials and Methods – Participants The USUKI study was prospective in nature and was conducted on community-dwelling older-adults without dementia from August 2015 to September 2019 in Usuki, Oita Prefecture, Japan. It was designed to explore the risk and protective lifestyle-factors of cognitive-decline in later-life.

Detailed designs and methods have been described elsewhere [31, 32]. The inclusion criteria were as follows: 1) age ≥ 65-years, 2) living in Usuki, 3) physically and psychologically healthy, 4) absence of dementia, and 5) independent function in activities-of-daily-living. All participants were required to wear a wristband-sensor for 7 consecutive days every 3 months (4 times per year) over 3 years of follow-up.

The valid sensing data were defined as at least 3 days in one period and at least two periods in a year. Table 1 shows the clinical and demographic characteristics of the participants. Between August 2015 and October 2017, 855 older individuals (317 men [37.1%] and 538 women [62.9%], median age [interquartile-range (IQR)]: 73 [69–78] years, median education duration [IQR]: 12 (11–12) years) satisfied the criteria, and they had valid sensing-data for analysis at baseline (data not shown).

Of 855 individuals, 122 (54 men [44.3%] and 68 women [55.7%], median age [IQR]: 75.50 [71.00–80.00] years, median education duration [IQR]: 12 [9–12] years) with MCI or subjective memory complaints [such as myself] (118 presented with MCI and 4 with subjective memory complaints) who underwent cognitive test and PiB-PET at baseline were recruited in the current study. The diagnosis of MCI was made based on the criteria of previous studies, which were as follows: 1) subjective and objective memory impairment, 2) Clinical Dementia Rating score of 0.5, and 3) absence of significant impairment in cognitive function or activities-of-daily-living.

Data on demographic characteristics and living environment and health behaviors were collected by trained medical staff using self-reported questionnaires (clinical history and medication). Moreover, the cognitive-test and PiB-PET were conducted on 99 individuals in the second year and 61 individuals in the third year (Table 1). In total, 282 eligible records with valid wearable-sensor, cognitive-test, and PiB-PET-findings and data on demographic characteristics and living environment and health behavior were used in the machine-learning-models.

Table 1 Clinical and demographic characteristics of the participants

Full-size table

Demographic Characteristics of the Participants Data on demographic characteristics, such as age, sex, education duration, and body-mass-index (BMI), were collected by trained medical staff annually. Moreover, a history of chronic disease, such as hypertension, diabetes mellitus, hyperlipidemia, stroke, heart disease, liver dysfunction, renal dysfunction, thyroid disease, and malignant tumor, was assessed based on clinical history and medications used.

Living environment and Health Behaviors Different living environment and health behaviors (Table 2) were collected using self-report questionnaires annually. We focused on family structure, living conditions (with a relative), transportation, engagement in paid work, hobby, exercise habits, cognitive activity, and social relationship. Dichotomous-variables such as drug and food allergy, pet ownership, gardening engagement, and cohabitants were used. Categorical-variables such as ever smoking or drinking (none, sometimes, everyday), history of chronic diseases (none, previous treatment, and current treatment), walking difficulty (none, walking with pain, need to use a cane), transportation mode (bicycle, driving one’s self in a private car or motorcycle, train or bus, and riding with family or friends or taxis), accompanying person (a person who can accompany the participants to outpatients visit: relative, friend, and others), caring about appearance (not at all, not often, very often, and most of the time), and denture (use of denture: none, partial denture, and full denture) were used.

The 4- or 5-point frequency-scale was used for the number of outing (none, 1–2 days a week, 3–4 days a week, and ≥ 5 days a week), reading a newspaper (none, 1–2 days a month, 1–2 days a week, 3–4 days a week, and ≥ 5 days a week), time spent on watching TV (none, < 3 h per day, > 3 h per day, and > 6 h per day), and lesson or class frequency (the frequency of taking lesson or class: none, 1–2 days a month, 1–2 days a week, 3–4 day a week, ≥ 5 days a week), communication frequency (the frequency of communication with friends or relatives: none, 1–2 days a month, 1–2 days a week, 3–4 a week, and ≥ 5 days a week), and primary or secondary hobby (none, 1–2 days a month, 1–2 days a week, 3–4 day a week, and ≥ 5 days a week). The number of family living together represents household-size, and the number of days working in a week (the average number of days spent on paid work), exercise frequency, and the number of days participating in community activity were treated as continuous variables.

Table 2 Explanatory variables with summary statistics in this study

Full-size table

Wearable-Sensor Data All participants wore a wristband-sensor (Silmee™ W20, TDK Corporation, Tokyo, Japan) continuously except when bathing. [I can also wear my Apple Ultra-2 in the shower and it measures all of these same parameters used in this research and more, the reason for my purchasing it before the patent-infringement suit.] Physical-activity, sleep, conversation time, and heart-rate were calculated by summing up the sensor data captured on each day and by averaging the entire measurement period annually. Data that indicated removing the wristband-sensor according to heart-rate were excluded. Physical-activity data were detected using a three-axis-accelerometer that measured acceleration in three perpendicular axes.

The steps and intensity-of-activity as metabolic-equivalents (METs) were captured. Physical-activity-intensity was divided into three categories, which were as follows: sedentary-behavior (≤ 1.5 METs), light-physical-activity (LPA) (1.6–2.9 METs), and moderate-to-vigorous-physical-activity (MVPA) (≥ 3.0 METs) [33]. The period of sedentary-behavior, LPA, and MVPA was evaluated during awaking.

Sleep–wake variables such as total-sleep-time (TST), sleep-efficiency, time-awake-after-sleep-onset (WASO), and awakening time count, were assessed from 18:00 to 5:59 the following day. Sleep-onset was defined as the first 20-min-block of resting-state-without-movement. Nocturnal-waking and waking during nap were defined as 5–90 min of continuous movement during a continuous sleep period. Sleep efficiency was calculated as the rate of TST versus the time spent in bed. WASO was defined as the total number of minutes awake after sleep onset during the night. Daytime napping was defined as the resting period without movement on the wearable-sensor from 6:00 in the morning to 17:59 in the evening. Nap efficiency was calculated as rate of naptime versus the time spent resting during daytime. Notably, WASO was not used as variable in machine-learning in this study.

Heart-rate was calculated by obtaining the average-pulses-per-minute in each day. Moreover, our wearable-sensor could detect sound pressure levels for utterances that originated within a 2-m radius from the device. The sound-pressure-level ranged from 55-to-75-dBA at this distance. Sound data were continuously captured via a microphone on the wearable-sensor and were analyzed to evaluate conversation-time. The microphone on the sensor could not detect the content of conversations. Active-Scale was calculated by counting the number of hours spent on walking at least 250-steps.

Cognitive Function Cognitive-assessments were performed using the Japanese version of the Montreal Cognitive Assessment (MoCA-J), with a score of 0–30. [Much better choice than the old and useless MMSE!] MoCA-J comprises 7 subscales, which are as follows: 1) visuospatial/executive function (alternate trail making, cube copying, and clock-drawing task, 5 points), 2) naming (three animal figures, 3 points), 3) memory (repetition only, no point), 4) attention (forward and backward digit span, target detection using tapping, and serial 7 subtraction, 6 points), 5) language (repetition, and verbal fluency, 3 points), 6) abstraction (2 points), 7) memory (enumeration of 5 nouns after approximately 5 min, 5 points), and 8) orientation (time and place, 6 points) [34]. If the education-duration was ≤ 12-years, we added one point to the total score of MoCA-J according to previous studies.

Apolipoprotein-E-Phenotype The ε4-allele of the apolipoprotein-E-gene is a genetic-risk-factor for late-onset-AD and is associated with brain-amyloid-deposition [35]. The apolipoprotein-E-gene could not be conducted. Hence, the apolipoprotein-E (APOE) phenotype was examined. The APOE-phenotype was identified using the Human Apolipoprotein E4/Pan-APOE ELISA kit (MBL Co., Ltd., Woburn, the USA), which measures the amount of APOE4 or total APOE specifically with high-sensitivity using affinity-purified polyclonal-antibody-against-APOE and monoclonal-antibody against APOE4 using sandwich-enzyme-linked-immunosorbent-assay. Moreover, it assesses differences among the homozygotes (i.e., ε4/ε4) and heterozygotes (ε2/ε4, ε3/ε4) [such as myself] of the APOE4 phenotypes and non-APOE4-zygotes (ε2/ε2, ε3/ε3, and ε2/ε3) based on the APOE level-to-APOE4 level-ratio. The homozygotes or heterozygotes of the ApoE4-phenotypes, and non-ApoE4-zygotes were defined based on a cutoff of 0.3 [36]. Notably, the ApoE4-phenotype was not used as variable in machine-learning in this study.

Positron-Emission-Tomography Scans Siemens Biograph mCT (Siemens) in the three-dimensional scanning mode was used in static ¹¹C-PiB-PET [that I helped to fund the development of] studies. The production of PET-tracers was performed based on good manufacturing standard at the PET Center of Oita University Hospital. Further, ¹¹C -PiB (mean: 547 MBq [SD: 60]) was injected intravenously as a rapid bolus with a saline flush, and radioactivity concentrations were measured from 50-to-70-min after injection. Radiation in pre- and post-dose samples was measured to define the exact injected dose using a radiation detector.

All imaging data were reconstructed using the following parameters: thick slice: 3.0-mm, matrix: 256 × 256, and magnification: 3.0 × . The pixel-size of the reconstructed images was 1.06 mm. Spatial normalization of PiB-scans was performed with a customized-PET-template at the Montreal Neurological Institute reference space using the Statistical Parametric Mapping version 8 (Wellcome Trust Centre for Neuroimaging). The region-of-interest (ROI), including the frontal-lobes, temporoparietal-lobes, posterior-cingulate-gyrus, and cerebellum, was determined using the MarsBaR (MRC Cognition and Brain Sciences Unit) ROI toolbox for Statistical Parametric Mapping, as described in a previous study [32].

These ROIs included areas with known amyloid deposition in patients with AD [37]. The average-ROI-values was obtained across both-hemispheres. The standardized-uptake-value-ratio (SUVR) was calculated from the voxel-number-weighted-average of the median-uptake in the frontal, temporoparietal, and posterior-cingulate ROIs relative to the ROI in the cerebellum [38]. The single-mean-value of all regions was combined to represent the global SUVR for PiB-PET. Amyloid-positivity was defined as a global PiB SUVR of ≥ 1.4. Meanwhile, amyloid-negativity was defined as a global PiB SUVR of < 1.4 for machine-learning-models. [I would have preferred mixed-modeling utilizing both continuous and dichotomous type variables instead of their procedure, but it is OK for this purpose and likely not feasible using this approach. I might be willing to fund the extension of this research in that direction, depending upon the estimated cost of doing so being less than US$100,000.]

Statistical Analysis Continuous and categorical variables were summarized as medians with IQR and ratios, respectively in Tables 1 and 2. To assess the distribution balance of explanatory-variables between the amyloid-positive and negative groups, we used Welch’s t-test for continuous-variables and Wilcoxon-rank-sum-test for categorical-variables. All statistical analyses were conducted using Python version 3.7.7. [Too bad Python appeared long after my time of using SAS and PL-1.]

Model Construction and Evaluation Models were constructed based on machine-learning-algorithms for predicting brain-amyloid-deposition. Three-steps were conducted in model-construction, i.e., Step-1: pre-selecting the machine-learning-algorithms, Step-2: feature-selection, and Step-3: training and evaluating the models. Step 1: To select the proper machine-learning algorithms, DataRobot 7.1.3 (DataRobot, the USA) was employed to explore different types of machine-learning-algorithms. We found that 3 algorithms (i.e., kernel Support Vector Machine [SVM], Elastic-Net, and logistic-regression) had a higher-accuracy after investigating more than 2,500 machine-learning-algorithms. SVM is supervised learning algorithms which is less-affected by outlier, kernel-SVM can be applied to nonlinear-classification by incorporating kernel-functions in the algorithm.

In this study, we used the radial-basis-function-kernel (RBF-kernel) as the kernel-function in the SVM. Elastic-Net is linear-regression-algorithm combined L1 and L2 regularization-methods. Logistic-regression is a classification-algorithm. In this study we set L2-penalty in logistic-regression as regularizer.

Step 2: To establish a predictive-model for amyloid-positivity using accumulated data collected using wearable-sensor and questionnaires, the selection of proper-number-of-variables is the key-point to prevent overfitting in the machine-learning-algorithm. We pre-selected 54-variables as the predictors from the cognitive test score variables and 111 variables including demographic characteristics (35), wearable-sensor (24), and living-environment and health-behaviors (52) before training the models.

In the pre-selection procedure, as a first-step, a comprehensive correlation calculation was conducted among 111 variables to avoid including statistically-similar variables in the models. One-of-two-variables whose correlation-coefficient equals 1.0 were excluded from the candidate predictive-variables.

Second, variables that were related to outcome in a clinical-point-of-view were selected from the candidate-variables that remained after the first-step. Table 2 shows summary-statistics of all predictive-variables in the machine-learning-algorithm. The variables with missing-values were imputed using only training-data before training. Thereafter, we explored different types of machine-learning-models using 52 variables that excluded MoCA-J-score and Active-Scale data from 54 variables for predicting-PET-amyloid-positivity (Model 0–3).

The reason why MoCA-J-score was excluded from the variables set was to validate the performance of models using variables that do not require a visit to the hospital. And the reason for the exclusion of Active-Scale was to compare the performances of models using Active-Scale with ones not using it and examine whether we could use Active-Scale as a substitute for other exercising features, which is easy to calculated from questionnaires and hence effective for future social implementation.

Step 3: For training and validation of the machine-learning-models, hyper-parameter-tuning was conducted using nested-cross-validation, the hyper-parameters in algorithms were chosen from some candidate values using grid-search method within the training dataset, and the exact grid of parameters in each algorithm were, gamma: [0.0001, 0.001, 0.01] and C: [1, 10, 100] in kernel-SVM, L1_ratio: [0.01, 0.1, 0.9] and C: [0.1, 1, 10] in Elastic-Net, C: [0.01, 0.03, 0.1, 0.3, 1, 3, 10, 30, 100] in logistic-regression, respectively.

All variables were preprocessed with standardization or smooth-rigid-transform depending on the types of variables. In model-construction, the Boruta-method was employed as the feature-selection technique for reducing variables that did not contribute to the discrimination of amyloid-positivity from the amyloid-negative group and improvement of explanation ability [39]. Using the Boruta-method, features were selected in the training-models.

Feature-selection-with-Boruta was also conducted within the cross-validation-loop using only training-data in this study. The other machine-learning-models were also constructed by combining variables after feature selection using the Boruta-method with MoCA-J-scores or Active-Scale-data (Model-4–6).

In order to avoid information-leakage of the test dataset and biased-evaluation of model-performance, we used the group-fivefold-cross-validation to assess our models, in which records from the same participant were assigned to the same fold. No holdout was set in training the models.

In this study, each model was trained with fivefold-cross-validation with 10-different-seeds (10-repeats), and mean of area-under-a-receiver-operating-characteristic-curve (ROC-AUC) was utilized for evaluation. Also, other evaluation-indices, precision, recall, and F1-score as a harmonic-mean of precision and recall, were calculated.

In this study, the cutoff-point to calculate these indices was selected using Youden-index computed from training-dataset. The ROC-curve for the fivefold-cross-validation in the fifth-constructed-models from the top-of-10-times-seeds-randomization were also outputted to confirm generalized-performance. In addition, the permutation importance of each variable finally remained in each model was computed as feature-importance [40]. The series-of-steps in model-training and performance validation were conducted using Python version 3.7.7.

Results - Clinical and Demographic Characteristics of the Participants Table 1 shows data on the demographic characteristics, wearable-sensor parameters, cognitive-function, and PiB-uptake-value of the participants. At baseline, the median-age (IQR) of the participants was 75.5 (71.00–80.00 years), and 68 (55.7%) were women. The median-education-duration (IQR) was 12 (9–12) years, and the median-BMI (IQR) was 23.19 (21.28–24.93). [Obviously much better than a random-sample would be from here in the fat-USA!]

In terms of physical-activity-variables, the median-daily-steps (IQR) was 4,425 (2,840–6,365), [my own being about 4,500 at age-81 with a timed-30-minute, 1-mile, daily-loop, walking exercise,] and the median-times spent on LPA, MVPA, and sedentary-behavior (IQR) were 16.46 (10.55–28.64), 21.46 (11.16–35.13), and 788.35 (739.65–830.27) min/day, respectively, [all also measured by my Apple Ultra-2 watch so it too should be useable in such a study as this or a future one like it.]

In terms of sleep-variables, the median daily TST (IQR) was 400.12 (355.96–445.53) min, the median daily sleep-efficiency (IQR) was 0.96 (0.93–0.97). The median daily conversation-time (IQR) was 213.79 (173.46–270.91) min, and the median-daily heart-rate (IQR) was 63.48 (59.63–68.79) beats/day.

The walking steps and sleep-parameters of our participants were similar to those in previous studies that assessed Japanese-adults of a similar-age [41, 42]. The median MoCA-J-score (IQR) was 22 (19–25). The median-PiB-SUVR (IQR) was 0.92 (0.83–1.32). Based on a PiB-PET-SUVR-cutoff of 1.4, 28 (23.0%) of 122 individuals were amyloid-positive. [I fall into an uncertain region using a 3-part-division and C₂N-proability-score.]

In addition, 17 (13.9%) individuals [such as myself] were APOE4-carriers. Therefore, the prevalence of individuals with abnormal-PiB-uptake was relatively-smaller than that of individuals in previous studies [10, 43]. Of individuals with MCI followed up to the second-year (N = 34) and the third-year (N = 61), 1 [of] 2 respectively converted to dementia in the second and the third years, while 3 [of] 17 reverted to normal-cognition. The average-annual-conversion and reversion-rate were 1.9% and 12.8%, respectively. [See, there is hope even after a questionable plasma-measure, such as I had with the earlier-version of now much-improved new version of that C₂N measure, which I have a physician’s order for but have yet to find a lab with the kit needed to perform it here in the Cincy/Covington area.]

Model Prediction Accuracy The current study applied machine-learning-models for assessed data on 54-variables collected from 282-records, which included valid wearable-sensor, cognitive-test, and PiB-PET-imaging-data, [all of which I currently have for myself and longitudinally-measured over a multi-year period.] Among 282-records, 68 (24.1%) and 214 (75.9%) indicated amyloid-positivity and negativity (Table 2).

Three machine-learning-models (namely, kernel-SVM, Elastic-Net, and logistic-regression) were constructed using 25-37-variables (25 for kernel-SVM, 35 for Elastic-Net, and 37 for logistic-regression). These variables were selected from initial 54-variables, except two variables of MoCA-J-score and Active-Scale using the feature-selection-algorithm Boruta-method. These variables were defined as basic-features.

The other machine-learning-models were also constructed using basic-features combined with MoCA-J-scores or Active-Scale data. The machine-learning-models were constructed not only by adding basic-features, but also by replacing exercise-variables of basic-features because Active-Scale information is a substituted-variable for exercise-habits.

In addition, to confirm the contribution of variables in basic-features, the machine-learning-models were also established using age, sex, education-years and BMI (Model-0 in Table 3), only wearable-sensor-variables or wearable-sensor and demographic-variables with feature-selection using the Boruta-method (Model-1–6 in Table 3).

Table 3 shows the mean-predictive-accuracy of each model and combination dataset. The mean ROC-AUC of 10-times-seeds were 0.79 for all three models, using basic-features (wearable-sensor, demographic-characteristics, health and life-environment questionnaire-features) (Model-3 in Table 3).

Precision in Model-3 were 0.49, 0.51, 0.51 for kernel-SVM, Elastic-Net, and logistic-regression, and regarding recall, the values showed 0.69, 0.63, 0.58 for kernel-SVM, Elastic-Net, and logistic-regression, respectively.

The F- score, a harmonic-mean of precision and recall, were 0.56, 0.55, 0.51 for kernel-SVM, Elastic-Net, and logistic-regression, respectively. Furthermore, MoCA-J was combined separately with basic-features and used to construct the three models.

The predictive-accuracy improved from 0.79 to 0.83 after combining basic-features with the total-MoCA-J-score (Model-4 in Table 3). Precision and recall of three algorithms in Model-4 went up-or-down comparing with Model-3, and the F1-score (Model-4 in Table 3) was slightly-higher than before combining the total MoCA-J-score (Model-3 in Table 3) in all 3 algorithms.

Moreover, the ROC-AUC of the models with basic-features and Active-Scale data was approximately 0.80, which was similar to that of Model-3, and the ROC-AUC was similar even after removing exercising-features including steps, period of sedentary-behavior, LPA, and MVPA (Models-5 & 6 in Table 3).

About Model-5 and Model-6, other evaluation-indices except ROC-AUC showed similar tendency of Model-3 and Model-4. The mean ROC-AUC of 10-times-seeds were 0.61, 0.70, 0.70 for kernel-SVM, Elastic-Net, and logistic-regression, respectively, using wearable-sensor-variables alone (Model-1 in Table 3). However, the values of precision, recall, F1-score showed considerably lower than the values of Model-3 in all algorithms.

The ROC-AUC of the models with wearable-sensor and demographic variables had fair performance (approximately 0.77, Model-2 in Table 3). As for precision, recall and F1-score of these models, not as much as them of Model-1, showed low values. In addition, the mean ROC-AUC were 0.72, 0.75, 0.74, precision were 0.41, 0.42, 0.43, recall were 0.63, 0.71, 0.69, and F1-score were 0.48, 0.52, 0.51 for kernel-SVM, Elastic-Net, and logistic-regression, respectively (Model 0 in Table 3).

Figure 1 shows the ROC-curves of each machine-learning-model using basic-features (Model-3 in Table 3) showed the variation of the fifth constructed models from the top-of-10-times-seeds-randomization for the fivefold-cross-validation.

The mean-AUC as the mean-cross-validation performance of our models, which had generalized-performance, were 0.80 ± 0.12 SD for kernel-SVM, 0.80 ± 0.10 SD for Elastic-Net, and 0.78 ± 0.13 SD for logistic-regression. Figure-2 shows the boxplot of the cross-validated-AUCs across all-10-repeats in each model, Model-3 and Model-4 demonstrated higher-performance than either Model-0 using age, sex, education-years and BMI or Model-1 using wearable-sensor-data only.

Table 3 Evaluation Indices of Amyloid-Positivity Prediction Models Using Machine-Learning Methods

Full-size table

Fig. 1

Receiver operating characteristic (ROC) curves in each three machine-learning-models. Every ROC curve represents the results in the fifth model from the top of 10 times seeds for the fivefold cross validation results of predicting amyloid-positivity. The blue line shows the mean ROC for the fivefold cross validation, and the red dot line shows the chance and ROC of each fivefold according to different colors. The gray shadow shows ± 1 standard deviation of the mean ROC. a kernel-SVM, b Elastic-Net, c logistic-regression

Full-size image

Fig. 2

The boxplot of the cross-validated AUCs across all 10 repeats. The flier points are those past the end of the whiskers extending from the box by 1.5 × the inter-quartile range (IQR). a kernel-SVM, b Elastic-Net, c logistic-regression

Full-size image

Visualization of Feature Importance To make the machine-learning interpretable and explainable, the feature-importance of all models were performed using the permutation-importance-estimate-method. Figure 3 shows the feature-importance listing for features. In total, 25 features were extracted in the kernel-SVM, which included the following: 8 features related demographic characteristics and chronic diseases (age, education duration, BMI, ever drinking, hypertension, stroke, heart-disease, and thyroid-disease); 3 features related to physical-activities (steps, LPA, and MVPA), 2 features related to sleep-parameters (nap-efficiency and awakening-time count), heart-rate, and conversation-time from wearable-sensor; 10 features related to living-environment and health-behaviors (living-with-spouse-or-children, transportation-mode, accompanying-person, number-of-days-working-in-a-week, primary-hobby, time-spent-on-watching-TV, number-of-days-participating-in-community-activity, communication-frequency, and number-of-outings).

Similarly, 34 features were extracted in the Elastic-Net model, which included the following: 12 features related to demographic-characteristics and chronic-diseases (age, sex, education-duration, BMI, ever-drinking, food-allergies, hypertension, diabetes-mellitus, hyperlipidemia, stroke, heart-disease, and thyroid-disease); 4 features related to physical-activities (steps, LPA, MVPA, and sedentary-behavior), 3 features related to sleep-parameters (naptime, nap-efficiency, and awakening-time-count-during-nap), heart-rate, and conversation-time from wearable-sensor; 13 features related to living-environment and health-behaviors (living with children, transportation mode, accompanying person, number of days working in a week, primary hobby, reading a newspaper, lesson or class frequency, number of days participating in community activity, communication frequency, number of outings, pet ownership, caring about appearance, and denture).

Third, 37 features were extracted in the logistic-regression-model, which included the following: 11 features related to demographic characteristics and chronic diseases (age, sex, education duration, BMI, ever drinking, food allergies, hypertension, diabetes mellitus, stroke, heart disease, and thyroid disease); 4 features related to physical activities (steps, LPA, MVPA, and sedentary behavior), 5 features related to sleep parameters (TST, sleep efficiency, naptime, nap efficiency, and awakening time count during nap), heart-rate, and conversation time from wearable-sensor data; 15 features related to living environment and health behaviors (living with spouse, children, or grandchildren, transportation mode, accompanying person, number of days working in a week, primary hobby, exercise frequency, reading a newspaper, lesson or class frequency, number of days participating in community activity, communication frequency, number of outings, pet ownership, and denture).

Finally, as shown in Fig. 4, 22 variables were common to all three machine-learning-models: 8 features were related to demographic-characteristics and chronic-diseases (such as age, education-duration, BMI, ever-drinking, hypertension, stroke, heart-disease, and thyroid-disease), 3 features related to physical-activities (such as steps, LPA, and MVPA), 1 feature related to sleep-parameter (such as nap-efficiency), heart-rate, and conversation-time, and 8 features related to living-environment and health-behaviors (such as living with children, transportation mode, accompanying person, number of days working in a week, primary hobby, number of days participating in community activity, communication frequency, and number of outings).

Fig. 3

The feature importance ranking table extracted in each three machine-learning-models. The vertical axis labels show the explanatory variables, and the horizontal axis labels depict the feature importance of each explanatory variable. a kernel-SVM, b Elastic-Net, c logistic-regression

Full-size image

Fig. 4

Variables that remained in each of the model. Venn diagram showing the variables that finally remained among the three models and 22 common variables in all three models

Full-size image

Discussion The current study established machine-learning-models for predicting PET-amyloid-positivity using 3-categories (objectively measured lifestyle-factors alone or in combination with demographic-characteristics and living-environment and health-behaviors). The machine-learning-technique was used to evaluate 54 variables collected from 122 participants without dementia at baseline (118 with MCI and 4 with subjective-memory-complaints). Further, the models integrating 3 categories had a better predictive-accuracy (AUC of 0.79) than the models using wearable-sensor-data alone (AUC of 0.70). Moreover, the predictive-accuracy (AUC of 0.83) slightly-improved with the integration of MoCA-J with 3-categories.

Given the performance of previous studies and the aim of our study model to be utilized for prescreening of amyloid-positivity in the brain, we considered the performance in this study (ROC-AUC = 0.79) to be fair and acceptable in real-world settings. To the best of our knowledge, this study first developed and validated models using machine-learning-techniques to predict-PET-amyloid-positivity with lifestyle-variables collected using wearable-sensors and questionnaires in community-dwelling-older-adults. [They haven’t gotten me into my planned eventual residence at White Sands or similar in La Jolla, CA.]

These machine-learning-models are useful for prescreening to detect-amyloid-positivity and can reduce costs and the number of unnecessary invasive lumbar-puncture and amyloid-PET in the clinical-trials in AD or clinical settings. The most interesting finding of the current study was that machine-learning-models using objectively-measured lifestyle-factors combined with demographic-characteristics and living-environment and health-behaviors had acceptable-performance on predicting-amyloid-positivity-on-PET. [The technique that I helped to fund the development of at the U. of Pittsburgh.]

These models, which are based on different machine-learning-techniques, had similar predictability (0.79 in all three algorithms). This finding suggested that the ability of each model is reliable for learning and making associations within and between data despite the utilization of different classification-techniques. Research on predictive-models have investigated several machine-learning techniques. The different machine-learning techniques have similar predictive performance for amyloid-positivity and negativity. This finding increases the reliability and generalizability of predictive-models [44]. However, the AUC-value slightly-increased if the MoCA-J-score was added (kernel-SVM: 0.83, Elastic-Net: 0.83, and logistic-regression: 0.82 respectively).

The assessment of MoCA-J required the trained medical staffs or clinical psychologists. Several studies have developed machine-learning-models to predict brain-amyloid-deposition using demographic-characteristics, neuropsychological-test-results, APOE-genotype, neuroimaging-findings, and blood-base-biomarkers, [such as C₂N’s newest blood-plasma measure, now approved for use in the U.S.]

Some studies have reported the use of two predictive-models using demographic-characteristics and neuropsychological-test results. The AUC of the model using age, family-history, online-cognitive-function-instrument-scores, and Cogstate was 0.806 [19]. The AUC of the model that used Mini-Mental State Examination (MMSE) score, Alzheimer’s Disease Assessment Scale, American National Adult Reading Test, Rey Auditory Verbal Learning Test, clock-drawing, and logical-memory-delayed-recall was 0.864 [20].

Moreover, 4 models have used the APOE-genotype combined with demographic-characteristics and neuropsychological-test. The AUC-values were 0.65 in the model using age, sex, education-duration, ApoE4, baseline-cognition, and longitudinal-cognitive-rate [21], 0.72 in the model using age, sex, education-duration, ApoE4, and neuropsychological-tests (such as MMSE-score, Alzheimer’s-disease-assessment-scale, and logical-memory-II) [22], and 0.83 in the model using age, 10-word-delayed-recall, and ApoE4 [23]. The model using age, sex, education duration, history of hypertension, ApoE4, and word-list-recall-test-score had the highest-accuracy at 0.873 [24].

Machine-learning-models using neuroimaging or blood-base-biomarkers combined with demographic-characteristics, ApoE4, and neuropsychological-testes had better performance in predicting-brain-amyloid-positivity. The AUC of the model using MRI radiomic-features combined with age, sex, and ApoE4 was 0.79 [25]. The AUC of the model using MRI-volumetrics combined with age, sex, education-duration, ApoE4, and neuropsychological-tests was 0.71 [22]. The AUC of the model using 6 blood-based-markers combined with age, ApoE4, and CDR was 0.87 [11]. The AUC of the model using plasma-Aβ42/Aβ40 combined with age, 10-word-delayed-recall-score, and ApoE4 was 0.85 [23].

The AUC of the model using blood-derived p-tau or amyloid-beta in a series of studies was exceeding 0.80 [12, 13]. There was the study reporting the capacity of plasma-Aβ42/Aβ40 measured using 6 different assays that could predict amyloid-positivity with good performance [14]. Another study also showed that good amyloid-PET-positivity prediction-accuracy by using MMSE, age, and APOE in addition to blood-test [26].

These findings suggested that machine-learning-models combined with demographic-characteristics, cognitive-test-results, APOE-genotype, neuroimaging, or blood-base-biomarkers can be used to identify individuals with elevated brain-amyloid-deposition. The model-prediction-accuracy of the current study was comparable to that in previous studies.

Note that some of the previous studies listed above differ with our study in that they used study-population including cognitively-normal-participants [11,12,13, 19,20,21]. Moreover, there was a study establishing the models using lifestyle-factors and other dominant-variables such as APOE, MMSE, and in its study, although study population is participants including cognitively-normal, the model using only-lifestyle-factors was also established [27]. The advantage of our models is that lifestyle-factors objectively-measured with wearable-sensor and living-environment and health-behaviors are completely noninvasive and easily-available in the community.

The 3 machine-learning techniques used common-variables, such as demographic-characteristics, chronic-diseases, physical-activity, nap, heart-rate, conversation-time, and living-environment and health-behaviors in the predictive-models. These variables are important for differentiating individuals who are positive-for-amyloid from those who are negative-for-amyloid.

The important-predictors included demographic-characteristics such as age, education-duration, BMI, ever-drinking, hypertension, stroke, heart-disease, and thyroid-disease. Most of the previous machine-learning-models for predicting-brain-amyloid-deposition included age, sex, and education-duration. It is well established that advanced-age is the greatest-risk-factor-for-AD [45] and associated with the higher-prevalence-of-Aβ-positivity [46, 47].

The proportion of individuals with normal-cognition who are positive-for-amyloid generally increased with aging [10, 47]. [Similar to my own case of amnestic-MCI-with-mixed-etiology.] Education is considered as an indicator of cognitive-reserve and individuals with higher-level-of-education have greater-brain-amyloid-burden than those with lower-levels-of-education [48,49,50].

Sex did not remain as common-variables in all 3 models. Although the incident-rate-of-AD is higher in women than man [45], the association between sex and brain-amyloid-burden is controversial [51, 52].

BMI may be bi-directionally associated with brain-amyloid-burden. It has been reported that greater-brain-amyloid-burden was associated with the subsequent-decline-in-BMI and higher-BMI was associated with greater-brain-amyloid-burden [53, 54].

Moderate-alcohol-intake was associated with a lower risk of cognitive impairment or brain-amyloid-deposition in older-adults [55, 56].

Moreover, a growing body of evidence has shown the association between vascular-risk-factors, including hypertension, and hyperlipidemia or cerebrovascular-disease, and AD-pathology [57, 58]. By contrast, the association between diabetes-mellitus and brain-amyloid-deposition is inconsistent [59, 60]. The two-hit-vascular-hypothesis-of-AD has proposed that vascular-risk-factors contribute to the dysregulation of the neurovascular-unit, thereby resulting in chronic-hypoperfusion or impaired-Aβ-clearance and increased-Aβ-production [58].

The association between cardiovascular-disease and AD-pathology remains unclear [61]. However, several autopsy-cases showed that coronary-artery-disease was associated with brain-amyloid-deposition [62].

Thyroid-state, [I’ve had a partial-thyroidectomy,] was associated with brain-Aβ-deposition [63, 64]. In particular, triiodothyronine-negatively regulates the gene-expression of amyloid-precursor-protein [64, 65].

The important-predictors in wearable-sensor data included physical-activity, nap-efficiency, heart-rate, and conversation-time. Physical-activity or exercise is associated with lower-brain-amyloid-deposition on PET and higher-Aβ42-levels in the cerebrospinal-fluid among older-adults-without-dementia [66].

The mechanisms underlying the association between physical-activity and brain-amyloid-deposition suggests that physical-activity inhibits-amyloid-production and enhances-amyloid-degradation-or-clearance [67, 68].

Although short-sleep-duration, poor-sleep-quality, and frequent-napping, [I’ve never ever napped, nor do I now,] are associated with higher-brain-amyloid-deposition [28], only few studies have investigated the role-of-nap-efficiency. Previous findings on the association between daytime-napping and cognitive-function have been contrasting. Self-reported daytime-napping reduced-the-risk-of-cognitive-decline [69]. Meanwhile, more-frequent-napping measured using actigraphy was associated with a poorer-cognitive-function [70]. Further studies should be conducted to validate the association between nap-efficiency and brain-amyloid-deposition.

A higher-resting-heart-rate, [my own is quite-low and sleeping it is very-low,] is a risk-factor for not only stroke or cardiovascular-disease but also cognitive-decline or dementia in older-adults [71]. Nevertheless, the association between heart-rate and brain-amyloid-deposition remains unclear.

The important predictors related to living-environment and health-behaviors included living-together-with-a-relative, transportation-mode, number-of-days-working-in-a-week, hobby, exercise-frequency, and social-relationship. Conversation-time, living-with-children, number-of-days-participating-in-community-activity, communication-frequency, and number-of-outings are related to social-isolation or loneliness. Older-adults with less-social-participation-and-contact and subjective-loneliness are at-higher-risk-for-cognitive-impairment-and-dementia [72]. Moreover, active-social-engagement, including contact-with-family-and-friends and positive-social-support-and-engagement-in-leisure-activities, play a role in preventing-cognitive-impairment [73]. The association between loneliness and brain-amyloid-deposition has been found in older-adults with healthy-cognition [74].

The transportation-mode and accompanying-person (the need for company during hospital-visits) were important predictive-variables in each model. The number of older-adults who retire from driving increased according to stringent-licensing-polices in Japan. Older-adults who stop-driving are at high-risk-of-depression, general-health-decline, cognitive-impairment, social-isolation, and mortality [75, 76]. [My ex-PCP keeps trying to force me to stop driving my c8-Corvette-HTC and move into White Sands. He most-recently forced me to spend $365 taking a Bick’s Driving Test evaluation that I passed with a near-perfect driving-score, with the only negative being that I “signaled too soon before an expressway exit!” I fired that Cincy PCP immediately afterwards, since he had previously had me waste an entire day’s time of both myself and my best-friend in doing that same driving-test. It cost me the last day with my friend, who passed-away shortly thereafter, robbing us of our last day together, because of this useless exercise, only to be replaced with demanding the additional Bick’s half-day, multi-factor, evaluation.]

Although alternative-transportation is required to maintain independent-mobility for shopping or social-connectedness, the public-transport-network is inadequate, particularly in rural-areas. A higher-level-of-AD-biomarkers-in-the-CSF could be a determiner of early-driving-cessation among older-adults [77]. Moreover, transportation with family or friend is attributed to impairment-in-instrumental-activities-of-daily-living, which is associated with brain-amyloid-deposition [78].

Number-of-days-working-in-a-week [I’ve worked every day of the week for the past two-decades on this effort,] is a protective-factor against decline-in-cognitive-function or basic-activities-of-daily-living. However, it is not a predictor of elevated-brain-amyloid-deposition in a longitudinal-observational-cohort-study [79]. This finding is inconsistent with that of our machine-learning-models probably due to differences in study-design, analytic-methods, or age-of-participants.

The current study had several strengths. That is, lifestyle-factors, such as physical-activity, sleep, and conversation, were continuously and accurately measured using a wearable-sensor in community-dwelling-older-adults. Further, brain-amyloid-deposition was assessed via PiB-PET. [Huray, since I helped to fund its development!]

Limitations The current study had several limitations that should be considered. First, the predictive-model for brain-amyloid-deposition in an independent-cohort was not-validated. However, this is also a common limitation in previous-studies. Hence, further large-scale, multicenter-studies should be conducted.

Second, we collected clinical-data to define the presence or absence of dementia-at-baseline and not all participants with possible-dementia could be excluded from the study. Our participants-with-MCI from the community had an average-annual-conversion-rate-of-1.9% and reversion-rate-of-12.8%. Annual-conversion-rate from MCI-to-dementia in community-based-studies was lower than that in clinic-based-studies [80], whereas annual-reversion-rate-from-MCI-to-normal-cognition-in-community-based-studies was higher than that in clinic-based-studies [81, 82]. The community-sample had a conversion-rate of approximately 3% to 6% and a reversion rate of approximately 25%-to-30% [80, 81]. Therefore, the conversion and reversion rates in our cohort were relatively-lower than those reported in previous community-based-studies.

Third, participants recruited from 855 community-dwelling-individuals in this study were 122 with MCI or subjective-memory-complaints and the number-of-individuals-with-MCI-who-presented-with-abnormal-PiB-levels was relatively-small. In addition, the sample-size between the amyloid-positive (n = 68) and the amyloid-negative group (n = 214) was imbalanced, with a ratio-of-1:3. This imbalance could lead to the construction of models with respect to the majority-class. Therefore, more-sampling-techniques and larger-cohorts are needed to address those sample-size and class-imbalance issues in future studies.

Conclusion In conclusion, we developed machine-learning-models for predicting-PET-amyloid-positivity using easily-available and noninvasive-variables without the need to visit a hospital. Our models are useful for prescreening on enrollment of subjects who seem with brain-amyloid-deposition to reduce screen-failure-rate and trial-costs in clinical-trials. Furthermore, our models are useful for clinicians to identify the individuals who really need to conduct lumbar-puncture or amyloid-PET-scan even after the disease-modifying-therapy being approved. [My thoughts exactly!]

Availability of Data and Materials Data cannot be shared publicly due to ethical restrictions. The participants signed an informed consent form, which states that their data are exclusively available for research institutions in an anonymized form. The raw data used in this study contains sensitive and identifying information on individuals including gender, age, and education level that could compromise the privacy of research participants. However, the data that support the findings of this study are available upon ethical approval by the local ethics committee of the Oita University Hospital. Please contact the ethics committee of the Oita University Hospital. Email: This email address is being protected from spambots. You need JavaScript enabled to view it..

References

van Dyck CH, Sabbagh M, Cohen S. Lecanemab in Early Alzheimer’s Disease. Reply N Engl J Med. 2023;388(17):1631–2. https://doi.org/10.1056/NEJMoa2212948.

Article Google Scholar

Dubois B, Feldman HH, Jacova C, Hampel H, Molinuevo JL, Blennow K, et al. Advancing research diagnostic criteria for Alzheimer’s disease: the IWG-2 criteria. Lancet Neurol. 2014;13(6):614–29. https://doi.org/10.1016/S1474-4422(14)70090-0.