Edited by: Roberto Monastero, University of Palermo, Italy
Reviewed by: Rufus Olusola Akinyemi, University of Ibadan, Nigeria; Kay Deckers, Maastricht University, Netherlands
This article was submitted to Dementia, a section of the journal Frontiers in Neurology
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Dementia is a major public health issue worldwide, with a serious burden for patients, caregivers, and society, as well as substantial economic impacts (
We have been conducting a community-based observation study focusing on lifestyle risk and preventive factors related to dementia in Usuki city, in southern Japan, since 2015. The proportion of the population over 65 years old in Usuki city has reached 38%, compared with 27.3% of the nationwide population in Japan. In the present study, public servants carried out public relations initiatives to recruit participants aged 65 or older without dementia from the entire city using electronic and paper-based media because the lifestyle factors such as physical activity and social isolation are closely related to late-life cognitive impairment (
Flow of participant recruitment.
All participants were asked to wear a wristband sensor (Silmee™ W20, TDK Corporation Tokyo, Japan) on the wrist, except while bathing. We excluded data if the heart rate count indicated that the wristband sensor had been removed. Sixteen of 98 participants who had inadequate sensing data (16.3%) did not wear wristband sensor during sufficient period for analysis. Our wearable sensor measured various lifestyle parameters, including heart rate, walking steps, conversation time, total sleep time (TST), sleep efficiency, time awake after sleep onset (WASO), awakening count, and napping time. These parameters were calculated by summing sensing data each day and averaging this over the whole measurement period.
Physical activity data were detected by a 3-axis accelerometer, which enabled measurement of acceleration in three perpendicular axes. The evaluation circuitry converted the output of a micromechanical acceleration-sensing structure, according to the differential capacitance principle. The accelerometer generated physical activity data regarding walking steps with composite acceleration of 3-axis measurement every time the wearable sensor was moved, and data were captured continuously and summarized in 1 min intervals. Walking steps were identified by capturing frequency bands ranging from 2 to 3 Hz, which synthesized acceleration by the accelerometer. The number of walking steps was calculated by summing the number of steps for each day and dividing this by the number of days of lifestyle data measurement. Therefore, the number of walking steps was represented as the average number of steps per day.
Sleep-wake parameters were assessed by the magnitude of synthesized acceleration of the 3-axis accelerometer and cumulative energy. The data were confirmed and adjusted by qualified technicians using visual inspection. Time in bed between bedtime and waking time was determined by the activity count recorded by the wristband sensor. Sleep parameters such as TST, WASO, and sleep efficiency, as well as the awakening count were measured from 18:00 in the evening to 5:59 the following morning (
Indices of wearable sensor.
Heart rate was detected by photoplethysmography. Pulse photoplethysmography is a simple and useful method for monitoring heart rate. Using this method, pulse measurement is based on the irradiation of 573 nm wavelength light, and the conversion of the intensity of reflected light to an electrical signal. The heart rate was calculated by summing pulses per min for each day and dividing this by the number of days in the lifestyle data measurement period. Therefore, heart rate was represented as the average number of pulses per day.
Our wearable sensor could not detect the content of conversation, but could detect utterances of the participant wearing the wearable sensor, and utterances of nearby people. Sound data were captured continuously during the presence or absence of a conversation every minute. Although the utterances of other individuals were included in the sound data, participating in the conversation was considered to be important for social activity in this study.
Sound data were collected by a microphone on the wearable sensor, and analyzed to evaluate the conversation time. Our wearable sensor detected sound pressure, which was produced by utterance within a 2-meter radius from the device. The sound pressure range was from 55 to 75 dBA at this distance. The conversation time was defined as the frequency components included in conversation data extracted by signal processing. In detail, the wearable sensor extracted a frequency band corresponding to a human voice from the sound data within the sound pressure range as a sound frame. Conversation was defined as more than four sound frames per minute, during a 1-min period. It is possible that the sound of television viewing or radio listening was detected as conversation due to the detection method based on the sound pressure and frequency. Therefore, we also quantified the detection rate of television viewing.
The accuracy of walking step detection was verified by comparing the sensing data and video observation data. Walking steps were simultaneously collected for 9 min by wristband sensor and continuous video monitoring in twenty healthy participants aged 60–80 years (11 men and 9 women). Significant correlation was found between walking steps measured by wristband sensor and those from video observation (
The accuracy of sleep duration detection was verified by comparing the sensing data and video observation data. Sleep duration was simultaneously collected during night time by wristband sensor and continuous video monitoring in five healthy participants aged 20–60 years (5 men). Significant correlation was found between sleep duration from wristband sensor and that from video observation (
The accuracy of conversation time detection was verified by comparing the sensing data and self-report data regarding conversation time. Sound data were captured for 50 h in healthy participants aged 30–40 years and analyzed in terms of precision, recall, and F-Measure (
RF regression analysis was used to examine the relationships between total daily sensing data and MMSE score. RF is an ensemble learning method that operates by constructing decision trees using bootstrap aggregation, and computes node impurity for every variable. This analysis can be used to rank the variables based on their predictive importance, such as %IncMSE and IncNodePurity. Moreover, confounding factor analysis was performed to build an unadjusted model (model 0), a model adjusted for demographic factors (model 1), and a model adjusted for demographic and vascular risk factors (model 2). The selected variables were used to assess the risk and protective factors for cognitive function using partial dependence plots (PDPs).
RF was conducted using R (version 3.4.1) and an RF package for Windows 10 (
Top N variables selection process.
All variables for RF analysis.
Age (years) |
Gender (0; Male, 1; Female) |
Education (years) |
BMI (kg/m2) |
Smoking status (0; Every day, 1; None, 2; Sometimes) |
Alcohol consumption (0; Every day, 1; None, 2; Sometimes) |
Hypertension (0; No, 1; Yes) |
Diabetes mellitus (0; No, 1; Yes) |
Hypercholesterolemia (0; No, 1; Yes) |
Walking steps (steps/day) |
Conversation time (mins/day) |
Heart rate (counts/mins/day) |
TST (mins/day) |
WASO (mins/day) |
Sleep efficiency (%/day) |
Awakening time count (counts/day) |
Nap time (mins/day) |
Potential confounding factors included age, gender, education levels, and BMI, alcohol consumption, smoking status, and vascular risk factors, which may affect both lifestyle factors and cognitive function (
Confounding factors in model 0, 1, and 2.
Unadjusted | Adjusted for age, gender, education year | Adjusted for age, gender, education year, BMI, hypertension, diabetes mellitus, hypercholesterolemia, alcohol consumption, smoking status |
The variables in model 1 identified by RF and confounding factor analysis were assessed for risk and protective factor analysis. In black box methods like RF analysis, functional relationships between each independent variable and the response variable are assessed using PDP ( For xj, sort the unique values V = {xj}iϵ{1,…,n} resulting in V*, where |V*| = K. Create K new matrices X(k) = (xj = V*k, X–j), ∀k = (1,…,K). Drop each of the K new datasets, X(k) down the fitted forest resulting in a predicted value for each observation in all K datasets: ∧y(k) = ∧f(X(k)), ∀k = (1,…,K). Average the predictions in each of the K datasets, ∧y* k = 1 nPN I = 1 ∧y (k) i, ∀k = (1,…,K). Visualize the relationship by plotting V* against ∧y*.
PDPs are generated from the RF model and used to define the variables as risk and protective factors for MMSE scores. The marginal prediction data were extracted from the PDP for all the variables in the RF model. The correlation value was calculated between each independent variable and MMSE score, using the marginal prediction data. Variables with positive correlation values were defined as protective factors, whereas those with negative correlation values were defined as risk factors. Moreover, PDPs for each selected variable were used to determine risk and protective factors.
The variables in model 1 identified by RF and confounding factor analysis were applied in multiple linear regression for quantifying the contributions of individual risk and protective variables. *
Moreover, the model prediction accuracy was verified by linear regression and RF analyses. The training models, including linear regression and RF algorithms, were built using the data set with the variables walking steps, conversation time, TST, heart rate, age, gender, and years of education, and were evaluated with the test data. The performance of the model was evaluated by the prediction RMSE with the lowest value. We used 90% of the 855 samples for training, and 10% of the 855 samples for testing.
Summary of demographic characteristics and wearable sensor data of participants.
Age (years) | 73.8 ± 5.8 |
Gender (M:F) | 317:538 |
Education (years) | 11.8 ± 2.1 |
BMI (kg/m2) | 23.2 ± 3.1 |
Median MMSE scores | 29 (20, 30) |
Ever smoker | 36 (4.2%) |
Ever drinker | 354 (41.4%) |
Hypertension | 429 (50.2%) |
Diabetes mellitus | 114 (13.3%) |
Hypercholesterolemia | 281 (32.9%) |
Walking steps (steps/day) | 5452.9 ± 2778.0 |
Conversation time (mins/day) | 219.7 ± 86.3 |
Heart rate (counts/mins/day) | 64.7 ± 6.3 |
TST (mins/day) | 408.4 ± 69.1 |
WASO (mins/day) | 22.1 ± 14.1 |
Sleep efficiency (%/day) | 1.0 ± 0.0 |
Awakening time count (counts/day) | 0.5 ± 0.3 |
Nap time (mins/day) | 48.7 ± 39.3 |
RF regression analysis using sensing data revealed that the five variables (walking steps, conversation time, TST, and WASO as well as heart rate) in model 0 were selected. Moreover, the top 10 variables (walking steps, conversation time, TST, and WASO, sleep efficiency, awakening time count, naptime, heart rate as well as age and education years) in model 1 were selected based on the lowest RMSE value (1.6).
In model 1, three variables, including sleep efficiency, awakening time count, naptime were newly selected and WASO exhibited a < 200% decrease in the estimated parametric values (−440.91%). Therefore, these variables were excluded for next step analysis. Finally, four variables (walking steps, conversation time, TST and heart rate) were included in model 1 for risk and protective factor analysis. The IncNodePurity value of each variable is 313.7 in walking steps, 258.8 in TST, 225.1 in heart rate, and 220.3 in conversation time (
Variable importance measure. The IncNodePurity value of each variable is 313.7 in walking steps, 258.8 in TST, 225.1 in heart rate, and 220.3 in conversation time.
Four variables in model 1 were assessed in the protective and risk factor analysis. The number of walking steps, conversation time, and heart rate exhibited positive correlations with MMSE score and were categorized as protective factors for cognitive function, whereas TST was categorized as a risk factor. PDPs of walking steps and heart rate revealed continuously increased MMSE scores. The inclination of the graph, however, began to reverse by the boundary of the specified threshold in the PDP of conversation time and TST (
Partial dependency plot for the actinography data. The number of walking steps
Correlation analysis between the number of walking steps and conversation time. The daily number of walking steps was not correlated with conversation time in participants exhibiting <1.125 min (transformed value, mapping value: 320 min) of conversation, and decreased with increasing conversation time in participants exhibiting more than 1.126 min (transformed value, mapping value: 321 min) of conversation.
To quantify the individual risk and protective variable contributions, multiple linear regression was performed using only the selected four variables (walking steps, TST, heart rate, and conversation time) after transforming the data to the normal distribution (
Risk and protective variables contribution.
Walking steps (steps/day) | 0.4116 | 0.0722 | 5.702 | 1.63e−08* |
TST (mins/day) | −0.1128 | 0.0731 | −1.542 | 0.123 |
Heart rate (counts/mins/day) | 0.1071 | 0.0726 | 1.476 | 0.140 |
Conversation time (mins/day) | 0.0612 | 0.0721 | 0.849 | 0.396 |
The prediction accuracy in the RF model was better than that in the linear regression model (
Model accuracy.
Linear regression | 0.119 | 3.155 | 1.776 | 0.175 | 2.955 | 1.719 |
RF regression | 0.775 | 0.796 | 0.892 | 0.759 | 0.863 | 0.929 |
The present study evaluated lifestyle components, including physical activity, sleep, and conversation, as well as heart rate, using a wearable sensor in a large sample of community-dwelling older people, and constructed a machine learning model to predict cognitive impairment. RF regression analysis adjusted for age, gender, and education years identified four significant variables: walking steps, conversation time, TST, and heart rate. Specifically, walking steps, conversation time, and heart rate remained significant after the RF regression analysis was adjusted for demographic and vascular risk factors. Moreover, the walking steps, conversation time, and heart rate were identified as protective factors for cognitive function, whereas TST was categorized as a risk factor for cognitive function. PDPs of walking steps and heart rate revealed continuously increased MMSE scores. TST and conversation time, however, indicated that the tendency in the PDP graph was reversed at the boundary of a specified threshold. Thus, these variables tended to have a protective effect on cognitive function within a particular range of time, whereas longer periods of time exceeding a particular threshold were risk factors for cognitive function. Because RF regression analysis and PDP graph cannot quantify individual variable contribution, multiple linear regression analysis was performed to quantify individual contribution of the selected variables and also to find its statistical significance. Although the multiple linear regression showed that only the number of walking steps was significant, both RF and linear regression analysis revealed similar results regarding the risk and protective factors for cognition. Moreover, the PDP graph exhibited a reversal at the boundary of a specified threshold in TST and conversation time, which was not detected by linear regression. Therefore, we selected four lifestyle variables related to MMSE scores by RF regression analysis and risk and protective factor analysis of each variable were performed using PDPs. The current findings highlight the importance of physical activity, sleep, and conversation in preventing cognitive impairment among community-dwelling older people.
Numerous studies have examined the beneficial effects of physical activity on cognitive function among older people. A meta-analysis of 15 prospective cohort studies reported protective effects of vigorous exercise against cognitive decline (hazard ratio 0.62, 95% CI 0.54–0.70) (
Sleep is important for brain plasticity and memory consolidation (
The present findings identified daily heart rate as a protective factor for cognitive function among community-dwelling older people. To our knowledge, no previous reports have examined the relationship between heart rate and cognitive function. A previous study reported relationships between resting heart rate, depression, and cognitive impairment in patients with ischemic stroke, and relationships between reduced heart rate variability and cognitive impairment among older women (
Importantly, the current results revealed that conversation time was an important predictive factor for MMSE score. Social isolation and subjective loneliness are increasingly recognized as risk factors for cognitive impairment and dementia among older people (
The present study has several limitations that should be considered. First, the study could not determine the causal direction of the association between lifestyle factors and cognitive function because of its cross-sectional design. Second, we were unable to exclude factors influencing cognitive reserve, such as past, or current occupation and engagement in cognitive and social activity, which may affect lifestyle and cognitive function. Third, cognitive function was evaluated only by MMSE and information regarding depression was not collected. The MMSE is a very crude measure of cognition and questionable accuracy for detecting dementia. Although we collected the clinical information to define the present or absence of dementia, the patients with possible dementia could not be excluded completely from participating in the current study. Therefore, further studies assessing a broader range of cognitive domains should be needed to confirm our results. Forth, it is possible that the sound of television viewing or radio listening was detected as conversation due to the detection method based on sound pressure and frequency. Therefore, further studies are needed to improve the reliability of our sensing data. Conversation time may have included sleep or nap time during television viewing or radio listening. The relationship between conversation time and sleep or nap time, however, suggested that the possibility of sleeping being included in the daily conversation time was only 6.4%, which would not be expected to influence the results.
In conclusion, the current study revealed that lifestyle factors such as physical activity, sleep, and social activity were associated with global cognitive function among older people. Physical activity and heart rate were positively associated with cognitive function. Moreover, an appropriate balance between the durations of sleep and conversation appears to be important for cognitive function. These results may contribute to the development of new evidence-based interventions for preventing cognitive impairment and improving health and wellbeing among older people.
This prospective study was conducted in accordance with the Declaration of Helsinki, and was approved by the local ethics committee at Oita University Hospital (UMIN000017442).
NK and EM conceived and designed the trial. SN and MK developed the wearable sensor. YA, KY, MI, DH, YS, AN, SU, HF, SI, and TK performed the PET study and conducted the data analysis. AE performed the neuropsychological assessment. KS and MJ helped with the recruitment of participants.
SN and MK were employed by TDK Corporation. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
We gratefully acknowledge the assistance of Usuki city employees for their efforts in recruiting participants. We thank Suzuki Co., Ltd. for their assistance with data collection and HCL Technologies, Confidential, Fusa Matsuzaki for database construction and data analysis. We received generous assistance from Kaori Hirano, Megumi Ogata, Mai Kishigami, and Oita University students with physical measurement and interview procedures. The authors are sincerely grateful to all participants who enrolled in this study. We thank Benjamin Knight, MSc., from Edanz Group (
The Supplementary Material for this article can be found online at:
Linear correlation analysis of walking step. Significant correlation was found between walking steps from wristband sensor and those from video observation (
Linear correlation analysis of sleep duration. Significant correlation was found between sleep duration from wristband sensor and that from video observation (
Linear correlation analysis of conversation time. Significant correlation was found between conversation time from wristband sensor and that from video observation (
Results of conversation detection.
Analysis of false detection rate.
body mass index
confidence interval
mini mental state examination
partial dependency plot
random forest
root mean squared error
total sleep time
wake after sleep onset.