2025.5.15

・Introduction

Since 2022, our institute has been carrying out the "Project of promoting the health of older adults through the utilization of digital devices such as smartwatches", for 3 years under support from the Tokyo Metropolitan Government. The smartphone ownership rate and skills among older adults are increasing yearly (NTT Docomo Mobile Society Research Institute 2024), making it more realistic to promote the health of older adults using digital devices, such as smartwatches and smartphones. In particular, preventing frailty and the need for long-term care, which is necessary for healthy longevity, targets the decline in daily living functions, such as motor functions, oral functions, nutritional status, and cognitive functions. Thus, the early detection of frailty signs through the measurement of these functions in daily living is expected to be helpful for preventive measures. As part of this project, we conducted a study on the development of a smartphone application capable of detecting frailty and promoting its preventive measures through the use of wearable sensors (sensors that can be worn on the body), such as smartwatches.

・Overview of the study

Fig. 1 shows the overview of the study. Our institute has been conducting large-scale cohort surveys, including the "Otassha Study." The participants were asked for their cooperation in collecting data from their daily lives using wearable sensors, and long-term life-log data, such as the number of steps taken, physical activity levels, and walking speed during daily living, were collected using a smartwatch, ankle-band accelerometer, and smartphone GPS. The cohort surveys have been performed annually to collect data on health conditions, including frailty, cognitive decline, depression, social isolation, and the need for long-term care. Using these as training data, an algorithm can be developed using AI to predict health conditions based on life-log data. A high-performance model, if it can be developed, will enable us to identify the signs of frailty in daily life without requiring participation in an annual survey.

Fig. 1. Overview of the study.

・Participants

Participants were recruited from the cohort surveys of our institute (Itabashi and Chiyoda cohorts) and the frailty outpatient clinic at our hospital, and 90-day data were collected from approximately 1,000 individuals. According to the Japanese Frailty Criteria (Japanese Cardiovascular Health Study (J-CHS) Criteria) (Foundation for Longevity Science Promotion 2023), the participants were assessed as healthy (43.9%), pre-frail (50.0%), or frail (6.1%) (Table 1). Since a national survey showed that 50.5%, 40.8%, and 8.7% of people in Japan were healthy, pre-frail, and frail, respectively (Murayama et al. 2020), our participants included slightly fewer frail individuals, more pre-frail individuals, and fewer healthy individuals compared to numbers in the national survey data. However, we believe that we have collected data that are representative of older adults in urban areas.

Table 1. Summary of participants.

	Healthy		Pre-frail		Frail		Total
	Number	%	Number	%	Number	%	Number
Itabashi cohort	372	43.2%	439	51.0%	50	5.8%	861
Chiyoda cohort	51	52.6%	41	42.3%	5	5.2%	97
Frailty outpatient clinic	48	41.4%	57	49.1%	11	9.5%	116
Total	471	43.9%	537	50.0%	66	6.1%	1074

・Identification of frailty using life-log data

Based on the life-log data of the participants, learning was performed with 15 AI models, and we examined how well frailty could be identified, an example of which is shown in Fig. 2. A larger area under the curve (AUC) indicates a higher performance in identifying frailty. The diagonal line represents an AUC of 0.5, which indicates 50% performance in identifying frailty, and the closer the curve is to the upper left, the higher the performance is. The results showed that we were able to construct a learning model with a score exceeding 0.9 (Obuchi et al. 2024). The model has since been repeatedly revised, and the study is ongoing to further improve its performance.

Fig. 2. Receiver Operating Characteristic (ROC) curves for the identification of frailty using AI models

・Classification of lifestyle patterns using life-log data

We further classified the lifestyle patterns of the participants based on their long-term life-log data. Minute-by-minute data on sleeping, walking, and talking were collected using a smartwatch and were labeled with three colors, and data for 1 month were displayed as a tapestry image, as shown on the left side of Fig. 3. The horizontal axis represents 0 to 1,440 minutes, whereas the vertical axis represents 1 to 30 days. Based on the images, the characteristics of lifestyle patterns were classified using machine learning, which divided the lifestyle patterns of the participants into six clusters. The characteristics of the clusters were as follows: those in Cluster 0 had a long walking time and a long talking time; those in Cluster 1 had a long sleeping time and a short walking time; those in Cluster 2 had a long walking time and a short talking time; those in Cluster 3 wore a device poorly; those in Cluster 4 did not wear a device at night; those in Cluster 5 had long sleeping and talking times. Cluster 0 was associated with the absence of frailty (Obuchi et al. 2025).

Fig. 3. Classification of lifestyle patterns based on tapestry images.

・Implementation of the study results in a smartphone application

The study results were implemented in a smartphone application. The application can calculate the "frailty prevention score," which means the "degree of non-frailty," based on the life-log data of the previous day. The number of steps taken, metabolic rate, amount of talking, skin temperature, sleep time, and resting pulse rate, which contributed greatly to the calculation, were compared with the average values for the same age group, and the smartphone application displayed advice based on the results. A tapestry image displaying the lifestyle pattern over 1 month.

・ Conclusions

The Tokyo Metropolitan Government is currently preparing a project that utilizes this application. We plan to be involved in verifying the effectiveness of the project.

References

1. Foundation for Longevity Science Promotion: Healthy Longevity Network "Diagnosis of Frailty." Updated August 2, 2023. https://www.tyojyu.or.jp/net/byouki/frailty/shindan.html (accessed March 7, 2025).
2. Murayama H, Kobayashi E, Okamoto S, Fukaya T, Ishizaki T, Liang J, Shinkai S. National prevalence of frailty in the older Japanese population: Findings from a nationally representative survey. Arch Gerontol Geriatr. 2020:91:104220.
3. Obuchi S, Kawai H, Ejiri A, Imamura K, Sasai H, Fujiwara Y, and Hirano H. Identifying frailty using a walking ability meter: SWING-Japan Study. 83rd Annual Meeting of the Japanese Society of Public Health, Sapporo. October 29-31, 2024.
4. Obuchi S, Kawai H, Gong R, Wakui T, Shirabe M, Tamura Y, Yasuraku M, Fujiwara Y, Akishita M, Toba K, SWING-Japan Member. Classification of lifestyle patterns based on life-log data of community-dwelling older adults and its relationship with the prevalence of frailty and cognitive decline: SWING-Japan Study, 67th Annual Meeting of the Japan Geriatrics Society, Makuhari. June 27-29, 2025 (presentation pending).
5. NTT Docomo Mobile Society Research Institute: Seniors' smartphone skills are on the rise, with a particularly large increase among those in their early 70s. April 15, 2024. https://www.moba-ken.jp/project/seniors/seniors20250124.html (accessed March 7, 2025).

Implementation of the study results in a smartphone application