- Approximately 90% accuracy in detecting alcohol consumption using biometric data -
Development of multivariate prediction model for alcohol consumption and evaluation 
of effects on daily life

================================================================================

TechDoctor Inc. (CEO: Kazumichi Minato, Headquarters: Chuo-ku, Tokyo; hereinafter "TechDoctor") presented research results on the development of a multivariate 
prediction model for alcohol consumption and evaluation of the effects of alcoholic 
beverage intake on daily life at the 32nd Annual Meeting of the Japanese Society 
of Behavioral Medicine held at Sagamihara City Industrial Hall on December 6-7, 
2025 (general session, poster presentation).

This research developed technology to automatically detect alcohol consumption 
using biometric data obtained from wearable devices and verified its accuracy and 
effectiveness. The results showed the possibility of identifying the presence or 
absence of alcohol consumption with approximately 90% accuracy. Heart rate 
variability during sleep was extracted as an important indicator, and although 
some variation in accuracy was observed, a tendency for alcohol consumption to 

affect sleep and activity levels was also confirmed.

--------------------------------------------------------------------------------

BACKGROUND

--------------------------------------------------------------------------------

Conventionally, the presence or absence and amount of alcohol consumption have been understood through self-reporting via questionnaires. However, retrospective responses such as "how much did you drink" and "how often do you drink" have challenges including memory bias and social bias. Against this background, this research aimed to develop a new approach that automatically detects alcohol consumption using data such as heart rate, sleep, and activity levels obtained from wearable devices. By establishing detection technology using objective biometric data, we aim to more accurately grasp drinking behavior and appropriately evaluate its health effects.

--------------------------------------------------------------------------------

RESEARCH OVERVIEW

--------------------------------------------------------------------------------

Research Period: May 1, 2022 - September 10, 2024
Research Subjects: 21 individuals aged 20 and above who provided consent

Evaluation Indicators:

• Subjective evaluation indicators:

  - Presence or absence of alcohol consumption the previous day
  - Amount of alcohol consumed (none / 1-2 units / 3-4 units / 5+ units)

• Wearable device data

Analysis Method:

• Using alcohol consumption information from daily questionnaires as correct labels, a machine learning model (XGBoost) was constructed to predict the  presence or absence of alcohol consumption using heart rate, sleep, and activity  level data obtained from wearable devices

• In addition to evaluating the model's prediction accuracy, major features were  analyzed to extract biometric indicators contributing to alcohol consumption estimation

• Accuracy differences (fairness) by gender, age, and amount of alcohol consumed were also examined to evaluate the model's scope of application and challenges

• The effects of alcohol consumption on sleep indicators and activity levels the  next day were analyzed to verify the relationship between alcohol consumption and daily life data

--------------------------------------------------------------------------------

RESEARCH RESULTS

--------------------------------------------------------------------------------

In this research, when the presence or absence of alcohol consumption was identified using biometric data obtained from wearable devices, high accuracy of AUC 0.92, approximately 90%, was confirmed. Changes in heart rate during sleep and heart rate elevation trends observed during non-walking periods were extracted as important indicators contributing to estimation, while cases where identification accuracy decreased depending on gender, age, and amount of alcohol consumed were also observed, indicating the need for further model improvement.

Additionally, while tendencies for changes in sleep patterns and activity levels after alcohol consumption were confirmed, it was also revealed that the degree of these changes varies greatly among individuals.

--------------------------------------------------------------------------------

SIGNIFICANCE AND FUTURE PROSPECTS

--------------------------------------------------------------------------------

These results provide initial findings demonstrating the possibility of objectively and accurately detecting alcohol consumption using consumer-grade wearable devices. This is expected to enable wide-ranging applications in the medical and health fields, including early understanding and prediction of health risks associated with alcohol consumption, preventive interventions, and individualized healthcare support based on individual drinking habits.

Additionally, in research on wearable data not directly related to alcohol consumption, application possibilities expand, such as improving analysis accuracy by removing the effects and noise of alcohol consumption on biometric data.

In the future, we plan to further collect data and improve algorithms to develop this into technology that diverse people can use with confidence. We are widely recruiting research collaborators and development partners who empathize with this research theme and will promote technology verification and applied research together.

The analysis technology developed in this research is patent pending as of December 2025.

--------------------------------------------------------------------------------

REFERENCE INFORMATION

--------------------------------------------------------------------------------

Conference Name: 32nd Annual Meeting of the Japanese Society of Behavioral Medicine
Conference Dates: December 6-7, 2025 (Saturday-Sunday)
Venue: Sagamihara City Industrial Hall
Official Website: https://pub.confit.atlas.jp/ja/event/jsbm32

--------------------------------------------------------------------------------

ABOUT TECHDOCTOR INC.

--------------------------------------------------------------------------------

TechDoctor Inc. operates under the vision "Towards an era where data improves well-being," developing and promoting social implementation of "digital biomarkers**" that derive health-related insights from daily sensing data collected by wearable devices and other means. We aim to realize AI-driven medicine based on data in collaboration with medical, pharmaceutical, food-related companies, and research institutions.

CEO: Kazumichi Minato
Headquarters: Kyobashi Ed-Grand 4F, 2-2-1 Kyobashi, Chuo-ku, Tokyo
Founded: June 21, 2019
Business: Development and operation of digital biomarker development platform 
"SelfBase," provision of digital medical solutions
URL: https://www.technology-doctor.com/

** About Digital Biomarkers

Digital biomarkers are indicators for continuously and objectively evaluating the presence or absence of disease, changes in disease status, and treatment effects based on daily biometric data obtained from smartphones, wearable devices, and other sources.

While traditional biomarkers were primarily "point data" measured temporarily at medical institutions, digital biomarkers are characterized by the ability to continuously acquire "line data" from daily life. Based on indicators such as exercise, sleep, and heart rate, they are expected to be used for early disease detection, treatment monitoring, and as new endpoints in drug development. Development has been progressing overseas since around 2019, and attention is also increasing in Japan.