S-ADL: Exploring Smartphone-based Activities of Daily Living to Detect Blood Alcohol Concentration in a Controlled Environment

Document Title

S-ADL: Exploring Smartphone-based Activities of Daily Living to Detect Blood Alcohol Concentration in a Controlled Environment

Document Information

• Subject Area: Human-Computer Interaction, Smartphone Function Assessment, Alcohol Concentration Detection
• Keywords: Alcohol Detection, Activities of Daily Living, Smartphone Applications, Functional Assessment, Machine Learning

Research Background and Problem

• Identified Issues or Challenges:

  • Alcohol abuse is prevalent among young populations, leading to health and social issues.
  • Current methods for detecting blood alcohol concentration (BAC) (e.g., self-reports, sensor monitoring, or breath tests) suffer from issues such as insufficient accuracy, high user burden, or lack of timeliness.
  • While some studies have attempted to use smartphones for functional assessment, they lack sensitivity to complex cognitive functions, particularly in detecting subtle effects of alcohol.

• Why This Problem Is Important:

  • Accurate BAC detection is crucial for preventing alcohol-related incidents (e.g., drunk driving), which is significant for public health and safety.
  • Providing a more convenient and real-time tool can help young individuals monitor alcohol intake anytime, reducing health risks.

• Research Motivation and Related Work:

  • Existing smartphone-based methods relying on motion sensors or simple reaction experiments are not sensitive enough to detect mild alcohol effects.
  • Smartphones have become indispensable tools in daily life, and people's everyday usage behaviors may reflect their cognitive functional states, especially changes in activity efficiency under alcohol influence.
  • Based on this, the authors propose combining "smartphone-based activities of daily living" (S-ADL) with machine learning models for alcohol detection.

Solution

• Proposed Method or Solution:

  • Develop and validate a new S-ADL method for BAC detection.
  • S-ADL is based on scenario-based design and functional assessment of daily interaction behaviors with common smartphone applications (e.g., replying to messages, web searching, financial transactions).
  • Extract performance metrics from smartphone interaction data and train machine learning models to classify different BAC levels.

• Innovative Aspects of the Solution:

  • Extends the traditional concept of activities of daily living (ADL) to smartphone interaction behaviors.
  • Proposes automated evaluation metrics that reflect changes in cognitive function (e.g., task completion time, typing speed).
  • The S-ADL method improves measurement accuracy by using tasks from commonly used smartphone applications, avoiding learning effects.

• Implementation Steps and Key Techniques:

  1. Define S-ADL task categories (e.g., communication, information searching, financial management).
  2. Design and test scenario-based S-ADL task scripts.
  3. Conduct a controlled experimental study with 40 healthy young participants, performing BAC-segmented experiments (BAC levels: 0%, 0.03%-0.04%, 0.07%-0.08%).
  4. Train classifiers based on S-ADL performance metrics using machine learning models (e.g., Random Forest, Gradient Boosting Machines).
  5. Analyze the contribution of key performance metrics to the model using SHAP values.

Research Findings

• Specific Achievements:

  • Machine learning models trained on S-ADL performance metrics effectively detected BAC, achieving an average AUC-ROC and accuracy of approximately 81% in binary classification tasks.
  • Compared to traditional computerized neuropsychological tests (CNT), the S-ADL method demonstrated better performance, higher sensitivity, and less influence from learning effects.
  • Using only two tasks (information searching and message replying), 80% detection accuracy could be achieved within one minute or less.

• Advantages Over Existing Solutions:

  • More convenient: Designed based on everyday smartphone applications, requiring no additional devices or complex setups.
  • More real-time: Capable of rapid BAC detection without hours of waiting or additional data input.
  • More sensitive: Particularly effective in detecting mild alcohol consumption (BAC 0.03%-0.04%).

• Experimental or Evaluation Results:

  • Key extracted metrics include task completion time, typing speed, etc.
  • Information searching and message replying were identified as the best task combination for predicting BAC.
  • Model robustness was evaluated using leave-one-out cross-validation.

• Limitations and Future Directions:

  • Limitations:

    • The experiment was conducted in a controlled environment and only involved young, healthy individuals, which may limit its generalizability to real-world scenarios or other populations.
    • Differences between smartphone operating systems (e.g., Android and iOS) require further validation.
    • Privacy protection issues need to be addressed with enhanced technical measures.

  • Future Directions:

    • Expand the participant pool to include diverse ages, ethnicities, and health conditions to validate the model's general applicability.
    • Explore the use of other smart device sensors or IoT technologies to extend BAC detection capabilities, such as home IoT or vehicle sensors.
    • Address background noise in real-world scenarios (e.g., distracting notifications, weather factors) and reduce detection errors using multi-source data.
    • Develop functional monitoring systems for alcohol dependence and mental health issues, leveraging S-ADL to assess long-term risks.