Skip to content

Wandering Detection Model

Wandering is a clear indicator of other issues, such as dementia or Alzheimer's disease. It can also be a sign of depression or anxiety. Detecting wandering is one of the primary purposes the wearable is designed for. An accurate model is one of the flagship features of the Golden Age Hub.

Foundation Checklist

Problem Definition

  • Clearly defined the specific problem to solve
  • Detect potential wandering behavior in seniors using location data
  • Distinguish between normal movement and potential wandering episodes
  • Integrate with alert throttling system to minimize caregiver fatigue

  • Respect privacy through minimal data collection and processing

  • Identified the type of ML task

  • Geofence-based event detection
  • Time-series analysis of movement patterns
  • Anomaly detection for unusual movement

  • (Future) Sequence modeling for trajectory prediction

  • Established clear success criteria

  • 85% precision in distinguishing wandering from normal movement

  • <15% false positive rate during normal activities
  • <5 minute detection latency for non-critical wandering
  • <60 second detection for boundary violations

  • 80% caregiver satisfaction with alert relevance

  • <5% battery impact from location services

  • Determined if ML is actually the right approach

  • Rule-based geofencing insufficient for complex movement patterns
  • ML can learn individual movement baselines
  • Hybrid approach (rules + ML) balances reliability and adaptability
  • Privacy-preserving techniques available for sensitive location data

Data and Assumptions

  • Inventoried all available data sources
  • Primary: Device location (GPS when available, else network-based)
  • Secondary: Geofence boundaries from caregiver app
  • Optional (future): Step count, activity type from motion sensors

  • Caregiver feedback on alert accuracy

  • Assessed data quantity and quality

  • Initial: Synthetic datasets for development
  • Phase 1: Partner data under IRB/DUA/HIPAA compliance
  • Continuous: Real-world data collection with user consent

  • Data augmentation for rare scenarios

  • Documented key assumptions

  • Users will have mobile devices with location services
  • Caregivers will set up appropriate geofences
  • Most wandering follows recognizable patterns

  • Privacy is a primary concern for users and caregivers

  • Identified potential biases

  • Limited data from advanced dementia cases
  • Potential underrepresentation of night-time wandering
  • Cultural differences in movement patterns
  • Variability in GPS accuracy across environments

Performance and Ethics

  • Selected appropriate evaluation metrics
  • Primary: Precision (minimize false alerts)
  • Secondary: Recall (detect actual wandering events)
  • Tertiary: Time-to-detection (critical for safety)

  • Additional: Battery usage, storage requirements

  • Established deployment thresholds

  • Precision: ≥85% (minimize false alerts)
  • Recall: ≥80% (detect wandering episodes)
  • Detection latency: <1 minute for critical events
  • Model size: <5MB for edge deployment

  • Inference time: <500ms on target hardware

  • Ethical considerations

  • Risk: Privacy concerns with location tracking
    • Mitigation: On-device processing, data minimization
  • Risk: False positives causing unnecessary concern
    • Mitigation: Multi-factor confirmation before alerts

  • Risk: Over-reliance on technology

    • Mitigation: Clear guidelines for human oversight

  • Fairness and bias mitigation

  • Tested across different mobility levels
  • Validated in various living environments
  • Considered cultural differences in movement

  • Regular bias audits with diverse test groups

  • Interpretability and explainability

  • Clear documentation of model decisions
  • Confidence scores with explanations
  • Caregiver-facing dashboards with model insights
  • Regular model cards updates

Personalized Model Training

Data Collection Strategy

  • Passive Data Collection
  • Continuous location tracking with privacy-preserving techniques
  • Movement patterns during different times of day
  • Routine establishment (frequently visited places, typical routes)

  • Battery-optimized location sampling (adaptive based on movement)

  • Active Data Collection

  • Caregiver-verified "safe zones" and "restricted areas"
  • User confirmation of intended destinations
  • Integration with calendar/appointment data
  • Family member check-ins for ground truth

Labeling Strategy

  • Automated Initial Labeling
  • Geofence-based pre-labeling
  • Anomaly detection in movement patterns
  • Time-series analysis of location history

  • Integration with activity recognition

  • Human-in-the-Loop Verification

  • Caregiver verification of wandering alerts
  • Simple mobile interface for confirming/denying wandering events
  • Contextual questions about the purpose of movement

  • Optional photo confirmation at destinations

  • Quality Control

  • Multi-modal verification (location + activity + time)
  • Confidence scoring for automated labels
  • Periodic review of false positives/negatives
  • Feedback loop for model improvement

Model Personalization

  • Initial Model
  • Pretrained on population-level wandering patterns
  • Common routes and destinations by time/location

  • General movement patterns for different mobility levels

  • Personalization Process

  • Learning Period (14-30 days)

    • Establish baseline movement patterns
    • Learn typical daily routines
    • Identify regular destinations and routes

  • Adaptation Phase

    • Adjust sensitivity based on individual patterns
    • Learn user's typical walking speed and range
    • Update model with verified wandering events

  • Continuous Learning

    • Adapt to gradual routine changes
    • Seasonal pattern adjustments
    • Health condition-aware updates

Addressing Challenges

  • Privacy-Preserving Techniques
  • On-device processing of location data
  • Differential privacy for location sharing
  • Geohashing for approximate locations

  • Local storage of sensitive location history

  • User-Friendly Verification

  • Simple one-tap verification
  • Voice-based confirmation ("I'm going for my usual walk")
  • Automatic check-ins at common destinations

  • Family-assisted verification for non-technical users

  • Handling Special Cases

  • Travel/vacation mode
  • Hospital stays or temporary relocations
  • Changes in mobility aids
  • Cognitive decline progression

Maintenance Planning

  • Performance monitoring
  • Real-time monitoring of prediction confidence
  • Weekly performance reports (precision, recall, F1)
  • User feedback collection and analysis

  • Alert system for performance degradation

  • Concept drift handling

  • Monthly analysis of feature distributions
  • Adaptive retraining based on performance metrics
  • Canary deployment for model updates

  • A/B testing framework for new features

  • Retraining strategy

  • Scheduled: Quarterly full retraining
  • Event-based: When new patterns emerge
  • Performance-based: When metrics degrade

  • Data-based: With significant new labeled data

  • Team responsibilities

  • Defined roles and responsibilities for model maintenance
  • ML Engineers: Model updates and optimization
  • Data Scientists: Pattern analysis and improvement
  • Caregivers: Ground truth verification

  • Product Team: User experience integration

  • Documented procedures for model updates and versioning

  • Semantic versioning for releases
  • Rollback procedures for critical issues
  • Change management documentation
  • Compliance with healthcare regulations