📊 Chapter 3: Digital Biomarkers and Monitoring

Understanding Digital Biomarkers

Digital biomarkers represent a paradigm shift in mental health monitoring. Unlike traditional biomarkers requiring laboratory tests or clinical visits, digital biomarkers are objective, quantifiable behavioral and physiological data collected passively through smartphones, wearables, and other connected devices. They enable continuous, real-world monitoring of mental health status, providing insights impossible to capture through periodic clinical assessments.

The FDA defines digital health technologies (DHTs) including digital biomarkers as "objective, quantifiable physiological and behavioral data that are collected and measured by means of digital devices." For depression detection, these biomarkers capture the behavioral and physiological manifestations of depressive states in everyday life.

Passive vs. Active Digital Biomarkers

Digital biomarkers fall into two categories based on collection methodology:

Passive Biomarkers

Collected automatically without user action, passive biomarkers provide continuous, objective data representing natural behavior. Examples include:

• Physical activity levels from accelerometers
• Sleep patterns from wearable sensors
• Communication metadata (call/text frequency, not content)
• Location and movement patterns from GPS
• Screen time and app usage patterns
• Heart rate and HRV from wearables

Advantages: No burden on users, captures natural behavior, continuous data, high compliance, minimal recall bias

Challenges: Privacy concerns, requires device integration, battery consumption, data interpretation complexity

Active Biomarkers

Require user engagement such as completing assessments or tasks. Examples include:

• Ecological Momentary Assessment (EMA) surveys
• Digital versions of clinical scales (PHQ-9, GAD-7)
• Cognitive tests and reaction time tasks
• Voice recordings for analysis
• Selfie-based mood tracking

Advantages: Direct symptom assessment, standardized clinical measures, user awareness and engagement

Challenges: Compliance burden, survey fatigue, response biases, not truly continuous

Smartphone-Based Behavioral Monitoring

Smartphones have become ubiquitous personal sensing platforms, with over 6.8 billion smartphone users globally. Their rich sensor arrays—accelerometers, GPS, microphones, cameras, and usage logs—enable comprehensive behavioral phenotyping for depression detection.

Activity and Movement Patterns

Physical inactivity is both a symptom and risk factor for depression. Smartphone accelerometers provide continuous activity monitoring, capturing:

• Daily Step Count: Depressed individuals average 2,500-4,000 fewer steps per day compared to healthy controls
• Activity Variability: Reduced day-to-day variation in activity levels correlates with depression severity (r = -0.48)
• Sedentary Time: Prolonged periods without movement (>2 hours) increase in depression
• Circadian Activity Patterns: Disrupted or delayed activity peaks indicate circadian rhythm dysfunction

class ActivityBiomarkerExtractor:
    """Extract depression-relevant features from accelerometer data"""
    
    def __init__(self, sampling_rate=50):
        self.sampling_rate = sampling_rate
        
    def calculate_daily_features(self, accel_data, timestamps):
        """
        Calculate daily activity features from raw accelerometer data
        
        Parameters:
        - accel_data: numpy array of shape (n_samples, 3) for x, y, z axes
        - timestamps: datetime array aligned with accel_data
        
        Returns:
        - Dictionary of activity features
        """
        import numpy as np
        from scipy import signal
        
        # Calculate movement magnitude
        magnitude = np.sqrt(np.sum(accel_data**2, axis=1))
        
        # Step detection using peak finding
        steps = self._detect_steps(magnitude)
        
        # Activity intensity levels
        sedentary_threshold = 0.2  # g-force threshold
        light_threshold = 0.5
        moderate_threshold = 1.0
        
        sedentary_time = np.sum(magnitude < sedentary_threshold) / self.sampling_rate / 60  # minutes
        light_activity = np.sum((magnitude >= sedentary_threshold) & (magnitude < light_threshold))
        moderate_activity = np.sum((magnitude >= light_threshold) & (magnitude < moderate_threshold))
        vigorous_activity = np.sum(magnitude >= moderate_threshold)
        
        # Circadian activity pattern
        hourly_activity = self._calculate_hourly_activity(magnitude, timestamps)
        peak_activity_hour = np.argmax(hourly_activity)
        
        return {
            'total_steps': len(steps),
            'sedentary_minutes': sedentary_time,
            'light_activity_minutes': light_activity / self.sampling_rate / 60,
            'moderate_activity_minutes': moderate_activity / self.sampling_rate / 60,
            'vigorous_activity_minutes': vigorous_activity / self.sampling_rate / 60,
            'peak_activity_hour': peak_activity_hour,
            'activity_entropy': self._calculate_entropy(hourly_activity),
            'movement_variability': np.std(magnitude)
        }
    
    def _detect_steps(self, magnitude):
        """Detect steps using peak detection algorithm"""
        from scipy.signal import find_peaks
        
        # Filter signal for step detection
        filtered = self._bandpass_filter(magnitude, 0.5, 3.0)
        
        # Find peaks with minimum prominence and distance
        peaks, _ = find_peaks(
            filtered,
            prominence=0.1,
            distance=int(0.3 * self.sampling_rate)  # Minimum 0.3s between steps
        )
        
        return peaks
    
    def calculate_weekly_patterns(self, daily_features):
        """
        Calculate weekly-level features indicating depression risk
        """
        steps_per_day = [day['total_steps'] for day in daily_features]
        
        return {
            'mean_daily_steps': np.mean(steps_per_day),
            'step_variability': np.std(steps_per_day),
            'low_activity_days': sum(1 for s in steps_per_day if s < 2000),
            'activity_trend': self._calculate_trend(steps_per_day),  # Declining = negative
            'mean_sedentary_hours': np.mean([day['sedentary_minutes']/60 for day in daily_features])
        }
        

Sleep Monitoring

Sleep disturbances affect 80-90% of individuals with depression. Smartphone-based sleep monitoring uses movement sensors and screen activity to estimate:

Wearable Device Integration

Wearable devices like smartwatches and fitness trackers provide physiological monitoring capabilities beyond smartphone sensors. Modern wearables track heart rate, heart rate variability, blood oxygen, skin temperature, and detailed sleep architecture.

Heart Rate Variability (HRV)

HRV measures variation in time intervals between heartbeats, reflecting autonomic nervous system function. Depression is associated with reduced HRV, indicating decreased parasympathetic (rest-and-digest) activity and increased sympathetic (fight-or-flight) activity.

Advanced Sleep Monitoring

Wearable devices with optical heart rate sensors can estimate sleep stages (light, deep, REM) using heart rate and movement patterns. Depression shows characteristic sleep architecture changes:

• Reduced REM Latency: Time to first REM period shortens from ~90 minutes to <60 minutes
• Increased REM Density: More rapid eye movements during REM sleep
• Reduced Slow-Wave Sleep: Less time in deep, restorative sleep stages
• Early Morning Awakening: Terminal insomnia with inability to return to sleep

Social Interaction Digital Phenotyping

Social withdrawal is a core feature of depression. Smartphone communication metadata (not content, preserving privacy) reveals social interaction patterns:

• Call Frequency: Depressed individuals make 40-60% fewer outgoing calls
• Text Messaging: Reduced texting frequency and shorter message lengths
• Conversation Duration: Shorter call durations, particularly for social (non-work) calls
• Unique Contacts: Decreased number of distinct communication partners
• Response Latency: Longer delays in responding to messages during depressive episodes
• Communication Reciprocity: Reduced bidirectional communication (more incoming than outgoing)

Location and Mobility Patterns

GPS and location data reveal mobility patterns that correlate with depression. While privacy-sensitive, when collected with informed consent, location entropy provides valuable insights:

Location Entropy and Diversity

class LocationBiomarkers:
    """Extract depression-relevant features from GPS location data"""
    
    def calculate_location_entropy(self, gps_points):
        """
        Calculate Shannon entropy of visited locations
        Lower entropy = less location diversity = potential depression indicator
        """
        import numpy as np
        from sklearn.cluster import DBSCAN
        
        # Cluster GPS points to identify distinct locations
        clustering = DBSCAN(eps=0.1, min_samples=5)  # ~100m radius clusters
        location_clusters = clustering.fit_predict(gps_points)
        
        # Calculate location visit frequency
        unique_locations, visit_counts = np.unique(location_clusters, return_counts=True)
        probabilities = visit_counts / np.sum(visit_counts)
        
        # Shannon entropy
        entropy = -np.sum(probabilities * np.log2(probabilities + 1e-10))
        
        return {
            'location_entropy': entropy,
            'unique_locations': len(unique_locations),
            'primary_location_percentage': np.max(probabilities) * 100,
            'location_diversity_score': len(unique_locations) / len(gps_points)
        }
    
    def calculate_home_time(self, gps_points, timestamps):
        """Calculate percentage of time spent at home location"""
        # Identify home location (most frequent location during 10pm-6am)
        night_points = [p for p, t in zip(gps_points, timestamps) 
                       if 22 <= t.hour or t.hour <= 6]
        home_location = self._identify_cluster_center(night_points)
        
        # Calculate time at home
        at_home = [self._distance(p, home_location) < 0.1 for p in gps_points]  # 100m radius
        home_time_percentage = sum(at_home) / len(at_home) * 100
        
        return home_time_percentage
        

Research findings on location patterns and depression:

• Location entropy decreases by 35-50% during depressive episodes
• Time spent at home increases from ~50% to >75% in moderate-severe depression
• Reduced variability in daily routes and destinations
• Decreased distance traveled per day (from ~15km to <5km average)
• Loss of regular routine locations (gym, social venues, restaurants)

Multimodal Biomarker Integration

The most accurate depression detection systems integrate multiple biomarker types, leveraging their complementary strengths. Single biomarkers typically achieve AUC of 0.65-0.75, while multimodal approaches reach AUC >0.85-0.90.