Workout Data Report

Overview

This report documents the workout data available from Garmin fitness exports, specifically focusing on strength training data for use in Bayesian time-series modeling. The data is loaded via the src.data.workout module, which processes Garmin's summarized activities JSON files.

Key Statistics (based on current dataset): - Total workouts: 139 strength training sessions - Date range: 2024-03-06 to 2026-01-12 (678 days) - Workout frequency: 0.21 workouts/day (approximately 1.5 workouts per week) - Most common workout day: Wednesday - Volume data quality: 86.3% of workouts have zero volume recorded - Total reps: 23,897 | Total sets: 2,164 - Primary metrics: Reps, sets, volume, duration, calories

Data Fields

Core Workout Metadata

Field	Type	Description	Example
`activity_id`	int64	Unique Garmin activity identifier	14288027894
`activity_type`	string	Type of activity (e.g., "strength_training")	"strength_training"
`name`	string	Workout name (user-defined)	"Morning Workout"
`start_time_gmt`	datetime	Start time in GMT (millisecond precision)	2023-01-15 14:30:00
`start_time_local`	datetime	Start time in local timezone	2023-01-15 09:30:00
`date`	datetime	Date extracted from local start time (day-level precision)	2023-01-15

Fitness Metrics

Field	Type	Description	Notes
`duration`	float64	Workout duration in seconds	3600 = 1 hour
`calories`	float64	Estimated calories burned	Likely most reliable magnitude metric - Garmin estimation based on HR, duration, user profile; consistently recorded
`avg_hr`	float64	Average heart rate during workout	Optional, may be null
`max_hr`	float64	Maximum heart rate during workout	Optional, may be null
`min_hr`	float64	Minimum heart rate during workout	Optional, may be null
`steps`	float64	Steps recorded during workout	May be zero for strength training

Strength-Specific Metrics

Field	Type	Description	Notes
`total_reps`	int64	Total repetitions across all exercises	Potential counting errors - may not capture effort intensity
`total_sets`	int64	Total sets across all exercises	Potential counting errors - may not capture effort intensity
`active_sets`	int64	Active sets (sets with recorded reps)	May equal total_sets
`total_volume`	int64	Total weight lifted (volume) in arbitrary units	Data quality issue: Many zero values

Exercise Details

Field	Type	Description	Structure
`exercise_details`	list[dict]	Detailed breakdown of individual exercises	List of exercise records

Exercise Record Fields: - category: string - Exercise category (e.g., "BENCH_PRESS", "UNKNOWN") - sub_category: string or None - More specific classification - reps: int - Number of repetitions for this exercise - volume: int - Weight lifted for this exercise (may be zero) - duration: float - Duration of this exercise in seconds - sets: int - Number of sets for this exercise - max_weight: int or None - Maximum weight used

Data Quality Assessment

Missing Values

min_hr: 79 out of 139 records missing (56.8%) - expected for strength training without continuous HR monitoring
Other heart rate metrics (avg_hr, max_hr): Some missing but fewer than min_hr
All other core fields have complete data

Volume Data Issues

The total_volume field shows significant data quality problems: - Many workouts have total_volume = 0 despite having total_reps and total_sets - Volume appears to be inconsistently recorded in Garmin exports - Recommendation: Use calories as primary magnitude metric; reps/sets as alternatives with caution

Metric Reliability Trade-offs

Different metrics capture different aspects of workout magnitude with varying reliability:

calories: Most reliable for magnitude - consistently recorded, estimates energy expenditure based on heart rate, duration, and user profile. However, it's an estimation that may have individual errors.
total_reps / total_sets: Subject to counting errors - may not accurately capture effort if reps are miscounted or sets are incomplete. Don't account for weight lifted.
total_volume: Poor data quality - 86.3% of workouts have zero volume recorded.
duration: Doesn't capture intensity - a short intense workout may have greater physiological impact than a long light one.
count: Simple frequency - useful for workout occurrence but not magnitude.

For cross-lagged modeling of workout effects on weight, calories is recommended as the primary magnitude metric due to consistent recording and relevance to energy expenditure.

Exercise Detail Completeness

All workouts include exercise_details list (when include_exercise_details=True)
Some exercises have category: "UNKNOWN" - Garmin may not classify all exercises
volume field within exercise details often zero

Aggregation Options

The prepare_workout_aggregates() function supports multiple aggregation strategies:

Time Aggregation

Daily: Aggregate metrics per calendar day
Weekly: Aggregate metrics per week (starting Monday)

Metric Aggregation

Metric	Column Used	Description	Recommended
`count`	None	Number of workouts per time period	Good for frequency analysis
`volume`	`total_volume`	Total weight lifted	Not recommended due to data quality
`reps`	`total_reps`	Total repetitions	Use with caution - potential counting errors, doesn't capture weight
`sets`	`total_sets`	Total sets	Use with caution - potential counting errors, doesn't capture weight
`calories`	`calories`	Total calories burned	Primary recommendation - most reliable magnitude metric, consistently recorded
`duration`	`duration`	Total workout duration	Useful for time commitment analysis

Handling Zero Days

When aggregating daily, many days have zero workouts. The prepare_workout_for_stan() function can fill missing days with zeros using fill_missing=True.

Usage in Cross-Lagged GP Models

Current Implementation

The cross-lagged Gaussian Process model (weight_gp_crosslagged.stan) uses: - Fixed lag parameter (lag_days) specified as data - Single lagged workout value affects weight at time t - β parameter represents causal effect size - Model tested with lags: 0, 1, 2, 3, 7 days

Data Preparation Pipeline

from src.data.workout import load_strength_training_data
from src.data.align import prepare_crosslagged_stan_data

# Load and aggregate workout data
df_workouts, stan_data_workout = load_strength_training_data(
    aggregation="daily",
    metric="calories"  # Most reliable magnitude metric
)

# Prepare cross-lagged data (requires weight data)
stan_data = prepare_crosslagged_stan_data(
    stan_data_weight,
    stan_data_workout,
    lag_days=2
)

Recommendations for Modeling

Primary metric: Use calories as the most reliable magnitude metric (consistently recorded, estimates energy expenditure)
Alternative metrics: Consider total_reps or total_sets but be aware of potential counting errors
Aggregation: Daily aggregation with zero-fill for missing days
Lag testing: Test multiple lags (0-7 days) using WAIC/LOO comparison
Sparse data: Account for many zero days in workout time series

Temporal Patterns

Workout Frequency

[To be populated from workout_report.py output]
Average workouts per week: [ ]
Most common workout days: [ ]
Monthly trends: [ ]

Exercise Distribution

Top exercise categories by volume: [ ]
Most frequent exercise types: [ ]
Volume trends over time: [ ]

Integration with Other Health Metrics

Workout data can be correlated with: 1. Weight: Muscle gain hypothesis (positive effect after 1-3 day lag) 2. Sleep: Impact of evening workouts on sleep quality 3. Heart Rate: Resting HR changes with training adaptation 4. VO2 Max: Cardiovascular fitness improvements

Limitations and Considerations

Garmin Export Limitations

Volume data incomplete: Many workouts lack weight information
Exercise classification: Some exercises marked "UNKNOWN"
Timing precision: Only start time recorded, not exercise timing within workout
Missing metrics: No RPE (rate of perceived exertion) or intensity measures

Modeling Challenges

Sparse time series: Many days without workouts (zeros)
Metric reliability trade-offs:
calories: Most consistently recorded but estimated (HR-based)
reps/sets: May have counting errors, don't capture weight
volume: Often zero (86.3% of workouts)
duration: Doesn't capture intensity
Variable workout intensity: Reps/sets don't capture weight lifted
Delayed effects: Physiological responses may have distributed lags
Individual variation: Response to training varies by person

Future Enhancements

Data Improvements

Supplement with manual training log data
Estimate volume from exercise categories and typical weights
Add derived metrics (volume × reps, intensity index)

Modeling Extensions

Estimated lag models: Treat lag as estimated parameter rather than fixed
Distributed lag models: Cumulative effects from multiple lag periods
Threshold models: Minimum workout intensity required for effect
Interaction models: Combined effects of workouts with sleep/nutrition

References

src/data/workout.py: Primary data loading module
src/analysis/workout_report.py: Comprehensive analysis and visualization
stan/weight_gp_crosslagged.stan: Cross-lagged GP model
src/data/align.py: Data alignment for cross-lagged modeling

Report generated: 2026-01-31 Data source: Garmin DI-Connect-Fitness export Analysis code: src/analysis/workout_report.py