Abnormal Exercise Detection

1. Data Collection

I have extracted data from the Health app on my iPhone for this analysis to identify data of abnormal exercise. Data ranges from 2018-2025.

Below is the list of available features I had to select from. I ended up choosing Steps (HKQuantityTypeIdentifierStepCount), Distanace (HKQuantityTypeIdentifierDistanceWalkingRunning), Speed (HKQuantityTypeIdentifierWalkingSpeed), and Flights Climbed (HKQuantityTypeIdentifierFlightsClimbed).

                    ['HKCategoryTypeIdentifierAppleStandHour',
                    'HKQuantityTypeIdentifierHeartRateVariabilitySDNN',
                    'HKQuantityTypeIdentifierActiveEnergyBurned',
                    'HKQuantityTypeIdentifierWalkingStepLength',
                    'HKQuantityTypeIdentifierWalkingSpeed',
                    'HKQuantityTypeIdentifierRestingHeartRate',
                    'HKQuantityTypeIdentifierFlightsClimbed',
                    'HKQuantityTypeIdentifierHeight', 
                    'HKQuantityTypeIdentifierWalkingAsymmetryPercentage',
                    'HKQuantityTypeIdentifierBodyMass',
                    'HKQuantityTypeIdentifierAppleExerciseTime',
                    'HKQuantityTypeIdentifierWalkingHeartRateAverage',
                    'HKQuantityTypeIdentifierWalkingDoubleSupportPercentage',
                    'HKQuantityTypeIdentifierBasalEnergyBurned',
                    'HKQuantityTypeIdentifierHeadphoneAudioExposure',
                    'HKQuantityTypeIdentifierStepCount',
                    'HKQuantityTypeIdentifierHeartRate',
                    'HKQuantityTypeIdentifierAppleWalkingSteadiness',
                    'HKQuantityTypeIdentifiMerDistanceWalkingRunning',
                    'HKCategoryTypeIdentifierHighHeartRateEvent']
                

2. Preprocessing

Data was aggregated to daily values using preprocess.py. Typically my first approach for finding anomalies in my data would be to find the mean and standard deviation, compute z-scores, and flag if the value is 3 standard deviations away from the mean. However this only works if the data is normally distributed. Let's first check distribution.

Histogram of Steps

Histogram of Distance

Histogram of Speed (mph)

Histogram of Flights Climbed

3. Data Adjustments

Looking at this, it seems as if Apple devices did not start tracking speed until late 2020, so records prior to that year were excluded to maintain consistency across all features.

Steps Over Time

Distance Over Time

Speed Over Time

Flights Over TIme

4. Model Training

Trained an Isolation Forest with a 5% contamination rate using model.py. All functionality (data I/O, model training, inference) was modularized in separate scripts (utils.py, preprocess.py, model.py, inference.py).

Here displays all datapoints and their anomaly scores:

Anomaly Scores

5. Results

The model flagged several anomalous days. Here are the anomalies plotted over time for each feature used.

Step Anomalies

Distance Anomalies

Speed Anomalies

Flights Anomalies

Here are scatterplots allowing us to see which features are influencing the anomaly detection. High and low ends for steps typically delineate an anomaly as well as low values for speed and high values for flights climbed.

Steps vs Distance

Steps vs Speed

Steps vs Flights

Speed vs Flights

As a final sanity check, I wanted to see if my recent trip to Korea would flag any of the days as anomalies as I knew I did a lot of walking on the trip. As seen below 7/8 of the days were identified as an anomalous exercise day.

Korea Trip Exercise