CS328 Mobile Health Sensing and Analytics · Companion Visualizations

Sensor Data Smoothing and Filtering

Time-domain smoothing

• Smoothing Filters by Hand
• Time-Domain Smoothing Filters
• Window Size vs Step Peaks

From time to frequency

• Building a Signal from Sinusoids
• Aliasing and Nyquist
• Signals and Noise in Frequency
• Time vs Frequency Domain Smoothing

Frequency-domain filtering

• Filter Shapes
• lfilter versus filtfilt

Denoising pipelines

• Denoising an ECG Recording
• Denoising a PPG Recording
• Denoising a Speech Recording
• Denoising a Walking Signal

Counting Steps and Reps

Counting steps

• What an Accelerometer Sees During Walking
• Collapsing Three Axes into One
• Filtering Out the Noise
• Finding Peaks
• Trying It on Real Data

Counting exercise repetitions

• Acceleration vs Rotation-Dominant Exercises
• What a Wrist Sensor Sees During Exercise
• The Counting Pipeline
• Why One Setting Cannot Count Every Exercise
• Which Sensor Counts Best Depends on the Exercise
• The Same Exercise Looks Different Across People

The Activity Recognition Pipeline

Sensor data

• Traces of Six Activities
• Activity Signals by Sensor and Placement
• Band-Pass Pre-processing

Windows and features

• Segmenting into Windows
• From Windows to a Feature Table
• The Pipeline Stage by Stage

Feature Extraction

Time and frequency domain

• Sliding-Window Time Features
• Spectrum and Dominant Frequency

Sequence similarity

• Dynamic Time Warping

Decision Trees

Choosing a split

• Information Gain on Cardiac Data
• Threshold Search on Features

Growing and using a tree

• Growing a Decision Tree
• Tree Depth and Overfitting
• Walking the Cardiac Risk Tree

Beyond Decision Trees

Ensembles

• Majority Voting
• Bagging and Bootstrap Samples
• Random Forests
• Boosting

Tuning and Evaluating Classifiers

Validation

• Holdout and Cross-Validation
• Tuning with GridSearchCV

Metrics

• Confusion Matrix and Metrics

Case Studies in Activity Recognition

Activity recognition

• Activity Windows and Features
• Subject Split and Tuning

Exercise recognition

• Feature Families for Exercises
• Sensors, Direction and 39 Classes

Voice-Based Health Analytics

Representing speech

• Framing and the Spectrogram
• The MFCC Pipeline

Voice features

• Pitch and Intensity Contours
• Jitter, Shimmer and HNR

Sensing Heart Rhythm with PPG

From PPG to heart rate

• Band-Pass Filtering the PPG
• Peaks to Heart Rate

Motion and variability

• Peak Detection through Motion
• Heart Rate Variability

Measuring Sleep Patterns

A night of data

• Epochs and the Hypnogram
• Overnight Motion and Heart Rate

Actigraphy

• Per-Epoch Motion Features
• Wake/Sleep Classifier