Wearable PPG Heart Rate & HRV (Synthetic Data)

Overview

This notebook is a compact, end-to-end demo of a wearable heart-rate pipeline using synthetic data. It creates a realistic photoplethysmogram (PPG) signal like what a smartwatch optical sensor records, then processes it to produce heart rate and heart rate variability (HRV) metrics with clear plots.

What the code does

1. Generate a synthetic PPG signal

   • Builds a timeline at 100 Hz for 5 minutes.
   • Creates beat times from a target mean heart rate with small random variability by sampling RR intervals and taking a cumulative sum.
   • For each beat, adds a pulse kernel with a fast rise and slow decay to mimic the systolic upstroke and diastolic falloff.
   • Adds realistic artifacts:
     • Respiratory baseline wander at about 0.25 Hz.
     • Low frequency drift and occasional spikes to simulate motion.
     • White noise to simulate sensor noise.
   • The result is a raw PPG that looks and behaves like real wearable data.

2. Preprocess the signal

   • Detrends to remove slow drift.
   • Applies a 0.5 to 5 Hz Butterworth bandpass filter to isolate the cardiac component of the waveform and suppress baseline and high-frequency noise.
   • Shows a side-by-side plot of raw versus filtered so you can see the improvement.

3. Detect beats and compute heart rate

   • Uses scipy.signal.find_peaks with a minimum distance and prominence threshold so it does not double count or chase noise.
   • Converts the time between consecutive detected peaks to instantaneous heart rate in beats per minute.
   • Plots heart rate across the whole session to verify it looks stable and plausible.

4. Compute HRV metrics

   • Time domain:
     • SDNN is the standard deviation of RR intervals in seconds.
     • RMSSD is the root mean square of successive differences of RR intervals.
     • Both are commonly reported in research and consumer products.
   • Frequency domain:
     • Interpolates RR intervals to an even grid at 4 Hz and runs Welch power spectral density.
     • Integrates power in standard bands:
       • LF 0.04 to 0.15 Hz
       • HF 0.15 to 0.40 Hz
     • Reports LF, HF, and LF to HF ratio.
   • Nonlinear view:
     • Plots a Poincaré scatter of RR_n against RR_{n+1} to visualize rhythm stability.

5. Export results and artifacts

   • Saves ppg_timeseries.csv with time, raw PPG, and filtered PPG columns.
   • Saves ppg_metrics.json with mean heart rate, SDNN, and RMSSD for quick inspection and for inclusion in the repo.

Why synthetic

• No privacy risk and no dataset hunting.
• Reproducible on any machine, including Colab, with standard scientific Python libraries.
• Still realistic enough to demonstrate signal processing, detection, and HRV workflows used in wearables and research.

What the plots show

• Raw vs Filtered PPG: effect of preprocessing
• Detected Beats: correctness of the peak detector
• Heart Rate over Time: the metric users and clinicians care about
• HRV Power Spectral Density: autonomic balance features in LF and HF bands
• Poincaré Plot: quick visual of variability and rhythm stability

Limitations to be aware of

• The data are simulated and will not include all edge cases like severe motion, poor perfusion, or arrhythmias.
• Peak detection thresholds are tuned for this synthetic signal and may need adjustment for real sensors.
• Frequency domain HRV is sensitive to interpolation choices and windowing. The implementation follows common defaults but is simplified for clarity.

Tech stack

• Python, NumPy, SciPy, Matplotlib, Pandas. Runs out of the box in Google Colab.

---

Features

• Signal Simulation: Generates a 5-minute PPG signal with physiological artifacts (respiration drift, motion noise, white noise).
• Preprocessing: Detrending + 0.5–5 Hz Butterworth bandpass filter to isolate pulse frequency band.
• Beat Detection: Peak detection with prominence/distance criteria to find heartbeats.
• Heart Rate Calculation: Instantaneous HR over time derived from RR intervals.
• HRV Metrics:
  • Time-domain: SDNN, RMSSD.
  • Frequency-domain: LF/HF power ratio via Welch PSD.
  • Poincaré plot for RR interval variability.
• Visualizations:
  1. Raw vs. Filtered PPG (10 s).
  2. Detected Beats (10 s window).
  3. Heart Rate over Time.
  4. HRV Power Spectral Density.
  5. Poincaré Plot.
• Outputs:
  • ppg_timeseries.csv — raw & filtered signals.
  • ppg_metrics.json — mean HR, SDNN, RMSSD.

---

Plots

1. Raw vs. Filtered PPG (10 s)

This plot compares the unprocessed (raw) simulated PPG signal to the bandpass-filtered version over a 10-second window.

• The raw signal includes baseline wander from respiration, low-frequency drift from motion, occasional spikes, and white noise — all common in real wearable sensor data.
• The filtered signal has these artifacts largely removed, isolating the pulsatile waveform corresponding to heartbeats.
• This step demonstrates the importance of preprocessing in biomedical signal analysis — without it, heartbeat detection would be highly unreliable.

2. Detected Beats (10 s)

Displays the filtered PPG from the same time window as above, but with markers placed on each detected systolic peak.

• These peaks represent the moments when blood volume in the tissue is at its maximum during each cardiac cycle.
• The peak detection algorithm uses both minimum distance (to avoid double-counting beats) and prominence thresholds (to avoid detecting noise).
• This plot visually validates that the detection logic is working and is an essential check before calculating HR or HRV.

3. Heart Rate over Time

Plots the instantaneous heart rate (in beats per minute) calculated from the inverse of RR intervals (time between consecutive detected peaks).

• This is essentially what a wearable device’s heart rate display would show over the course of the 5-minute simulation.
• A stable HR line indicates consistent beat detection, while sharp drops or spikes may reveal signal disturbances or physiological variability.
• This plot connects the raw signal processing to a metric that is both clinically and commercially relevant.

4. HRV Power Spectral Density

A frequency-domain representation of heart rate variability, calculated using Welch’s method on the evenly resampled RR interval series.

• The power spectrum is divided into:
  • Low Frequency (LF): 0.04–0.15 Hz, associated with both sympathetic and parasympathetic modulation.
  • High Frequency (HF): 0.15–0.40 Hz, associated with parasympathetic (vagal) activity and respiratory sinus arrhythmia.
• The LF/HF ratio is often interpreted as a measure of autonomic balance.
• This plot is critical in HRV research and is widely used in sports science, stress monitoring, and medical diagnostics.

5. Poincaré Plot

A scatter plot of each RR interval (RRₙ) against the next consecutive RR interval (RRₙ₊₁).

• Produces an ellipse-like shape when the rhythm is steady, with wider dispersion indicating greater variability.
• Tight, narrow clouds suggest low HRV (often seen in stress or certain disease states), while broader, more dispersed clouds indicate high HRV (often seen in well-conditioned or relaxed states).
• This plot provides a quick, intuitive way to assess rhythm stability and variability without delving into statistical calculations.

---

Requirements

• Python 3.8+
• numpy
• scipy
• pandas
• matplotlib

All are pre-installed in Google Colab.

---

Use Cases

• Demonstrating biomedical signal processing skills.
• Showcasing end-to-end wearable device data analysis.
• Teaching HR/HRV analysis concepts.

---

License

MIT License — see LICENSE for details.