Roni Bandini — Buenos Aires, Argentina Project started November 2024 · inference sketch January 2025

HR-HRV is an experimental wearable Machine Learning project that classifies worker stress from ECG signals directly on an Arduino Portenta H7.

The signal is acquired with a SparkFun AD8232 Single Lead Heart Rate Monitor and processed using the Edge Impulse HR/HRV Features block.

The resulting physiological features feed a neural-network classifier with two output classes:

stress
working

Inference runs locally on the Portenta H7. The onboard RGB LEDs provide immediate visual feedback.

Experimental project for exploring ECG, HR/HRV feature extraction and Edge Machine Learning. It is not intended as a medical diagnostic device.

• ❤️ Single-lead ECG acquisition
• 🧠 On-device stress classification
• 📈 Edge Impulse HR/HRV DSP block
• ⏱️ 40-second inference window
• 📊 50 Hz ECG acquisition
• 🔬 HR and interbeat-interval feature extraction
• 🚨 85% stress confidence threshold
• 🔴 Red LED for detected stress
• 🟢 Green LED for normal working state
• 📦 Compiled Arduino HR library included
• 🔌 Leads-off electrode detection
• 🧪 Alternative Elemyo ECG implementation included

flowchart LR
    BODY["❤️ ECG Electrodes"]
    AD["AD8232 ECG Front End"]
    PORTENTA["Arduino Portenta H7"]
    HRV["Edge Impulse<br/>HR / HRV DSP"]
    NN["🧠 Neural Network"]
    CLASS{"Classification"}
    RED["🔴 Stress"]
    GREEN["🟢 Working"]

    BODY --> AD
    AD -->|"A0"| PORTENTA
    PORTENTA --> HRV
    HRV --> NN
    NN --> CLASS
    CLASS -->|"stress > 85%"| RED
    CLASS -->|"otherwise"| GREEN

The complete inference pipeline executes on the Portenta H7 without requiring remote inference.

The project uses the Arduino Portenta H7.

Its Cortex-M7 provides enough processing power to run the HR/HRV feature extraction and classifier locally.

The original version uses the:

👉 SparkFun Single Lead Heart Rate Monitor — AD8232

The AD8232 amplifies and filters the small electrical signals captured by three ECG electrodes.

The project uses:

OUTPUT → A0
LO-    → D2
LO+    → D3

The LO+ and LO- outputs detect disconnected electrodes.

Electrode placement used in the original experiment:

Yellow → Left
Red    → Right
Green  → Below red

Complete reference:

👉 Wearable for Monitoring Worker Stress — Edge Impulse

Training acquisition sketch:

👉 acquisition.ino

The sketch defines:

#define FREQUENCY_HZ 50

and streams the analog ECG value:

Serial.println(
    analogRead(A0)
);

only while both leads-off inputs indicate a valid electrode connection.

Serial speed:

Serial.begin(115200);

The onboard LEDs also indicate electrode state.

After uploading acquisition.ino, run:

edge-impulse-data-forwarder

The expected acquisition frequency is:

50 Hz

The original workflow collected approximately:

10–20 samples per class
120 seconds per sample
10% reserved for testing

Typical classes:

working
stress

Official walkthrough:

👉 ECG HR/HRV Portenta Tutorial

The project uses the Edge Impulse HR/HRV Features block.

HR is the number of heartbeats per minute.

HRV describes variations between consecutive beats — the interbeat interval / RR interval.

ECG waveform
     ↓
R peaks
     ↓
Interbeat intervals
     ↓
HR / HRV features
     ↓
Machine Learning

The Edge Impulse block can produce features including:

HR Mean
HR Slope
IBI Slope
RMSSD
RMSSD Slope
AVNN
SDNN
Range NN
MAD NN
pNN50

VLF Energy
LF Energy
HF Energy
Total Energy
Relative VLF / LF / HF
LF/HF Ratio
Peak VLF / LF / HF

The configuration used by this project is intentionally compact:

Signal: ECG
Filter preset: 1
Window: 40 seconds
HRV feature group: none

The HR-derived features are then passed to the classifier.

Public project:

👉 Heart Rate Analysis — Edge Impulse Project #558080

Current impulse:

Current reported Portenta H7 performance for the float32 model:

The custom Keras architecture is compact:

20 input features
        ↓
Dense 20 / ReLU
        ↓
Dense 10 / ReLU
        ↓
2-class Softmax

Training configuration used for the published model:

Training cycles: 40
Learning rate:   0.005
Batch size:      30
Auto weighting:  Disabled

The resulting validation accuracy was:

89.3%

flowchart LR
    ECG["❤️ ECG @ 50 Hz"]
    HRV["HR / HRV Block"]
    FEATURES["20 Features"]
    NN["Dense Neural Network"]
    STRESS["stress"]
    WORK["working"]

    ECG --> HRV
    HRV --> FEATURES
    FEATURES --> NN
    NN --> STRESS
    NN --> WORK

The repository contains the compiled Arduino library:

👉 heart-rate-analysis-hr-library-portenta-compiled.zip

The original Edge Impulse deployment procedure is:

1. Export the model as an Arduino library.
2. Unzip it locally.
3. Download: edgeimpulse/example-hr-lib-arm
4. Copy those support files into the deployment folder.
5. Zip the complete folder again.
6. Install it through:

Arduino IDE
→ Sketch
→ Include Library
→ Add .ZIP Library

The resulting library exposes:

#include <arduino-hrv.h>

Current Edge Impulse documentation notes that HR/HRV block evaluation is available broadly, while hardware deployment requires Enterprise enablement.

Main application:

👉 ecg2.ino

The model expects:

float features[2000];

corresponding to:

40 seconds × 50 Hz = 2000 ECG samples

The sketch continuously fills this buffer from:

analogRead(A0);

and passes it to:

run_classifier(
    &features_signal,
    &result,
    false
);

The current source uses:

float threesold = 0.85;

A stress event is detected when:

result.classification[ix].value > threesold

and:

result.classification[ix].label == "stress"

Therefore:

stress confidence > 85%
          ↓
       STRESS

The Portenta onboard LEDs are used as the local interface.

digitalWrite(LEDR, LOW);
digitalWrite(LEDG, HIGH);

Result:

🔴 Red ON
🟢 Green OFF

digitalWrite(LEDG, LOW);
digitalWrite(LEDR, HIGH);

Result:

🟢 Green ON
🔴 Red OFF

No external screen is required.

flowchart TD
    ECG["❤️ Read ECG"]
    LEADS{"Electrodes connected?"}
    BUFFER["Store in 2000-sample buffer"]
    HRV["HR / HRV Processing"]
    ML["🧠 Classification"]
    SCORE{"stress > 0.85?"}
    RED["🔴 Red LED"]
    GREEN["🟢 Green LED"]

    ECG --> LEADS
    LEADS -->|"Yes"| BUFFER
    LEADS -->|"No"| ECG
    BUFFER --> HRV
    HRV --> ML
    ML --> SCORE
    SCORE -->|"Yes"| RED
    SCORE -->|"No"| GREEN
    RED --> ECG
    GREEN --> ECG

Machine Learning platform:

👉 Edge Impulse

👉 Arduino IDE

From Boards Manager install:

Arduino Mbed OS Portenta Boards

Select:

Arduino Portenta H7 (M7 core)

Hardware documentation:

👉 Arduino Portenta H7

git clone https://github.com/ronibandini/hr-hrv.git
cd hr-hrv

Repository:

👉 github.com/ronibandini/hr-hrv

Add:

👉 heart-rate-analysis-hr-library-portenta-compiled.zip

through:

Sketch
→ Include Library
→ Add .ZIP Library

For data collection:

👉 acquisition.ino

For stress inference:

👉 ecg2.ino

Serial Monitor:

115200 baud

Startup:

Stress detection started
Roni Bandini, Jan 2025

The repository also contains a later adaptation for an Elemyo MYO-series EMG/ECG module:

📂 elemyo/

Files:

• elemyo/acquisition3.ino
• elemyo/ecg3.ino
• elemyo/readme.txt

The acquisition version uses:

#define CSpin 7
#define sensorInPin A1

and applies notch filters around:

50 Hz
100 Hz

through the Elemyo library before forwarding ECG samples.

The adaptation was added in May 2025.

hr-hrv/
│
├── acquisition.ino
├── ecg2.ino
├── heart-rate-analysis-hr-library-portenta-compiled.zip
│
├── elemyo/
│   ├── acquisition3.ino
│   ├── ecg3.ino
│   └── readme.txt
│
├── README.md
└── LICENSE

• 📊 acquisition.ino — AD8232 data forwarding
• ❤️ ecg2.ino — on-device stress classifier
• 📦 Compiled HR library — Portenta deployment
• 🧪 elemyo/ — alternate sensor implementation
• ⚖️ LICENSE — MIT License

Complete official tutorial for this project:

👉 Wearable for Monitoring Worker Stress using HR/HRV DSP Block — Arduino Portenta

Dataset, impulse, HR/HRV processing block, classifier and model testing:

👉 Heart Rate Analysis — Project #558080

Full list of HR and HRV time-domain and frequency-domain features:

👉 HR/HRV Features Block

Technical background:

👉 Getting the Pulse: Technical Insights Into Edge Impulse's HR and HRV Blocks

Official Portenta H7 documentation:

👉 Arduino Portenta H7

AD8232 ECG sensor hookup, electrode placement and example waveforms:

👉 AD8232 Heart Rate Monitor Hookup Guide

Support library required by the original HR/HRV Arduino deployment workflow:

👉 edgeimpulse/example-hr-lib-arm

TinyML worker fall detection with Arduino Nano 33 BLE Sense, Edge Impulse, BLE and Raspberry Pi.

👉 github.com/ronibandini/BTFall

TinyML shoulder-movement recognition using accelerometer data and Edge Impulse.

👉 github.com/ronibandini/rotatorCuffRecovery

Contactless heart-rate and respiration monitoring with mmWave radar, UNIHIKER and ESP32-C6.

👉 github.com/ronibandini/heartRespirationMonitor

Motion-pattern classification using XIAO nRF52840 Sense and Edge Impulse.

👉 github.com/ronibandini/XIAOnRF52840TireCare

Contracultura Maker is a book by Roni Bandini about maker culture, experimental electronics, AI, physical computing and technological autonomy.

📂 Contracultura Maker — GitHub repository

📕 Download Contracultura Maker PDF

Roni Bandini Maker · AI Developer · Writer Buenos Aires, Argentina

• 🐙 GitHub — @ronibandini
• 🌐 Medium — @ronibandini
• 𝕏 X / Twitter — @RoniBandini
• 📸 Instagram — @ronibandini
• ▶️ YouTube — @RoniBandini
• 💼 LinkedIn — Roni Bandini

Built with ❤️ + ECG + HR/HRV + Edge Machine Learning.