7.7.14. ECG Classification

Edge AI solution for real-time ECG signal classification on the MSPM0G5187 with NPU. Classifies ECG signals into cardiac conditions for portable, low-power heart monitoring.

7.7.14.1. Overview

Cardiovascular diseases (CVDs) are the leading cause of death globally, responsible for approximately 17.9 million deaths per year. Early detection through continuous ECG monitoring significantly improves patient outcomes. However, traditional ECG analysis requires expensive medical equipment and trained professionals, limiting access to timely diagnosis.

This example demonstrates how Edge AI enables real-time cardiac classification on portable, low-power devices – without the need for cloud connectivity. By deploying a CNN-based model directly on the MSPM0G5187 microcontroller with NPU acceleration, ECG signals can be classified into multiple cardiac conditions at the point of care.

Application: Wearable health monitors, portable cardiac monitoring devices
Task Type: Time Series Classification
Primary Device: MSPM0G5187 (with NPU)
Also Supported: AM13E2, F28P55

Note

Device-specific configuration files are provided as config_<device>.yaml. See the Available Configurations section below for details.

7.7.14.2. Key Performance Targets

Real-time ECG classification on-device
Low power consumption suitable for wearables
Greater than 95% classification accuracy
LED feedback for classification results

7.7.14.3. System Components

Hardware

MSPM0G5187 – Microcontroller with integrated NPU
AFE1594 Analog Front End – ECG signal acquisition

Software

Code Composer Studio (CCS) 12.x or later
MSPM0 SDK 2.08.00 or later
TI Edge AI Studio

7.7.14.4. Dataset Details

Dataset: ecg_classification_4class

The dataset contains ECG recordings labeled into the following classes:

Class	Description
Normal	Healthy sinus rhythm
Mild	Minor cardiac abnormalities
Other	Other cardiac conditions

7.7.14.5. Feature Extraction

The feature extraction pipeline for MSPM0 (preset ECG2500Input_Roundoff_1Frame) processes the raw ECG signal as follows:

Sampling: 2500 samples per frame from the ECG input
Normalization: Signal normalization by rounding off
Framing: Single frame input to the model

data_processing_feature_extraction:
  feature_extraction_name: ECG2500Input_Roundoff_1Frame

7.7.14.6. Model

ECG_55k_NPU (Default)

A CNN architecture optimized for the MSPM0G5187 NPU, with approximately 55K parameters. This model is designed for NPU-compatible inference, enabling efficient real-time classification on the target device.

Model	Parameters	Description
`ECG_55k_NPU`	~55,000	CNN optimized for NPU, default for MSPM0

Expected Accuracy: ~97%

7.7.14.7. Training Configuration

The default training hyperparameters for the MSPM0 configuration:

Parameter	Value
Batch size	12
Learning rate	0.001
Optimizer	Adam
Weight decay	4e-5
Epochs	25

Quantization: INT8 quantization is enabled by default for deployment.

Compilation: Uses the TI Neural Network Compiler (NNC), optimized for the MSPM0G5187 NPU.

7.7.14.8. Configuration (MSPM0)

common:
  target_module: 'timeseries'
  task_type: 'ecg_classification'
  target_device: 'MSPM0G5187'

dataset:
  dataset_name: ecg_classification_4class

data_processing_feature_extraction:
  feature_extraction_name: ECG2500Input_Roundoff_1Frame

training:
  model_name: ECG_55k_NPU
  batch_size: 12
  learning_rate: 0.001
  weight_decay: 4e-5
  optimizer: 'adam'
  training_epochs: 25
  num_gpus: 0

testing:
  enable: True

compilation:
  enable: True

7.7.14.9. Running the Example

# MSPM0G5187 (primary target)
cd tinyml-modelzoo
./run_tinyml_modelzoo.sh examples/ecg_classification/config_MSPM0.yaml

# AM13E2
./run_tinyml_modelzoo.sh examples/ecg_classification/config.yaml

# Anomaly detection mode
./run_tinyml_modelzoo.sh examples/ecg_classification/config_anomaly_detection.yaml

cd tinyml-modelzoo

REM MSPM0G5187 (primary target)
.\run_tinyml_modelzoo.bat examples\ecg_classification\config_MSPM0.yaml

REM AM13E2
.\run_tinyml_modelzoo.bat examples\ecg_classification\config.yaml

REM Anomaly detection mode
.\run_tinyml_modelzoo.bat examples\ecg_classification\config_anomaly_detection.yaml

7.7.14.10. Available Configurations

Config File	Target Device	Mode
`config_MSPM0.yaml`	MSPM0G5187	Classification (4-class)
`config.yaml`	AM13E2	Classification
`config_anomaly_detection.yaml`	F28P55	Anomaly Detection

7.7.14.11. Anomaly Detection Mode

A separate anomaly detection configuration is available for ECG monitoring using the config_anomaly_detection.yaml file. This mode trains on normal ECG recordings only and detects anomalies based on reconstruction error, which can be useful when labeled fault data is not available.

common:
  task_type: 'generic_timeseries_anomalydetection'
  target_device: 'F28P55'

dataset:
  dataset_name: 'ecg_ad'

training:
  model_name: 'AD_Linear'
  training_epochs: 200

7.7.14.12. References

7.7.14.13. See Also

Anomaly Detection Example – Anomaly detection tutorial
Example on GitHub