7.9.15. ECG Classification

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

Example folder: examples/ecg_classification (GitHub)

7.9.15.1. Overview

Cardiovascular diseases are the leading cause of death globally (about 17.9 million deaths per year). Continuous ECG monitoring helps early detection, but traditional analysis needs expensive equipment and trained professionals. This example shows how a CNN running on the MSPM0G5187 NPU can classify ECG signals at the point of care without cloud connectivity.

Application: wearable health monitors, portable cardiac monitoring
Primary device: MSPM0G5187 (with NPU); also AM13E2 and F28P55 (anomaly detection)

Targets of the reference solution: real-time on-device classification, low power consumption, more than 95% classification accuracy, and LED feedback of the result.

7.9.15.2. System Components

Hardware: MSPM0G5187 and the AFE1594 analog front end for ECG acquisition
Software: Code Composer Studio 12.x or later, MSPM0 SDK 2.08.00 or later

7.9.15.3. Available Configurations

Config file	Device	Model	Mode and dataset
`config_MSPM0.yaml`	MSPM0G5187	`ECG_55k_NPU`	`task_type: ecg_classification`, dataset `ecg_classification_4class`, preset `ECG2500Input_Roundoff_1Frame`
`config.yaml`	AM13E2	`CLS_4k_NPU`	`generic_timeseries_classification`, dataset `ecg` (`ecg_classification_2class.zip`), 128-sample frames, 50 epochs
`config_anomaly_detection.yaml`	F28P55	`AD_Linear`	`generic_timeseries_anomalydetection`, dataset `ecg_ad` (`ecg_ad_dataset.zip`), 140-sample frames, 200 epochs

7.9.15.4. Running the Example

Install Tiny ML ModelZoo first (Installation, Quickstart), then run from the tinyml-modelzoo folder:

# MSPM0G5187 (primary target)
./run_tinyml_modelzoo.sh examples/ecg_classification/config_MSPM0.yaml
# AM13E2
./run_tinyml_modelzoo.sh examples/ecg_classification/config.yaml
# Anomaly detection
./run_tinyml_modelzoo.sh examples/ecg_classification/config_anomaly_detection.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\ecg_classification\config_MSPM0.yaml

7.9.15.5. Dataset (MSPM0 config)

ecg_classification_4class contains labeled ECG recordings. The example readme.md lists these categories:

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

7.9.15.6. Feature Extraction

The preset ECG2500Input_Roundoff_1Frame takes 2500 ECG samples per frame, normalizes the signal by rounding off, and feeds a single frame to the model.

7.9.15.7. Model and Training (MSPM0 config)

ECG_55k_NPU is a CNN with about 55,000 parameters designed for the MSPM0G5187 NPU.

common:
  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

Running the config performs feature extraction, training, INT8 quantization and compilation with the TI Neural Network Compiler, producing the model library and interface headers for the MSPM0 SDK (see Deploying a ModelZoo Run to an SDK Example (Golden-Vector Check)).

7.9.15.8. Results

The example readme.md states an accuracy of about 97% for the MSPM0 configuration. In the automated 1.5.0 config test run:

Config	Test accuracy	AUC ROC
`config_MSPM0.yaml` (`ECG_55k_NPU`)	97.19%	0.995
`config.yaml` (AM13E2, `CLS_4k_NPU`)	97.21%	0.995
`config_anomaly_detection.yaml` (`AD_Linear`)	F1 score 98.02%	not applicable

7.9.15.9. Anomaly Detection Mode

config_anomaly_detection.yaml trains an autoencoder on normal ECG recordings only and flags anomalies by reconstruction error. This is useful when labeled abnormal data is not available. It compiles with compile_preset_name: forced_soft_npu_preset.