Skip to content
heartKIT
User guide
HELIA

Train ECG Arrhythmia Classifier

Download notebook View source Open in Colab

This example displays saved outputs. Building the documentation does not run training. Check dataset paths for your notebook working directory before running.

Date created: 2024/07/17

Last Modified: 2024/07/17

Description: Train, evaluate, and export 4-stage ECG arrhythmia classifier

Overview

Section titled “Overview”

In this guide, we will train an ECG-based arrhythmia classifier that uses an EfficientNetV2 inspired model. The classifier is trained on raw ECG data and is able to discern normal sinus rhythm (NSR), sinus bradycardia (SBRAD), atrial fibrillation (AFIB), and general supraventricular tachycardia (GSVT).

Input

  • Sensor: ECG
  • Location: Wrist
  • Sampling Rate: 100 Hz
  • Frame Size: 5 seconds

Class Mapping

Identify rhythm into one of four categories: SR, SBRAD, AFIB, GSVT.

Base ClassTarget ClassLabel
0-SR0Sinus Rhythm (SR)
1-SBRAD1Sinus Bradycardia (SBRAD)
7-AFIB, 8-AFL2AFIB/AFL (AFIB)
2-STACH, 5-SVT3General supraventricular tachycardia (GSVT)

Datasets

  • LSAD: The Large Scale Rhythm Database (LSAD) is a large publicly available electrocardiography dataset. It contains 10 second, 12-lead ECGs of 45,152 patients with a 500 Hz sampling rate. The ECGs are sampled at 500 Hz and are annotated by up to two cardiologists.

Setup

Section titled “Setup”
import os


os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
import IPython
import contextlib
from pathlib import Path
import tempfile
import keras
import heartkit as hk
import helia_edge as helia
# Be sure to set the dataset path to the correct location
datasets_dir = Path(os.getenv("HK_DATASET_PATH", "./datasets"))


plot_theme = hk.utils.dark_theme
helia.utils.silence_tensorflow()
hk.utils.setup_plotting(plot_theme)
logger = helia.utils.setup_logger(__name__)

Configure datasets

Section titled “Configure datasets”

We are going to train our model using the Large scale Arrhythmia dataset. This dataset uses the slug lsad within heartKIT. We will download the dataset if it is not already available.

datasets = [
    hk.NamedParams(
        name="lsad",
        params=dict(
            path=datasets_dir / "lsad",
        ),
    )
]

Target classes

Section titled “Target classes”

For this task, we are going to classify ECG signals into one of four classes:

  • Sinus Rhytm: Normal ECG signal (SR)
  • Sinus Bradycardia: Slow heart rate (SB)
  • Atrial Flutter/Fibrillation: Irregular heart rate (AFIB)
  • General Supra-Ventricular Tachycardia: Fast heart rate (GSVT)

heartKIT already provides a number of heart rhythms. We will provide a class mapping for the four classes we are interested in. We will also provide class names for display purposes.

class_map = {
    hk.tasks.HKRhythm.sr: 0,
    hk.tasks.HKRhythm.sbrad: 1,
    hk.tasks.HKRhythm.afib: 2,
    hk.tasks.HKRhythm.aflut: 2,
    hk.tasks.HKRhythm.stach: 3,
    hk.tasks.HKRhythm.svt: 3,
}


class_names = ["SR", "SB", "AFIB", "GSVT"]

Preprocess pipeline

Section titled “Preprocess pipeline”

We will preprocess the ECG signals by applying the following steps:

  • Apply Z-score normalization w/ epsilon to avoid division by zero

The task accepts a list of preprocessing functions that will be applied to the input data.

preprocesses = [hk.NamedParams(name="layer_norm", params=dict(epsilon=0.01, name="znorm"))]

Define EfficientNetV2 model architecture

Section titled “Define EfficientNetV2 model architecture”

For this task, we are going to leverage a customized EfficientNetV2 model architecture that is smaller and can handle 1D signals. The model consists of 6 MBConv style blocks with a depth of 2. Each block leverages squeeze-and-excitation mechanism w/ ratio of 4 to improve the model’s performance.

architecture = hk.NamedParams(
    name="efficientnetv2",
    params=dict(
        input_filters=24,
        input_kernel_size=(1, 9),
        input_strides=(1, 2),
        blocks=[
            dict(filters=24, depth=1, kernel_size=(1, 9), strides=(1, 2), ex_ratio=1, se_ratio=4),
            dict(filters=32, depth=1, kernel_size=(1, 9), strides=(1, 2), ex_ratio=1, se_ratio=4),
            dict(filters=48, depth=1, kernel_size=(1, 9), strides=(1, 2), ex_ratio=1, se_ratio=4),
            dict(filters=64, depth=1, kernel_size=(1, 9), strides=(1, 2), ex_ratio=1, se_ratio=4),
            dict(filters=80, depth=1, kernel_size=(1, 9), strides=(1, 2), ex_ratio=1, se_ratio=4),
            dict(filters=96, depth=1, kernel_size=(1, 9), strides=(1, 2), ex_ratio=1, se_ratio=4),
        ],
        output_filters=0,
        include_top=True,
        use_logits=True,
    ),
)

Task configuration

Section titled “Task configuration”

Here we provide the complete configuration for the task. This includes the dataset configuration, preprocessing pipeline, model architecture, and training parameters.

params = hk.HKTaskParams(
    # Common arguments
    name="hk-4-stage-rhythm",
    job_dir=Path(tempfile.gettempdir()) / "hk-4-stage-rhythm",
    # Dataset arguments
    datasets=datasets,
    # Signal arguments
    sampling_rate=100,
    frame_size=800,
    # Dataloader arguments
    samples_per_patient=5,
    val_samples_per_patient=5,
    test_samples_per_patient=5,
    # Preprocessing/Augmentation arguments
    preprocesses=preprocesses,
    # Class arguments
    num_classes=len(class_names),
    class_map=class_map,
    class_names=class_names,
    # Split arguments
    val_patients=0.2,
    val_size=20000,
    test_size=20000,
    val_file="val.tfds",
    test_file="val.tfds",
    # Model arguments
    model_file="model.keras",
    architecture=architecture,
    # Training parameters
    lr_rate=1e-3,
    lr_cycles=1,
    batch_size=256,
    buffer_size=25000,
    epochs=100,
    steps_per_epoch=50,
    val_metric="loss",
    class_weights="balanced",
    # Evaluation arguments
    threshold=0.5,
    test_metric_threshold=0.98,
    # Export parameters
    tflm_var_name="ecg_rhythm_flatbuffer",
    tflm_file="ecg_rhythm_flatbuffer.h",
    # Demo params
    backend="pc",
    demo_size=800,
    display_report=False,
    # Extra arguments
    verbose=1,
    seed=42,
)

Load rhythm task

Section titled “Load rhythm task”

heartKIT provides a TaskFactory that includes a number ready-to-use tasks. Each task provides methods for training, evaluating, exporting, and demoing. We will grab the rhythm task and configure it for our use case.

task = hk.TaskFactory.get("rhythm")

Download datasets

Section titled “Download datasets”

We will download the datasets needed for the task.

task.download(params=params)

Visualize the model

Section titled “Visualize the model”

Lets quickly instantiate and visualize the first few layers of the model.

model = helia.models.efficientnet.efficientnetv2_from_object(
    x=keras.Input(shape=(params.frame_size, 1), name="inputs"), params=architecture.params, num_classes=len(class_names)
)
model.summary(layer_range=("inputs", model.layers[10].name))
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1723841711.256117  789182 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723841711.276561  789182 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723841711.276664  789182 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723841711.278058  789182 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723841711.278130  789182 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723841711.278176  789182 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723841711.327697  789182 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723841711.327778  789182 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
I0000 00:00:1723841711.327833  789182 cuda_executor.cc:1015] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero. See more at https://github.com/torvalds/linux/blob/v6.0/Documentation/ABI/testing/sysfs-bus-pci#L344-L355
Model: "EfficientNetV2"
Saved output 
┏━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━┓
┃ Layer (type)        ┃ Output Shape      ┃    Param # ┃ Connected to      ┃
┡━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━┩
│ inputs (InputLayer) │ (None, 800, 1)    │          0 │ -                 │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ reshape (Reshape)   │ (None, 1, 800, 1) │          0 │ inputs[0][0]      │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stem.conv (Conv2D)  │ (None, 1, 400,    │        216 │ reshape[0][0]     │
│                     │ 24)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stem.bn             │ (None, 1, 400,    │         96 │ stem.conv[0][0]   │
│ (BatchNormalizatio… │ 24)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stem.act            │ (None, 1, 400,    │          0 │ stem.bn[0][0]     │
│ (Activation)        │ 24)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.dp   │ (None, 1, 400,    │        216 │ stem.act[0][0]    │
│ (DepthwiseConv2D)   │ 24)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.dp.… │ (None, 1, 400,    │         96 │ stage1.mbconv1.d… │
│ (BatchNormalizatio… │ 24)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.dp.… │ (None, 1, 400,    │          0 │ stage1.mbconv1.d… │
│ (Activation)        │ 24)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ max_pooling2d       │ (None, 1, 200,    │          0 │ stage1.mbconv1.d… │
│ (MaxPooling2D)      │ 24)               │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.se.… │ (None, 1, 1, 24)  │          0 │ max_pooling2d[0]… │
│ (GlobalAveragePool… │                   │            │                   │
├─────────────────────┼───────────────────┼────────────┼───────────────────┤
│ stage1.mbconv1.se.… │ (None, 1, 1, 6)   │        150 │ stage1.mbconv1.s… │
│ (Conv2D)            │                   │            │                   │
└─────────────────────┴───────────────────┴────────────┴───────────────────┘
 Total params: 32,192 (125.75 KB)
 Trainable params: 30,912 (120.75 KB)
 Non-trainable params: 1,280 (5.00 KB)

Train the model

Section titled “Train the model”

Now let’s train the model using the LSAD dataset. We will train the model for 100 epochs.

task.train(params)
Sorting lsad labels: 100%|██████████| 36120/36120 [00:07<00:00, 4746.99it/s]
Sorting lsad labels: 100%|██████████| 36120/36120 [00:07<00:00, 4552.33it/s]
INFO     Validation steps per epoch: 78                                                             datasets.py:105
INFO     Saving validation dataset to /tmp/hk-4-stage-rhythm/val.tfds                                   train.py:57
Training:   0%|           0/100 ETA: ?s,  ?epochs/sWARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1723841770.965635  789335 service.cc:146] XLA service 0x79d8f400b520 initialized for platform CUDA (this does not guarantee that XLA will be used). Devices:
I0000 00:00:1723841770.965656  789335 service.cc:154]   StreamExecutor device (0): NVIDIA GeForce RTX 4090, Compute Capability 8.9
I0000 00:00:1723841777.014004  789335 device_compiler.h:188] Compiled cluster using XLA!  This line is logged at most once for the lifetime of the process.
Training: 100%|██████████ 100/100 ETA: 00:00s,   1.94s/epochs
78/78 ━━━━━━━━━━━━━━━━━━━━ 1s 806us/step
78/78 ━━━━━━━━━━━━━━━━━━━━ 0s 915us/step - acc: 0.9442 - f1: 0.9444 - loss: 0.0903
INFO     [VAL SET] ACC=0.9444, F1=0.9445, LOSS=0.0920                                                  train.py:190

Saved figure from cell 24

Model evaluation

Section titled “Model evaluation”

Now that we have trained the model, we will evaluate the model on the test dataset. Similar to training, we will provide the same task parameters.

task.evaluate(params)
INFO     Loading validation dataset from /tmp/hk-4-stage-rhythm/val.tfds                             evaluate.py:33
78/78 ━━━━━━━━━━━━━━━━━━━━ 1s 2ms/step - acc: 0.9442 - f1: 0.9444 - loss: 0.0903
INFO     [TEST SET] ACC=0.9444, F1=0.9445, LOSS=0.0920                                               evaluate.py:50
624/624 ━━━━━━━━━━━━━━━━━━━━ 2s 2ms/step
613/613 ━━━━━━━━━━━━━━━━━━━━ 2s 1ms/step - acc: 0.9518 - f1: 0.9520 - loss: 0.0804
INFO     [TEST SET] THRESH=50.00%, DROP=1.76%                                                        evaluate.py:62
INFO     [TEST SET] ACC=0.9520, F1=0.9521, LOSS=0.0822                                               evaluate.py:63

Confusion matrix

Section titled “Confusion matrix”

Let’s visualize the confusion matrix to understand the model’s performance on each class.

IPython.display.Image(filename=params.job_dir / "confusion_matrix_test.png", width=500)

Saved figure from cell 28

Export model to TF Lite / TFLM

Section titled “Export model to TF Lite / TFLM”

Once we have trained and evaluated the model, we need to export the model into a format that can be used for inference on the edge. Currently, we export the model to TensorFlow Lite flatbuffer format. This will also generate a C header file that can be used with TensorFlow Lite for Microcontrollers (TFLM).

Apply post-training quantization (PTQ)

Section titled “Apply post-training quantization (PTQ)”

For running on bare metal, we will perform post-training quantization to convert the model to an 8-bit integer model. The weights and activations will be quantized to 8-bits and biases will be quantized to 32-bits. This will reduce the model size and improve the inference speed.

quantization = hk.QuantizationParams(
    enabled=True,
    format="INT8",
    io_type="int8",
    conversion="KERAS",
)
params.quantization = quantization
# TF dumps a lot of info to stdout, so we redirect it to /dev/null
with open(os.devnull, "w") as devnull:
    with contextlib.redirect_stdout(devnull), contextlib.redirect_stderr(devnull):
        task.export(params=params)
W0000 00:00:1723841958.715237  789182 tf_tfl_flatbuffer_helpers.cc:392] Ignored output_format.
W0000 00:00:1723841958.715249  789182 tf_tfl_flatbuffer_helpers.cc:395] Ignored drop_control_dependency.
INFO     Validating model results                                                                      export.py:94
fully_quantize: 0, inference_type: 6, input_inference_type: INT8, output_inference_type: INT8
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
INFO     [TF METRICS] LOSS=0.2360 ACC=0.9657 F1=0.9761                                                export.py:101
INFO     [TFL METRICS] LOSS=0.2441 ACC=0.9639 F1=0.9752                                               export.py:102
INFO     Validation passed (0.0080)                                                                   export.py:110

Run inference demo

Section titled “Run inference demo”

We will run a demo on the PC to verify that the model is working as expected. The demo will load the model and run inferences across a randomly selected ECG signal. The demo will also provide the model’s prediction and the corresponding class name.

task.demo(params=params)
Inference: 100%|██████████| 1/1 [00:00<00:00,  1.86it/s]

Saved figure from cell 33