ECG anomaly detection with the LSTM-AD SageMaker algorithm¶
Detecting anomalies in electrocardiogram (ECG) signals is critical for the diagnosis and treatment of cardiovascular diseases. The introduction of wearable ECG sensors enables long-term continuous remote monitoring of patients' cardiac activity. However, it is unfeasible for cardiologists to manually review the large amount of data generated by real-time ECG sensors. Machine learning algorithms can automate the process of ECG analysis, reducing the need for manual inspection by healthcare providers.
Different supervised and unsupervised machine learning algorithms have been studied in the literature on ECG anomaly detection [1]. In this post, we will focus on the Long Short-Term Memory Network for Anomaly Detection (LSTM-AD) [2], a standard deep learning framework for detecting anomalies in time series data. We will demonstrate how to use our Amazon SageMaker implementation of the LSTM-AD model, the LSTM-AD SageMaker algorithm, for detecting anomalies in an ECG trace.
Model¶
The LSTM-AD model predicts the future values of the time series with a stacked LSTM model. The model parameters are learned on a training set containing only normal data (i.e. without anomalies) by minimizing the Mean Squared Error (MSE) between the actual and predicted values of the time series. After the model has been trained, a Gaussian distribution is fitted to the model’s prediction errors on an independent validation set (also without anomalies) using Maximum Likelihood Estimation (MLE).
At inference time, the model predicts the values of the time series (which can now include anomalies) at each time step, and calculates the likelihood of the model’s prediction errors under the fitted Gaussian distribution. The computed Gaussian likelihood is then used as a normality score: the lower the Gaussian likelihood at a given a time step, the more likely the time step is to be an anomaly.
Note
If enough labelled anomalous data is available, an optimal threshold on the normality score can be determined by maximizing the F-beta score between the actual and predicted anomaly labels.
Data¶
We use dataset number 179 from the Hexagon ML / UCR Time Series Anomaly Detection Archive. The dataset includes a single time series representing a human subject ECG trace sourced from record s30791 in the Long Term ST Database (LTST DB). The length of the time series is 55,000 observations. The first 23,000 observations are included in the training set, while the remaining 32,000 observations are included in the test set. The training set contains only normal data, while the test set contains an anomalous heartbeat between observations 52,600 and 52,800.
Hexagon ML / UCR dataset N°179 (combined training and test sets).
Code¶
Warning
To be able to run the code below, you need to have an active subscription to the LSTM-AD SageMaker algorithm. You can subscribe to a free trial from the AWS Marketplace in order to get your Amazon Resource Name (ARN). In this post we use version 1.11 of the LSTM-FCN SageMaker algorithm, which runs in the PyTorch 2.1.0 Python 3.10 deep learning container.
Environment Set-Up¶
We start by importing all the requirements and setting up the SageMaker environment.
import io
import sagemaker
import pandas as pd
import numpy as np
# SageMaker algorithm ARN, replace the placeholder below with your AWS Marketplace ARN
algo_arn = "arn:aws:sagemaker:<...>"
# SageMaker session
sagemaker_session = sagemaker.Session()
# SageMaker role
role = sagemaker.get_execution_role()
# S3 bucket
bucket = sagemaker_session.default_bucket()
# EC2 instance
instance_type = "ml.m5.2xlarge"
Data Preparation¶
After that we load the dataset and split it into training and test datasets, which we save to S3.
dataset_name = "179_UCR_Anomaly_ltstdbs30791AS_23000_52600_52800"
cutoff = 23000 # train-test cutoff
start = 52600 # start of anomalous time interval
end = 52800 # end of anomalous time interval
# load the dataset
dataset = pd.DataFrame(data=np.genfromtxt(f"{dataset_name}.txt"))
# extract the training dataset
training_dataset = dataset.iloc[:cutoff]
# extract the test dataset
test_dataset = dataset.iloc[cutoff:]
# save the training dataset in S3
training_data = sagemaker_session.upload_string_as_file_body(
body=training_dataset.to_csv(index=False, header=False),
bucket=bucket,
key=f"{dataset_name}_train.csv"
)
# save the test dataset in S3
test_data = sagemaker_session.upload_string_as_file_body(
body=test_dataset.to_csv(index=False, header=False),
bucket=bucket,
key=f"{dataset_name}_test.csv"
)
Training¶
Now that the training dataset is available in an accessible S3 bucket, we are ready to fit the model.
Note
The algorithm uses the first 80% of the training dataset for learning the LSTM parameters, and the remaining 20% of the training dataset for estimating the Gaussian distribution parameters.
# create the estimator
estimator = sagemaker.algorithm.AlgorithmEstimator(
algorithm_arn=algo_arn,
role=role,
instance_count=1,
instance_type=instance_type,
input_mode="File",
sagemaker_session=sagemaker_session,
hyperparameters={
"context-length": 100,
"prediction-length": 10,
"sequence-stride": 10,
"hidden-size": 32,
"num-layers": 2,
"dropout": 0.5,
"lr": 0.001,
"batch-size": 128,
"epochs": 100,
},
)
# run the training job
estimator.fit({"training": training_data})
Inference¶
Once the training job has completed, we can run a batch transform job on the test dataset.
# create the transformer
transformer = estimator.transformer(
instance_count=1,
instance_type=instance_type,
max_payload=100,
)
# run the transform job
transformer.transform(
data=test_data,
content_type="text/csv",
)
The results are saved in an output file in S3 with the same name
as the input file and with the ".out" file extension.
The output file contains the normality scores in the first
column, and the predicted values of the time series in the
subsequent columns.
# load the model outputs from S3
predictions = sagemaker_session.read_s3_file(
bucket=bucket,
key_prefix=f"{transformer.latest_transform_job.name}/{dataset_name}_test.csv.out"
)
# convert the model outputs to data frame
predictions = pd.read_csv(io.StringIO(predictions), header=None, dtype=float)
After loading the normality scores and the predicted values from S3, we can visualize the results.
Note
The algorithm defines the normality scores using the Gaussian log-likelihood instead of the likelihood.
Results on Hexagon ML / UCR dataset №179 (test set).
We find that the model correctly identifies the anomalies, as the normality score exhibits the largest downward spikes on the same time steps where the anomalies are observed.
After the analysis has been completed, we can delete the model.
# delete the model
transformer.delete_model()
References¶
[1] Li, H., & Boulanger, P. (2020). A survey of heart anomaly detection using ambulatory electrocardiogram (ECG). Sensors 2020, 20, 1461. doi: 10.3390/s20051461.
[2] Malhotra, P., Vig, L., Shroff, G., & Agarwal, P. (2015). Long Short Term Memory Networks for Anomaly Detection in Time Series. In European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning. Bruges (Belgium), 22-24 April 2015, pp. 89-94.