AI detects congestive heart failure with one heartbeat

By Dr. Liji Thomas, MDSep 12 2019

A new study has reported success in identifying severe heart failure in 100% of cases using a single heartbeat recording from an electrocardiogram (ECG). Medically, the condition called congestive heart failure (CHF) refers to a chronic loss of pumping power in the heart which is progressive. It is fairly common, causes significant illness and disability, and pushes up the costs of medical care. It affects about 26 million people around the world, and is more common in the elderly. It causes a considerable number of deaths, with about 40% mortality among the most severe cases. Even with treatment, relapses are common. It costs about 2% to 3% of total healthcare budgets.

Artificial neuron concept. Image Credit: ktsdesign / Shutterstock

How was the current study done?

The current study used an artificial intelligence (AI) approach called Convolutional Neural Networks (CNN) which makes use of neural networks arranged in layers of increasing complexity, similar to the visual pathway. These are able to detect data patterns and structures at extremely high efficiency, and have been used to perform speech recognition, arrhythmia detection and general time series classification.

The data for the control group was taken from the MIT-BIH Normal Sinus Rhythm Database containing 18 ECG recordings of healthy subjects, and for the CHF group from the BIDMC Congestive Heart Failure Database, containing 15 ECG recordings of patients with severe CHF. One heartbeat was randomly selected from each 5s segment. The dataset was split into three parts for training, validation and testing, and individual heartbeats were used in only one subset at a time.

The heartbeats were evaluated both individually, and on 5 minutes of ECG excerpts by a so-called majority voting scheme, using the number of heartbeats that were classified as normal or as CHF and assigning a final class.

Advantages of the current model

The current model uses raw ECG heartbeats that are processed by advanced signal processing and machine learning methods. This is a significant advance on existing machine learning methods which use heart rate variability (HRV) on ECG to detect CHF. HRV refers to the variation in the period between heartbeats recorded on ECG. These techniques have increased error rates and take more time to complete detection.

The current technique uses a 1-D CNN to allow the time series subsequences that act as the input to be visualized. This in turn allows the network to discriminate between CHF and healthy subjects, which not only allows faster detection but also helps us understand how certain tissue behaviour is related to the signals recorded. The sensitivity and specificity are also above 99% as compared to the model based on short-term HRV.

Earlier CNN-based models have also been proposed, but these used the recordings for specified subjects for both training and testing the models, rather than for either testing or training, as in the current model. This model is also simpler and less computationally complex, perhaps lending it suitability for mobile applications. Individual heartbeats are used as the input here, in contrast to the 2s ECG excerpts used as inputs in the earlier model; this helps researchers to explore the changes in the form of the heartbeat in CHF. In fact, this is the earliest evidence that segments in individual heartbeats can help discriminate a diagnosis of CHF.

Again, the current model highlights the features that are learned automatically while the process of classification is proceeding. Comparing the performance of this model with earlier classification methods, the error is found to be nil.

Researcher Sebastiano Massaro said: "We trained and tested the CNN model on large publicly available ECG datasets featuring subjects with CHF as well as healthy, non-arrhythmic hearts. Our model delivered 100% accuracy: by checking just one heartbeat we are able detect whether or not a person has heart failure. Our model is also one of the first known to be able to identify the ECG' s morphological features specifically associated to the severity of the condition.”

Future applications

The authors envision the adaptation of this system to wearable devices that may be able to perform prediction and detection of CHF using interim ECG recordings, by looking at individual heartbeat morphology. This could allow not only cardiologists but even patients and their caregivers, nurses, trainees and GPs to take part in the detection process. With the completion of the training phase, the network works very rapidly, making it fit for deployment into cloud systems or adaptation to mobile devices.

The authors point out the need for more studies in view of the fact that this study used only patients with severe CHF, which could affect the discrimination results for less severe heart failure. The population subset may need to be enlarged for a more generalizable model. Finally, work should be done on the potential of this system to predict CHF rather than merely detect it, to save life and reduce medical costs.

The study was published online in the Biomedical Signal Processing and Control Journal on September 3, 2019.