The future of medical diagnosis is brighter than ever before. Imagine what would happen if the medical industry could predict strokes or heart failures? This is possible today, thanks to artificial intelligence technologies. Mayo Clinic recently published research [1] in the latest issue of The Lancet, which showed that machine learning models can identify individuals with atrial fibrillation, or abnormal heart rhythm, with an accuracy of 79–83%. We will dive into this research to understand what the team did well, what they can improve upon, what is the timeline for seeing this technology in consumer’s hands,

Let’s analyze what the Mayo Clinic clinic did well. The team used a large dataset: 650,000 ECGs of cardiac rhythm scans were acquired from almost 181,000 patients. The authors used a deep learning algorithm called convolutional neural network (CNN) which is suited for image analysis. If we analyze state-of-the-art methods for image classification, CNNs are often used by academia to define new industry standards for model performance.

Let’s analyze what others could improve upon. The approach used to train a deep learning model was supervised learning. While this is a good starting point, it is important to this approach does not take into account medical errors. Unsupervised deep learning methods include clustering analysis, sample specificity learning, self-supervised learning, generative models and anchor neighborhood discovery. An alternative to image classification is image clustering, a machine learning model that clusters images based on similarity. Similarity means similar images, similar size, similar pixel distribution, similar background, etc.

While the authors used image classification based on ECG data, it is important to understand that ECG data is classified into two ways: ECG beat classification and ECG signal classification.

While the approach used by Mayo Clinic is not clear, one may assume that ECG signal classification was used. It is more difficult compared to beat classification because normal ECG signals may differ for each person, sometimes one disease has dissimilar signs on different ECG signals and two distinct diseases may have approximately identical effects on ECG signals. [3]

Using ECG data is suited for consumers, not just medical professionals. The benefit of allowing consumers to receive in-app information about their personal health will encourage users to reach out to their doctors. The mobile apps, however, must comply with legal and medical standards in order to offer such information. Scheduling a follow-up with a doctor or a medical expert can be made on-demand with the help of mobile apps. How fast can this technology make way into the consumer’s hands?

  1. model results can be replicated if the authors release the source code their work and the dataset used for research purposes.
  2. dataset creation is possible in as little as 3–6 months
  3. model can be developed in as little as 3–9 months
  4. existing mobile devices that measure ECGs, such as AliveCor [2], can accelerate market adoption by giving this technology to consumers who make daily recordings

The medical industry is seeing massive improvements with the help of artificial intelligence technologies. The Mayo Clinic study shows us that it’s possible to use AI and machine learning to drastically improve medical diagnosis.

What are you waiting for? It’s time to apply AI and machine learning to solve other healthcare use cases. Message Produvia if you need a boost.

About Produvia

At Produvia, we produce intelligent software. We specialize in developing machine learning, deep learning, and natural language processing software. Since 2013, we partnered with entrepreneurs, small businesses, mid and large-sized enterprises to accelerate the adoption of AI.

References

  1. Attia, Z., Noseworthy, P., Lopez-Jimenez, F., Asirvatham, S., Deshmukh, A., & Gersh, B. et al. (2019). An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm: a retrospective analysis of outcome prediction. The Lancet. doi:10.1016/s0140–6736(19)31721–0
  2. AliveCor. (2019). Alivecor.com. Retrieved 3 August 2019, from https://www.alivecor.com/
  3. Classification of ECG signals using machine learning techniques: A survey — IEEE Conference Publication. (2019). Ieeexplore.ieee.org. Retrieved 3 August 2019, from https://ieeexplore.ieee.org/abstract/document/7164783