“The future is ours to shape. I feel we are in a race that we need to win. It’s a race between the growing power of the technology and the growing wisdom we need to manage it.”

Max Tegmark, Swedish American physicist and AI expert

This excellent overview of the state of machine learning in healthcare from Nature Biomedical Engineering journal is both insightful as well as timely. The authors should be congratulated on such a comprehensive review of machine learning that is very helpful for anyone in this domain. The article focuses mainly on the central role of healthcare data in assuring that models mature from development to deployment. The myriad of topics in this data-centric review range from deep generative models to federated learning as well as transformer models designed for larger datasets and clinical text.

Generative adversarial networks can create large amounts of synthetic yet realistic data to not only augment datasets but also mitigate the problems of privacy restricted or unbalanced datasets. As the original GAN had issues with unstable training and low image diversity quality, newer GANs have arrived with newer architectures that minimize sensitivity to hyper parameter and to mode collapse. An example is deep convolutional GAN (DCGAN) that is used for medical image tasks with a feature of introducing key architecture design decisions. Many other GANs congeners are listed in a table. In addition, attention mechanisms in GAN architecture have become more common as these can capture longer range global and spatial relations from the input data. In addition, GANs can augment training data in order to improve model performance. Lastly, GANs can protect patient privacy by being used as a patient anatomization tool to generate synthetic data for model training. An innovative use of GANs is its capability to do image-to-image translation.  

The authors went on to describe the usefulness of federated learning, which is a tool for training models when decentralized data are used collaboratively under a central server. The two types of federated learning, cross-silo and cross-device (data generated at the edge), are described in the paper and accompanied by a useful table. The data aspect of federated learning is that data will need to be normalized and standardized.

The transformer has revolutionized natural language processing with its more parallelizable and less computationally complex model (compared to RNN and CNN models). The transformer model uses the aforementioned attention mechanisms within its encoders and decoders. This capability of the transformer has improved the transfer learning as well as clinical text understanding of NLP in healthcare. In addition, transformers are also used for modeling of clinical events. 

The authors conclude this detailed overview with a discussion on data limiting factors that can affect model performance. One issue has been the difficulty with transforming raw unstructured and heterogenous clinical data into structured data that can be input into the models in the data pipeline. Another challenge has been deployment in the presence of data shifts that can lead to algorithm bias. Two limitations of ML models are that models can propagate biases and that insufficient diversity can lead to inadequate generalization of models.

While sections and tables are more for the technically-savvy readers, the overall framework and content of this article are excellent and worthwhile to read for any clinician interested in AI in clinical medicine and healthcare.  

Read the full paper here

This fascinating topic of AI strategies for healthcare leaders, along with others will be discussed at the annual AIMed Global Summit 2023. Book your place now! 

Explore related articles in our Tools and Deployment content hub here

We at AIMed believe in changing healthcare one connection at a time. If you are interested in discussing the contents of this article or connecting, please drop me a line – [email protected]