Types of Deep Learning & Their Uses in Healthcare

May 06, 2022 - Deep learning (DL), which is also known as deep structured learning or hierarchical learning, is a subset of machine learning. It is loosely based on the way neurons connect to one another to process information in animal brains. To imitate these connections, DL uses layered algorithmic architecture known as artificial neural networks (ANNs) to analyze data. By analyzing how data is filtered through an ANN’s layers and how the layers interact with one another, a DL algorithm can 'learn' to make correlations and connections in the data.

These capabilities make DL algorithms innovative tools with the potential to change healthcare. The most common types found in the industry have various use cases.

DEEP NEURAL NETWORKS

A deep neural network (DNN) is a type of ANN, but is classified as 'deep' because it has a greater depth of layers than some other neural networks. These layers perform the mathematical translation tasks that allow raw data to be translated into meaningful output. Additional layers allow for slightly different translations to be performed in each layer with the intention of further refining the output.

A recent example of a DNN use case shows how these algorithms can be used to predict pediatric appointment no-shows.

In a study published in npj Digital Medicine, researchers hypothesized that patient EHR data and local weather information could be used as predictors for pediatric no-shows and that a prediction model utilizing this information could help providers implement no-show prevention measures.

The researchers developed their model by retrospectively collecting EHR data for 19,450 patients between Jan. 10, 2015, and Sept. 9, 2016, at Boston Children's Hospital’s primary care pediatric clinic. Of the 161,822 appointments connected to these patients, 20.3 percent were no-shows. Information about the local weather on the day of the appointment was also included.

The DNN outperformed the conventional no-show prediction approach, and it showed that atmospheric pressure and the history of patients’ no-show records were the most important predictors of whether patients would show up to their next appointment.

By using a DNN, researchers could more easily evaluate the relationship between various factors related to a patient no-show and determine which factors were most strongly correlated with the outcome.

CONVOLUTIONAL NEURAL NETWORKS

Convolutional neural networks (CNNs) are a type of DNN used to understand visual data. CNNs analyze images and extract features that they can use to classify images into categories.

Classification is key in arenas like medical imaging, where a clinician will look at an image, such as a CT scan or an X-ray, to diagnose a variety of conditions. Implementing a CNN algorithm to assist with medical imaging tasks has the potential to improve clinical decision support and address issues related to population health.

In a study that examined the relationship between retinal disease and lack of access to care, researchers sought to develop a CNN model that could detect multiple retinal diseases simultaneously.

They developed their model using 120,002 ocular fundus images that were evaluated by a group of certified ophthalmologists and labelled according to retinal disease diagnosis. To evaluate the model’s ability to accurately detect retinal disease, its performance was compared to that of a group of retinal specialists.

The CNN achieved an accuracy higher than or equal to that of the specialist group in seven of the 10 retinal diseases being evaluated. The model also outperformed clinicians when comparing image assessment speed, analyzing one image in less than a second while the fastest specialist took 7.68 seconds with the same image.

The researchers posited that the success of their model could help address access to care in underdeveloped regions where lower screening rates and late diagnosis of retinal disease are associated with increased risk of irreversible vision loss.

RECURRENT NEURAL NETWORKS

Recurrent neural networks (RNNs) are another type of ANN that use sequential or temporal data. They are often used for problems related to language translation, natural language processing, speech recognition, and image captioning. Unlike other neural networks, where inputs and outputs are independent of one another, RNNs take information from inputs in prior layers to influence the current inputs and outputs.

RNNs are useful for healthcare providers to assist with tasks such as clinical trial cohort selection.

In clinical trials, a cohort or group of patients who share similar relevant characteristics are selected for participation in research. Because the success of a clinical trial relies on accurate selection of a patient cohort, the process can time consuming and costly.

Researchers seeking to reduce time and costs set out to test whether various DL models could successfully identify key features for cohort selection. They trained and tested a simple CNN, a deep CNN, an RNN, and a hybrid model that combined both CNN and RNN. Models were given patient records manually labelled by experts to indicate whether patients met one or more criteria out of 13 possible options for a clinical trial.

Overall, the RNN and the hybrid model significantly outperformed the CNN models. However, the researchers noted that their study was limited by the small dataset they used and indicated that further studies would be needed before a cohort selection model could be implemented.

GENERATIVE ADVERSARIAL NETWORKS

Generative adversarial networks (GANs) use two neural networks to generate synthetic data that can be used in the place of real data. GANs are commonly used in image, video, and voice generation.

GANs have great potential for use in healthcare because of their ability to generate synthetic MRI images. Using medical images to train AI models for diagnostics and predictive analytics creates multiple challenges for researchers because their quality may vary, they may be subject to patient privacy regulations, and image datasets are often imbalanced.

To address these concerns, researchers have explored using MRI images created by GANs to train deep-learning models for clinical decision support. In one study, researchers trained a GAN to create abnormal MRI images using publicly available scans from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and the Multimodal Brain Tumor Image Segmentation Benchmark (BRATS).

The GAN-created MRIs did contain some features that allowed the researchers to distinguish them from real MRIs, but the team pointed out that a larger training dataset could eliminate that issue in the future.

Synthetic medical images would potentially allow researchers to create large datasets with high-quality images and a more balanced distribution of pathological findings.

In addition to improving the imaging data used to train other AI, GAN-created medical images could help protect patient privacy by decreasing the need to use real medical images for research.