Marpai’s deep learning algorithms represent the greatest health state prediction available today in employer health plans. Marpai can predict potential near-term health events related to chronic illness (like Type 2 Diabetes) and major procedures (like knee surgery) and intervene early to prevent costly claims.

Powered by a deep neural network brain that mimics the logic and learning of the human brain, Marpai’s deep learning platform anticipates and predicts potential health events with unmatched speed and accuracy.

Deep learning is the most advanced subset of AI, leveraging deep neural networks that take inspiration from how the human brain works. It is artificial intelligence method that imitates the way the human brain works in the sense of processing data and creating patterns for use in decision making. It utilizes a hierarchical level of artificial neural networks to carry out the learning process involved in machine learning. The artificial neural networks are built like the human brain, with neuron nodes connected like an interconnected web.

As more data is fed into the neural network, it becomes better at intuitively understanding the meaning of new data. This allows it to predict and prevent increasingly advanced threats such as health risks. It does not require a (human) expert to help it understand the significance of new features.

The advantage of deep learning over other forms of machine learning is the end-to-end processing of data. These three factors contribute to deep learning’s greater level of accuracy.

Machine Learning systems rely on feature engineering which is limited to the knowledge of the expert who has to handcraft the features for detection. Machine learning-based solutions are still producing low detection rates and high false-positive rates. With Deep Learning, the algorithm analyzes all the raw data in a file, and is not limited by an expert’s capabilities. This represents a quantum leap in computer science. For future health states, this enables a more advanced level of prediction; with higher detection rates, lower false-positive rates and the ability to detect health risks effectively in zero-time.

Data scientists prepare data samples to train the “brain” or deep neural network. During the “training” phase, the deep neural network is exposed to all the available raw data in a file from which it learns to instinctively identify health risks. This process takes place within 24-48 hours. Once the network has reached the prediction stage, it can quickly and efficiently predict where a health risk exists or not. The input agnostic algorithm can apply this knowledge to any sort of file. Next, the brain is compressed into a lightweight agent where Terabytes of insights are turned into Megabytes of instincts. The agent is always working to detect health risks.

We are creating a robust deep learning infrastructure that will allow us to analyze nearly any type of data. More importantly, this infrastructure is expected to have the capability to automatically train on multiple data types at once and obtain unified predictions from that data. Specifically, we expect our deep learning capabilities to be used for the following data types:

Structured data: This includes any form of data that results in a tabular database. Examples of such data include laboratory tests, claims data, and payment details to healthcare providers. Our deep learning models for structured data, which rely on fully connected neural networks, allow us to automatically train on any such structured databases and incorporate those insights into predictive models.

Images: While traditional image processing requires cumbersome development for each specific use case, deep learning-based models can directly operate on any image data without having to go through a preprocessing feature extraction phase. We are developing a series of deep learning models based on convolutional neural networks that could be extended to process medical images, including radiology imaging such as MRIs, CT scans, x-rays, ultrasounds, and PET scans, as well as other types of medical images, such as those used in dermatology, pathology, and ophthalmology.

Text: Traditionally, analyzing text data requires lengthy natural language processing, or “NLP.” Recent breakthroughs in deep learning, especially the advent of transformer networks (e.g., BERT), allow for the end-to-end training of language models, which results in significantly higher accuracy than previous NLP methods. Our infrastructure includes support for these deep learning-based language models, allowing us to automatically process textual data as well. A prime example of textual medical data is the plain text description of doctors’ notes, which include information about symptoms, diagnosis, and treatment. Even though some of this data, such as disease codes, appears in structured databases as well, the textual data contains a large amount of information that does not appear elsewhere.

Multi-type data: Deep learning can also be applied to multiple data sources and data types. During the first stage, the relevant deep learning models are applied to each data type (e.g., CNN for images, Transformers for text, and so on), and then, during the second stage, the processed information from those various data sources is fed into a secondary deep learning model that provides unified predictions. This is one of the major advantages of deep learning over traditional AI, as it allows the incorporation of multiple big data sources into a single unified prediction model. This approach far exceeds the accuracy rate achieved by traditional methods.

Data fusion: With deep learning, we have the ability to study different data types together, such as data in the form of images (e.g., CT scans) with data in tabular form (e.g., figures from healthcare claims). Putting these different data types together is referred to as data fusion.