The Artificial Intelligence Continuum in healthcare

2019-07-26

What is Artificial Intelligence (AI) in medicine? Most people will answer with specific examples: Back in the 1970s, it would have been associated with something like MYCIN, a rule-based expert system for infectious diseases from Stanford University. Today, most people will mention software to detect pulmonary nodules on chest X-rays and CT studies, chatbots that are used in triage situations, or algorithms that predict which patient will suffer from heart failure, experience a seizure, or develop post-operative bleeding.

All these are excellent examples, but what about the broader picture? Let us do some localizing. The illustration below can be thought of as providing an expedition map into a world that is only partly explored and that could be called the Artificial Intelligence Continuum. A useful first step is to give present-day medical AI its proper place within the “grand narrative” of AI, i.e. the evolution from task-specific Artificial Narrow Intelligence (ANI) towards Artificial Super Intelligence (ASI).

Artificial Narrow Intelligence (ANI) is already in place in many different ways. ANI is programmed to perform a single task, based on specific data sets within a pre-determined range, faster and potentially more accurate than a human being can. Artificial Super Intelligence (ASI) is a thing of a far future. Click through the animation to learn about the whole continuum of AI.

ASI theoretically will surpass human intelligence. Attractive as it is, it will not become a reality in any foreseeable future. As for now, in medicine, we are in the realm of ANI, characterized by AI solutions that perform single tasks based on specific datasets within a predefined range. However, within this area, there are many advantageous applications to be explored. Many ANI solutions can perform tasks much faster and far more accurately than a human being ever could – and a machine will never tire, no matter how narrow and repetitive a task is.

In addition, the technology suffers from a mislabeling. ANI might sound narrow, but it is pretty diverse, and it offers plenty of opportunities for innovators. On closer inspection, it becomes clear that not only do we have an overarching trajectory from ANI to ASI. There is another trajectory within ANI, marked by increasing complexity, thanks in large parts to more data being integrated and aggregated from more and more data sources.

With this in mind, we can map different application fields of AI in medicine along the ANI trajectory, starting with fields in which algorithms are chiefly being used for automation and quantification, often embedded into medical devices and software solutions. Further along the curve, we approach more complex fields in which AI algorithms perform advanced analytics and predictions, be it on the level of the individual, of a certain population, or of some other cohort.

Before we continue, we would like to briefly outline five application fields on the ANI trajectory. Even at the very beginning of the digitalization process, we mainly talk about improving data acquisition and data generation. Take for instance CT imaging: A deep learning powered 3D camera can improve patient positioning, and can thus help to reduce radiation exposure while achieving higher quality CT scans even for stressed or novice operators. Similarly, a camera-based AI system can optimize the placement or routing of lab tubes in laboratory analyzers, which will improve workflows.

Algorithms like these can enhance the average quality of the digital data collected by any medical device. With the AI Continuum in mind, there is an interesting aspect here: Improved and standardized data quality will likely make it easier to use more complex algorithms further up the curve. In other words, data acquisition algorithms have immediate benefits to patients and user. But, they can also be seen as a prerequisite for making progress in higher ANI applications.

With increasing data complexity, we reach the fields of data processing/interpretation and data mapping/fusion. Many groups around the globe conduct research on data processing and interpretation. In imaging, we talk about digital segmentation and characterization tools, and about algorithms that automatically visualize, measure and classify. Beyond imaging, think of medical data mining in plain text documents, such as reports, or of interpreting other medical data of almost any kind.

Data mapping and fusion is where we leave single datasets and start to aggregate and apply AI algorithms on data from different sources, albeit from sources that still belong to a similar category. The most obvious example is imaging: Fusing live ultrasound with 3D MRI datasets in order to visualize a catheter during an intervention in real-time. It can be an enormous gain for both doctors and patients in terms of workflow optimization and reducing unwarranted variations.

This becomes different when we ascend the ANI trajectory further, towards where AI is being used for advanced analytics and prediction. This is happening now - it is the frontier of AI in medicine of our time, the big data realm of patient-centric predictions and cohort analytics.

Patient-centric predictive simulations are colloquially called "digital twins". High-level digital twins are lifelong physiological data models. They make use of all kind of available patient data - imaging, clinical records, and lab data including -omics, and they might also draw on behavioral data and on social determinants of health. These data sets can be integrated and analyzed, not only to provide a multi-dimensional risk model, but also to run simulations on the course of a disease and on the outcomes of treatments. Digital twins will also feed into sophisticated decision support systems which apply guideline recommendations and the most recent clinical trial knowledge to the individual patient data set, providing clinical guidance that is as tailored and as precise as possible.

Complexity further increases when predictive algorithms are used to compare the digital twin data set of an individual with those of similar patients. However, in order for this scenario to become a reality, there are still some technical as well as legal hurdles to be cleared. What we are talking about today are low-level digital twin models of individual organs like the heart. These are based on a manageable number of data sources - in the example of the heart, for instance, there is MRI data for dynamic mechanical and fluid-mechanical modeling, electrophysiological data, and vital data such as blood pressure. Models like these are currently being evaluated in clinical trials.

In summary, the concept of an AI continuum outlined above helps us to classify medical applications of artificial intelligence along a trajectory marked by rising complexity, increasing number of data sources, and closer patient involvement. On the lower end of the continuum, AI tools embedded into medical technology are available today and ready to increase quality of care and patient safety. With R&D ongoing, they will be supplemented by increasingly sophisticated predictive tools that help us to come closer to a future of healthcare in which precision medicine will be the new standard of care.