AI: Machine Learning vs. Expert Systems

At Neurocern, we use an “expert system” (a type of artificial intelligence) to connect fragmented health data and enrich it for actionable insights to improve health and financial outcomes — but what does that mean, really? 

What do we mean by AI, what exactly is an expert system, and why is AI so relevant in healthcare? Read on for a deep dive into different types of AI and how an expert system stacks up.

When people talk about artificial intelligence (AI), they frequently imagine the droids from Star Wars or a robot with human behaviors and reactions. While this is the end goal of some types of AI research, the reality is generally more mundane. Current goals in AI frequently involve classification, which can be thought of as “pattern matching,” or the process of finding specific information from a jumble of facts. Research in this area is currently in vogue; applications of this technology can be seen in areas as diverse as speech recognition, natural language processing, and tumor identification from medical imaging data. This type of AI deals with large amounts of information that is expected to follow certain patterns; much of the research goes into identifying the patterns that are necessary to provide structure to seemingly random information.

Machine learning (ML) is a phrase frequently used in conjunction with classification. Unlike pattern matching, this is a technique rather than a type of artificial intelligence. The idea behind ML is that the patterns we want to analyze and on which classification will be based can be difficult to spot; having a computer do this work makes the entire process simpler and can yield new insights that might be missed by a human. 

However, the efficacy and accuracy of ML is dependent on the data that is used to “train” the model. To accurately train a system, a very large amount of data is required. Finding enough data that is both unbiased and fully representative of the target population is difficult; any inaccuracies or biases in the data being used to train ML systems will be “learned” by the system and propagate incorrect responses.

Expert systems, which have been in development since the 1970s and earlier, take a different approach to AI. Rather than attempting to identify new patterns and classify data based on statistics or operational experience, they are designed to replicate the problem-solving process of experts in the field. For example, MYCIN was developed to identify bacteria based on the research and experience of physicians with relevant experience and deep expertise in the area.

Rather than finding and matching data to known patterns (or attempting to derive new patterns to fit data), such a system is coded with the diagnostic logic of these experts, and attempts to follow the same real-world process that may be used by a bacteriologist, in the case of MYCIN. This requires significant skills and domain expertise; it is not a statistical process or a way of learning new techniques — it is a direct application of existing knowledge. An expert system uses technology to apply human expertise on a broad scale.

While the concept of a computer that can identify new patterns and gain new insights is appealing, the process can be misleading and inefficient. Just weeks ago, Defense One published an article titled “This Air Force Targeting AI Thought It Had a 90% Success Rate. It Was More Like 25%.” Microsoft introduced an AI chatbot named “Tay” in 2016 only to take it down hours after it launched due to inappropriate and offensive outputs. It had been “trained” with inaccurate input that was submitted to it on the Internet, which Tay then learned and repeated. These are not isolated incidents; ML is hard, and ambitious projects have a much greater likelihood of failure than success.

Expert systems are not prone to such issues, since they do not attempt to learn directly from the data provided to them during operation. In fields such as neurology — and medicine in general — using systems that rely on data inputs can have significantly negative effects if the data used for training or learning is not “clean” and representative; a situation such as what happened with Tay can cost lives if permitted to occur in medical practice.

At Neurocern, we have implemented a suite of sophisticated expert systems based on the deep and specific experience of our medical advisory team and our Founder, Dr. Anitha Rao. Our clinical team’s years worth of clinical expertise allows us to develop software based directly on medical research and best practices. While Neurocern reviews information processed by the system, all changes are evaluated based on current medical knowledge prior to being incorporated into the expert system algorithms, which always reflect the diagnostic approach of the medical community.