By Russ Michael, senior director, Methodology and Data Science, Blue Health Intelligence® (BHI®)
Artificial intelligence (AI)-driven technologies are increasingly prevalent in all parts of our lives, including healthcare and health insurance. The pace of change in healthcare AI and data analytics has been breathtaking, as new products and technologies promise to improve quality, reduce costs, and help health plans better meet customers’ needs.
Yet the terms AI, machine learning (ML), and deep learning (DL) often are used interchangeably – leading to confusion. In supporting BHI’s commitment to transparently sharing our analytics, we want to help our partners and collaborators better understand the differences between and promise of various forms of AI.
Artificial Intelligence Defined
AI is the larger field of data science that encompasses the subdisciplines of machine learning and deep learning. AI refers to training machines to mimic human intelligence and perform tasks. Machines use an algorithm or mathematical model to interpret the environment, discover relationships between factors, and predict future events.
Early AI systems often involved machines responding to expert-developed rules about the relationship between factors. More recent AI involves using rules developed through machine learning. Over time, as the machine perceives input from its environment, it can improve its ability to successfully achieve its goal.
Everyday applications: Customer service chatbots, which operate in real time, are powered by AI. AI also is used to eliminate mundane work, such as data entry.
Health plan applications: Health insurers are using AI-powered processing to speed the acceptance or denial of claims, and to detect fraud. AI also is being used to support actuarial functions.
Machine Learning Defined
ML is a subset of AI. Data scientists create ML algorithms to enable machines to “learn” by processing data without explicitly being programmed to learn. This allows machines to make determinations and predictions, rapidly perform calculations, or process a huge amount of data.
As a result, ML can rapidly discover unexpected data correlations. For example, ML might discover a relationship between the number of physicians a member sees, a member’s access to transportation, and a member’s disease outcomes. Of course, these correlations must be validated against clinical expertise, but if correct, they can be used to guide clinical interventions and possibly, AI applications.
Everyday applications: ML powers recommendations from Netflix or Amazon about which shows to watch, based on your viewing history.
Health plan applications: ML-powered AI is helping insurers predict when a member is at risk of suffering from a severe healthcare event, such as an ED visit, as well as predict the right moment to intervene. This can maximize the possibility of a good outcome.
Deep Learning Defined
If machine learning is about discovering relationships between factors such as causes and effects, DL is based on the premise that we may not know all the factors within relationships, so we might need to probe patterns within patterns. Much like we imagine that neurons in the brain are connected, DL creates numerous layers of algorithms, each providing a different interpretation of the data upon which they feed.
DL techniques include artificial neural networks, adversarial networks, and deep reinforcement learning. Such models are powerful and usually the most difficult to understand.
Everyday applications: Driver-assistance aids in vehicles, such as hearing a sound when reversing over a white line, were produced using neural networks. These aids are trained to distinguish between any white line and a hazard.
Health plan applications: Predicting metastatic cancer in at-risk members, an immensely complex task, would help a plan optimize care management. Traditional regression models and machine learning cannot perform this prediction, but DL may be able to unlock this mystery in order to guide earlier intervention.
Infusing AI with Equity
AI models predict the way the world is most likely to work, not the way it ought to work.. Therefore, data scientists must identify and mitigate these biases within the data in order to prevent them being built into AI models.
At BHI, we adjust our data models to compensate for bias. We also infuse the latest AI and ML capabilities as well as commitments to transparency and data equity in our software-as-a-service tools, analytic methodologies, and consulting engagements. Read more about how we achieve equity in healthcare modeling.