When a new technology is marketed as “AI-powered,” what does that mean? Without a clear understanding of this label, it becomes difficult to appropriately evaluate, trust, and adopt new technology. The first step toward responsibly using and integrating AI-enabled technologies into healthcare is to understand what artificial intelligence or “AI” is.
The term Artificial Intelligence was coined in 1955 by John McCarthy, who defined it as “the science and engineering of making intelligent machines.”8 (p. 2) Since then, dozens of definitions have emerged, many of which remain similarly broad and vague. The term “AI” continues to act as an umbrella term—a catch-all for any technologies that emulate human thinking or behavior. As a result, the utilization of the term “AI” to describe a technology does little to explain how the technology works or what it is capable of. More meaningful insight comes from understanding the different underlying methods by which AI emulates human thinking; it is the advancements in these methods (often referred to as “models” or “algorithms”) that have led to substantial improvement in the utility of AI.9 This section will briefly introduce three major AI methods, all of which exist in healthcare technologies in practice today: rule-based systems, traditional machine learning, and deep learning.
Rule-Based AI
Early AI applications relied on inflexible, deterministic “if-then” rules to emulate human reasoning. A familiar healthcare example is the medication contraindication alert: an electronic health record might warn a provider that a new Warfarin prescription may lead to an increased risk of bleeding for a patient taking Aspirin. Another classic example is DXplain, a clinical decision support system that generates differential diagnoses based on a structured set of expert-defined rules and heuristics. While such systems mimic human-like intelligence in appearance and interaction, they involve no learning, thinking, or deeper processing beyond their predefined rules. Rule-based AI applications remain abundant in healthcare but are rarely described as “AI” due to their relatively simplistic functions. In essence, rule-based systems represent intelligence symbolically, as they contain no adaptive or data-driven learning.
Traditional Machine Learning
The next major advancement in AI was the introduction of machine learning, a subset of AI that uses statistical methods to “learn” from historical data and make predictions on new, unseen inputs. Machine learning (ML) is itself an umbrella term, encompassing both traditional algorithms and more recent methods like deep learning.10 Traditional ML models rely on structured, transparent algorithms (e.g., linear regression, decision trees) to generate predictions.
Developing these models typically requires domain expertise to select inputs, or variables that are logically connected to the outcome being predicted. For example, traditional ML has been used to build tools that predict early-onset sepsis using biometric data.11 To build an effective model, the selected biometric data typically include physiological variables known to indicate infection, such as heart rate, respiratory rate, temperature, and blood pressure. Another use-case includes tools designed to predict the duration of a procedure using patient information, procedure information, and personnel information in order to improve operating room scheduling efficiency.12,13 These models are trained on historical data where the outcome (e.g., case duration) is already known, and then used to predict outcomes for future cases.
Similarly, many widely used clinical risk calculators, such as the CHA₂DS₂-VASc score for stroke risk or the Wells score for pulmonary embolism, are examples of traditional ML models trained on structured clinical data to predict diagnostic outcomes. While the development process may be more labor-intensive and reliant on expert input, the benefit is that traditional ML models often provide insight into why a prediction was made. For instance, a model might indicate that a patient’s risk of stroke increases by 2 percent for each decade of age, an interpretable relationship that clinicians can easily understand and validate. Despite being more transparent and statistically grounded than contemporary deep learning models, tools built with traditional ML are still commonly referred to as “AI” in healthcare. While it is categorically accurate to label these tools as “AI,” the implementation considerations (i.e., risks) for traditional ML models differ markedly from those of deep learning. As such, referring to tools by their underlying model type rather than the broad label of AI may promote clearer understanding and more appropriate use.
Deep Learning
The current wave of AI advancement is largely driven by deep learning. Like traditional ML, deep learning models are trained on historical data to make predictions about new cases. However, instead of relying on a manually selected set of inputs and interpretable statistical relationships, deep learning effectively uses a web of virtual brain cells (called a “neural network”) to automatically detect complex patterns in large and often unstructured data. In healthcare, this web enables deep learning models to incorporate a much broader range of inputs, including thousands of data points from vital signs, lab results, medication histories, and even clinical notes. For example, a sepsis prediction model powered by deep learning utilizes a wide array of physiological and laboratory variables—including lymphocyte count, lymphocyte ratio, white blood cell count, neutrophil count, thrombocyte count, c-reactive protein, procalcitonin, and more14—some of which may not be definitive indicators of infection when considered in isolation.
The strength of deep learning lies in its ability to analyze vast and diverse inputs to identify subtle, non-obvious patterns that may be unknown to clinicians. For instance, deep learning has been used to predict protein folding, an achievement once thought to be unattainable, which has important implications for drug discovery and treatment development.15 However, this power comes with a tradeoff. The complexity and opacity of these models make it challenging (and, in many cases, currently impossible) to fully understand why a particular prediction was made, which can raise concerns around trust, safety, and accountability in clinical practice. Much of the recent momentum in AI can be traced to major breakthroughs in deep learning, particularly the introduction of transformer architectures in 2017,16 which have enabled systems capable of assisting with diagnosis, documentation, and decision-making across a wide range of healthcare settings.
How Deep Learning Works
Understanding the basics of how deep learning works is important for anyone interacting with AI-powered tools. While these models are often described as “black boxes” due to their complex and opaque internal logic, even a preliminary understanding of how they function can help illuminate both their potential and their limitations.
At its core, deep learning involves training a model on a large amount of data so it can recognize patterns and make predictions. One way to conceptualize this pattern recognition process is through the analogy of a toddler playing with a shape sorter toy—a common developmental activity in which children are given various blocks (like spheres, cubes, and pyramids) and must fit them into matching holes on a container. Upon first interacting with the toy, the child may pick up a spherical block and attempt to place it into the triangle-shaped hole without success. The child may then attempt to place the spherical block into the square-shaped hole, also without success.
Eventually, the child will slide the spherical block through the circle-shaped hole, indicating a successful attempt. Through repeated trial and error, the child learns that the sphere fits into the circle-shaped hole. With time, the child becomes more efficient at matching all the shapes to their corresponding holes. In a similar way, deep learning models adjust their internal structure over many iterations to reduce errors and improve at recognizing complex patterns. Just as the toddler learns from experience, the model becomes more effective at identifying the types of patterns it was trained to detect. While this description offers only a simplified overview, numerous introductory textbooks,17 videos,18 and tutorials are available online for those seeking deeper understanding.
