Comparison of Generative Artificial Intelligence and Predictive Artificial Intelligence

Artificial intelligence (AI) is “the science and engineering of creating intelligent machines that have the ability to achieve goals like humans via a constellation of technologies.”1(p69) The term was coined in the 1950s with high hopes for rapid, widespread use and important scientific breakthroughs. Progress in AI slowed, however, due to limitations in computing, analytics, and data. The landscape around these challenges has significantly evolved in the last 70 years, ushering in a new wave of opportunities. Among the opportunities are 2 AI technologies you will see in nursing practice, predictive AI and generative AI. Predictive AI uses data, such as patient data, to make predictions about future events or trends. Generative AI learns patterns in existing data and generates new data based on those patterns to create original content. This column of Technology Today column compares how these 2 AI technologies work and provides cases to illustrate how each may be employed to improve patient outcomes and advance nursing practice.Just a few years ago, researchers were largely focused on predictive AI that used particular datasets to find solutions for clinical and administrative issues, such as early identification of patient deterioration allowing for earlier intervention.2 The introduction of ChatGPT, a generative AI, moved a large portion of research efforts to focus on generative AI. Generative AI is advantageous in creativity and also uses significantly larger datasets than predictive AI—hundreds of millions or even billions of data points—allowing it to be useful in health care by creating efficiency in information gathering by combining and synthesizing multiple sources based on specific questions or prompts. ChatGPT has also recently removed its initial limitation of including only data before September 20, 2021; it now has unrestricted access to the past and current internet, allowing it to provide real-time, authoritative information that can now be linked to sources.3 Despite individual advantages, it is important to note that predictive AI and generative AI serve 2 distinct purposes (Table).In many industries—including health care—availability, accessibility, and analysis of structured data have made predictive AI easier and more useful to use than generative AI. For example, predictive AI uses structured data contained within a targeted dataset (eg, flowsheet cell data or laboratory results) relevant to the outcome of interest in order to make predictions that can inform clinical decisions.4 However, 80% of available and accessible data in health care is estimated to be unstructured (eg, progress notes).7 This is readily exemplified in electronic health records data. Newer predictive AI tools allow for the use of unstructured data.Generative AI analyzes unstructured data, using and learning from pattern recognition to create something original.8 Rather than focusing on one dataset, generative AI uses multiple data sources. For example, ChatGPT uses data from public sources, licensed thirdparty data, and information from human reviewers.9 The bottom line is that both predictive and generative AI analyze data and apply sophisticated algorithms to achieve a specific desired outcome, one that is more defined in predictive AI and more innovative in generative AI.Examples of outcomes from using predictive AI include predicting outbreaks of infections, identifying patients at risk of particular conditions, and discerning the timing for nursing interventions.4 Generative AI may solve problems that have long plagued nursing such as inadequate staffing and cumbersome documentation requirements. Through its creative focus, it may generate a solution for a more efficient and safe staffing model that is highly satisfactory to both nurses and employers. It is also being tested on its ability to create accurate clinical documentation from multiple data sources such as voice, images, sensors, and more. Generation of these new potential solutions will likely be aided by the ability of generative AI to learn from large amounts of unstructured data, making it more flexible and adaptable as it changes in response to new data.The most important thing to keep in mind for predictive AI and generative AI for nursing practice is accuracy. As I’ve previously discussed in this column, with any innovation there are downsides. Think about cars. As cars continue to advance, there are routine automaker recalls, often for safety reasons. Similarly, both predictive and generative AI tools are vulnerable to bias and errors. Although these areas are being addressed, we are not where we need to be yet. It highlights the most important role nurses play in using AI: Nursing practice should always be considered AI-enabled but nurse-driven. The importance of this approach is demonstrated in the following cases.In June 2021, a study published in JAMA Internal Medicine by University of Michigan School of Medicine researchers reported that the Epic AI sepsis tool failed to predict 1709 (67%) sepsis cases in the Michigan Medicine health system between December 6, 2018, and October 20, 2019.10 Clinicians missed only 7%.Researchers at Atrium Health Wake Forest Medicine in North Carolina assessed the validity of Epic’s AI sepsis tool by comparing it to existing sepsis prediction tools, Systemic Inflammatory Response Syndrome, Sequential Organ Failure Assessment, and Sepsis-Related Organ Failure Assessment.11 The retrospective study involved acute care hospitals in a single US health system between June 5, 2019, and December 31, 2020. The researchers reported the Epic tool was more accurate in prediction at the highest scoring thresholds but had poor timeliness for sepsis prediction. Researchers stated, “It appears to predict sepsis long after the clinician has recognized possible sepsis and acted on that suspicion.”11(p7)In 2022, Epic released a re-engineered version of the sepsis prediction model after changing the data variables, the definition of sepsis onset, and guidance for tuning the algorithm to local patients.12 No studies using this updated version were identified in a search of the literature.The ongoing case of Epic’s sepsis prediction model is a cautionary one that is not unique to Epic and affords important learning opportunities in the early stages of using AI in clinical settings. The tool is intended to support clinicians in the early identification and intervention of sepsis, the largest contributor to mortality and cost of hospitals in the United States. Survival rates for sepsis depend on early identification and treatment. Rigorous studies and publication in peer-reviewed journals are essential before the release of AI tools to ensure reliability and validity. Otherwise, the need to rigorously pilot new AI tools in practice settings where they will be used is imperative.Although this case is not specific to nursing, it involves the ability of generative AI to assist with an area of importance to nursing: clinical documentation.13 In a blinded, randomized controlled trial, researchers examined the use of ChatGPT to assist with documenting the history of patients’ present illnesses based on a standardized patient. The intervention consisted of 1 of 3 randomly assigned documentation techniques: ChatGPT, dictation, and typing. The Physician Documentation Quality Instrument (PDQI-9) was used to measure the quality of the documented history of present illness (HPI).14 The tool measures the accuracy, completeness, and organization of clinical documentation.In this study, ChatGPT produced more detailed documentation based on word count with a mean of 135 words for ChatGPT, 89 words for dictation, and 67 words for typing (P < .001).13 ChatGPT was reported to be intermediate in speed, with a mean time of 69 seconds, but this was not significantly different from dictation (P = .14). ChatGPT had the highest PDQI score at 35.6, followed by dictation at 31.6 and typing at 30.4 (P = .001).13It is important to note that ChatGPT also generated erroneous information in 4 of 11 documents (36%). These errors represented ChatGPT adding physical examination findings not included in a patient history, and 2 patient history details not included in the standardized HPI script used and not added by the person documenting the history.13 The information was not related to the patient or elements of an HPI. It was created by Chat-GPT, a so-called hallucination, and is thus inaccurate, creating concerns about the reliability of ChatGPT in HPI documentation.As nursing continues to evolve in the digital world, keeping pace with change as well as planning for future changes will be essential. The goals for using specific AI-enabled tools in clinical practice should be clear to nurses and valued by them. These goals may include using AI to support decision-making and problem-solving; enhance safety and minimize errors; communicate with others in person and electronically; document, access, and share information; make nursing practice more efficient and less stressful; improve the patient experience and outcomes; and advance the profession.15Nurse leaders should create a culture of positive attitudes toward new technology. Equally important, they should ensure nurses are prepared and the technology is safe before implementation. When using AI-enabled tools, start with the problem, not the technology.16 If the solution to the problem involves making a prediction, predictive AI tools perform better than generative AI.2 If you are looking for new ideas on how to improve nursing practice, generative AI is the first choice to consider.Google chief executive officer Sundar Pichai describes the impact of AI to human society as being more profound than fire or electricity.17 He goes on to say AI technology will be the biggest technological shift in our lifetime, perhaps even bigger than the internet. AI will redefine how we interact with machines, data, and patients. Generative AI can help us identify new solutions to age-old problems, such as defining nursing workflows we have not yet envisioned. Predictive AI can empower nurses by supporting data-driven decisions, such as intervening earlier to prevent patient deterioration. Both have the potential to enable us to glean significantly greater value from data. As AI evolves, the distinctions between the 2 types of AI are likely to diminish—AI systems are already being developed that merge the 2.6 The combination of prediction and creativity in one AI tool will further enhance the value of data as well as the science and practice of nursing.