Generative AI and Patient Care: A Systematic Review Examining Applications, Limitations, and Future Directions for ChatGPT in Healthcare

Theoretical Concepts

Generative AI refers to a class of artificial intelligence systems that can generate new content like text, images, or video, rather than solely analyzing existing data [2]. This distinguishes them from traditional predictive or analytical AI. One example is ChatGPT, a large language model developed by OpenAI to produce human-like text through reinforcement learning techniques [3].

Key theoretical frameworks guiding this review include human-AI interaction and ethical AI design. The human-AI interaction framework recognizes that AI systems do not operate in isolation – their integration in healthcare settings has bidirectional effects on clinical workflows, human behaviors, and patient outcomes [4]. Factors like trust, usability, and transparency are vital for safe and effective human-AI collaboration. The ethical AI framework posits that AI systems should be designed and evaluated based on principles of justice, beneficence, and non-maleficence [5]. Applied to healthcare, this entails ensuring equitable access, promoting patient wellbeing, and avoiding harms through rigorous testing.

This review focuses on two primary applications of generative AI in healthcare:

• Clinical decision support – Providing clinicians with patient-specific assessments, diagnoses, or treatment recommendations to enhance medical decision-making [6].
• Patient outcome improvement – Using AI to directly improve metrics like mortality, readmission rates, or recovery times that reflect patient health results [7]. Additionally, the emerging techniques of visual computing and natural language processing for analyzing medical images and generating written content were reviewed [8, 9]. Overall, this review synthesizes the current evidence on generative AI in healthcare through the theoretical lenses of human-AI collaboration and ethically aligned design. The aim is to provide data-driven insights on capabilities, limitations, and key considerations for translational efforts.

Literature Search Strategy

A systematic search was conducted of the Web of Science, Scopus, and ProQuest databases to identify relevant studies published in 2023. The search strategy involved three search themes using the terms:

• “generative artificial intelligence” or “chatgpt” or “large language models” or “generative ai” or “chatbot” or “conversational agents”
• “healthcare system” OR “clinical decision support”
• “patient record analysis” OR “disease diagnosis” OR “treatment planning”
• The final search strategy was: (“generative artificial intelligence” OR “ChatGPT”) AND (“healthcare system” OR “clinical decision support”) AND (“patient record analysis” OR “disease diagnosis” OR “treatment planning”).

Study Selection

The database search yielded 191 articles. Studies were included if they: (1) were published in 2023, (2) were original research or review articles, (3) were in English,(4) examined applications of generative AI in healthcare.

Editorials, books, conference proceedings, and non-English studies were excluded from the review. After screening, 37 studies remained for inclusion in the review (Figure 1).

Click to enlarge

Data Extraction

Each study was reviewed to consider the review’s research questions and an extraction tool was built covering key variables like study design, demographics, AI system details, and outcomes. Drawing from existing frameworks, both quantitative and qualitative fields were included. Before extracting data from all studies, a pilot test of the tool on a sample was performed, finding areas needing clarification or new fields, like AI model type. Refining the tool accordingly made it more comprehensive.

For each study, data was extracted in passes, starting with general info before capturing more details. To minimize errors, data was thoroughly reviewed, revisiting studies when needed to ensure understanding of methods or outcomes. A structured data handling approach was designed, entering extractions into a secure spreadsheet with quality checks. I reviewed each study’s data a minimum of three times.

Quality Assessment

Multiple tools assessed the methodological quality and risk of bias of included studies:

• AXIS for cross-sectional studies
• ROBIS for systematic reviews
• JBI checklists for other study designs

Data Synthesis

Given the diverse study designs, a quantitative meta- analysis was not feasible. Instead, a qualitative comparative analysis approach was used to synthesize findings, identify themes, and evaluate the overall strength of evidence. Tables and narrative descriptions summarize the evidence.

ChatGPT for Clinical Training

Some studies explored supervised use of ChatGPT for training clinicians. Chervenak J, et al. [7] described using ChatGPT to generate detailed case simulations for surgical training, providing exposure to complex scenarios learners may not frequently encounter. With oversight from experienced clinicians, this application appeared beneficial for building clinical reasoning skills. Temsah MH, et al. [10] found that ChatGPT could efficiently generate novel practice cases across medical disciplines, though they advised careful review to catch any potential inaccuracies.

In a survey of surgical trainees, Karabacak M, et al. [11] found 70% viewed ChatGPT-assisted learning positively, reporting it improved their medical knowledge. However, they noted the importance of governance frameworks to avoid misuse. Zumsteg JM, et al. [12] similarly concluded that ChatGPT showed promise as an adjunctive training tool in surgical specialties but emphasized the need for supervisor guidance to mitigate risks to patients. Overall, findings indicate supervised ChatGPT use could provide valuable training opportunities, but direct oversight is critical to avoid potential harm.

ChatGPT for Clinical Decision Support

Several studies evaluated the risks around using ChatGPT for clinical decision-making. Chavez MR, et al. [6] found that while ChatGPT could generate diagnostic and treatment suggestions for clinical scenarios, the outputs contained concerning factual errors, inaccurate or outdated guidelines, and inappropriate recommendations. They concluded that ChatGPT’s lack of medical training precludes independent use for clinical decision support until the technology matures. Xiao D, et al. [13] systematically reviewed ChatGPT’s responses to complex medical questions and concluded the information was often incomplete or partially incorrect. They advised thorough review of ChatGPT’s responses against current medical literature before considering acting on any recommendations.

Mello MM, et al. [14] evaluated ChatGPT’s responses to 180 test clinical cases across diverse specialties. While responses were articulate and sounded plausible, approximately one-third contained serious errors that could lead to patient harm if used for decision-making without oversight. The lack of transparency around how ChatGPT draws conclusions further complicates use in practice. Şendur HN, et al. [15] similarly found that while ChatGPT could provide coherent responses to clinical queries, only 65% were fully accurate and appropriate for guiding treatment decisions. Taken together, these findings indicate significant risks if ChatGPT outputs are applied clinically without proper validation (Table 1).

ChatGPT for Patient Communication

A small number of studies have examined ChatGPT’s potential for simulating patient communication scenarios. Cadamuro J, et al. [5] and Xie Y, et al. [17] found that ChatGPT could clearly explain complex medical procedures, risks, and benefits when prompted, showing an ability to simplify healthcare jargon. Other research by Giannos P, et al. [18] and Mallio CA, et al. [19] demonstrated ChatGPT’s capacity for providing coherent, appropriate responses to common patient questions in simulated digital clinic interfaces. However, more advanced patient-provider discussions involving empathy, sensitivity, and personalized care recommendations proved challenging for the current AI system. Mondal H, et al. [20] concluded significant work is still required to achieve the nuanced bidirectional dialogue exemplified in human clinical conversations. This technology shows promising competence in rudimentary patient education tasks, however, ChatGPT has important limitations in replicating multifaceted patient- provider discussions. Early studies provide preliminary support for its potential uses in simplified communication scenarios, but further rigorous evaluation is needed across diverse settings and interaction types to fully determine ChatGPT’s capabilities and real-world utility for enhancing patient care through improved communication.

ChatGPT for Medical Guidance and Education

Some studies explored ChatGPT’s potential to enhance medical guidance and education (Table 2).

While promising for streamlining certain data management tasks, findings reinforce the importance of governance and validation of ChatGPT-generated outputs to avoid potential patient harm. Further rigorous inquiry into the contextual factors driving variability in performance is warranted.

Synopsis of Research

This systematic review of 33 studies focused on three key potential applications of generative AI in healthcare:

• Clinical Documentation and Data Management a) Generative AI like ChatGPT shows promise for automating clinical documentation tasks, such as synthesizing patient records into coherent summaries or generating procedure notes. b) However, risks of erroneous outputs reinforce that human oversight is necessary when implementing for clinical uses.
• Clinical Decision Support a) Some studies found ChatGPT could provide appropriate evidence-based diagnostic and treatment suggestions when given patient data. b) But other analyses revealed high rates of factual errors, outdated information, and potentially harmful recommendations from ChatGPT. c) Vetting of ChatGPT outputs against medical literature is currently advised before clinical application given variable accuracy.
• Medical Education and Training a) Supervised use of ChatGPT-generated case simulations appears beneficial for building clinical reasoning skills. b) However, oversight by experienced clinicians remains critical to avoid potential patient harm through incorrect training content.

Areas like clinical documentation automation appear promising for near-term, controlled implementation, findings reinforce that ChatGPT requires significant improvement and governance before safely integrating for direct decision support or independent use in clinical care. Ongoing advances in AI may expand applications, but human expertise remains essential in healthcare.