Qiu, J., Lam, K., Li, G. et al.
Nat Mach Intell (2024).
Large language model-based agentic systems can process input information, plan and decide, recall and reflect, interact and collaborate, leverage various tools and act. This opens up a wealth of opportunities within medicine and healthcare, ranging from clinical workflow automation to multi-agent-aided diagnosis.
Large language models (LLMs) exhibit generalist intelligence in following instructions and providing information. In medicine, they have been employed in tasks from writing discharge summaries to clinical note-taking. LLMs are typically created via a three-stage process: first, pre-training using vast web-scale data to obtain a base model; second, fine-tuning the base model using high-quality question-and-answer data to generate a conversational assistant model; and third, reinforcement learning from human feedback to align the assistant model with human values and improve responses. LLMs are essentially text-completion models that provide responses by predicting words following the prompt. Although this next-word prediction mechanism allows LLMs to respond rapidly, it does not guarantee depth or accuracy of their outputs. LLMs are currently limited by the recency, validity and breadth of their training data, and their outputs are dependent on prompt quality. They also lack persistent memory, owing to their intrinsically limited context window, which leads to difficulties in maintaining continuity across longer interactions or across sessions; this, in turn, leads to challenges in providing personalized responses based on past interactions. Furthermore, LLMs are inherently unimodal. These limitations restrict their applications in medicine and healthcare, which often require problem-solving skills beyond linguistic proficiency alone.
Here are some thoughts:
Large language model (LLM)-based agentic systems are emerging as powerful tools in medicine and healthcare, offering capabilities that go beyond simple text generation. These systems can process information, make decisions, and interact with various tools, leading to advancements in clinical workflows and diagnostics. LLM agents are created through a three-stage process involving pre-training, fine-tuning, and reinforcement learning. They overcome limitations of standalone LLMs by incorporating external modules for perception, memory, and action, enabling them to handle complex tasks and collaborate with other agents. Four key opportunities for LLM agents in healthcare include clinical workflow automation, trustworthy medical AI, multi-agent-aided diagnosis, and health digital twins. Despite their potential, these systems also pose challenges such as safety concerns, bias amplification, and the need for new regulatory frameworks.
This development is important to psychologists for several reasons. First, LLM agents could revolutionize mental health care by providing personalized, round-the-clock support to patients, potentially improving treatment outcomes and accessibility. Second, these systems could assist psychologists in analyzing complex patient data, leading to more accurate diagnoses and tailored treatment plans. Third, LLM agents could automate administrative tasks, allowing psychologists to focus more on direct patient care. Fourth, the multi-agent collaboration feature could facilitate interdisciplinary approaches in mental health, bringing together insights from various specialties. Finally, the ethical implications and potential biases of these systems present new areas of study for psychologists, particularly in understanding how AI-human interactions may impact mental health and therapeutic relationships.
