Introduction
Chronic diseases, including cancer, diabetes, and cardiovascular conditions, represent a staggering global health burden, accounting for nearly 75% of deaths worldwide. Managing these conditions is a complex, long-term endeavor that requires high patient adherence and continuous data integration. However, approximately half of all patients struggle to follow prescribed treatment regimens, often due to a lack of timely access to medical guidance. Large Language Models (LLMs) like ChatGPT and Bard have emerged as potential tools to bridge this gap, yet robust evidence regarding their clinical efficacy has remained sparse. This systematic review synthesizes current evidence on the feasibility, opportunities, and challenges of LLMs across the chronic disease management spectrum.
Methodology
Following PRISMA guidelines, researchers conducted a mixed-methods systematic review across 11 major databases, including MEDLINE, Embase, and IEEE Xplore, through April 2024. The study analyzed 20 papers—comprising 17 general-purpose LLMs and 3 Retrieval-Augmented Generation (RAG)-enhanced models. The review utilized a dual approach:
- Narrative Synthesis: A qualitative evaluation of the opportunities (such as social support) and systemic challenges (such as privacy).
- Meta-Analysis: A quantitative pooling of results to determine overall accuracy rates and compare the performance of "vanilla" models against those using RAG or fine-tuning.
The researchers also evaluated the readability of AI-generated health advice using standardized metrics like the Flesch-Kincaid Grade Level to determine if the outputs were accessible to the general public.
Key Findings
The Power of Retrieval-Augmented Generation (RAG)
One of the most significant findings was the performance leap associated with RAG. Models capable of "looking up" specific medical textbooks or clinical guidelines were nearly three times more likely (Odds Ratio 2.89) to provide accurate answers compared to general-purpose models. This suggests that while base models are powerful, grounding them in verified medical knowledge is essential for clinical safety.
Democratizing Health Information
LLMs demonstrated a superior ability to translate complex medical jargon into understandable language. While traditional medical websites often require an 11th-grade reading level, LLM responses averaged a 9th-grade level. This improvement in readability makes critical health data significantly more accessible to patients with varying levels of health literacy.
Feasibility and Accuracy
Across all management tasks—from prevention and screening to long-term care—LLMs achieved a pooled accuracy rate of 71%. In specific specialized tasks, such as cochlear implant care, some models achieved 100% accuracy. However, performance remained inconsistent in high-stakes clinical areas like liver transplant thresholds, where models occasionally failed to identify critical clinical cut-offs.
Challenges and Clinical Limitations
Despite the promising data, the review highlights that LLM application in clinical settings is still in its infancy. Several "red flags" persist:
- Hallucinations: Models occasionally cite non-existent studies or provide incorrect screening intervals, a phenomenon observed in over 50% of cases in one colorectal cancer study.
- Lack of Multimodality: Most current models are text-based and cannot yet integrate radiology images or auditory data (like heart murmurs) into a holistic diagnosis.
- The Empathy Paradox: While LLMs provide a non-judgmental space for patients to express concerns, they lack the deep cultural context and genuine human touch required for therapeutic relationships.
- Privacy Concerns: There is currently a lack of robust data protection guidelines for patients sharing sensitive health information with AI models.
Conclusion
LLMs have the potential to transform chronic disease management at the individual and healthcare system levels by providing accessible, accurate, and compassionate support. However, their direct clinical application requires a multifaceted approach. Future evolution must focus on integrating LLMs with wearable devices for real-time physiological monitoring and developing multimodal systems capable of processing lab results and imaging alongside patient symptoms. Until then, LLMs serve best as an adjunct to, rather than a replacement for, professional medical oversight.
Summary updated