JMIR Publications

ABSTRACT

Background:

The exponential growth of biomedical literature has created a critical bottleneck in translational research. Conventional systematic reviews are constrained by manual curation, human screening bias, and time-intensive data extraction. While cloud-based artificial intelligence (AI) offers high-throughput processing, its application in medical informatics introduces severe data privacy risks and potential stochastic hallucinations

Objective:

To engineer, deploy, and validate a secure, fully automated natural language processing (NLP) pipeline utilizing an edge-hosted large language model (LLM) to accelerate systematic reviews, using preventive regenerative medicine interventions for the "Mibyo" (systemic resilience) state as a proof-of-concept.

Methods:

A 120-billion parameter open-weights LLM was deployed on a high-performance, strictly offline edge computing infrastructure (NVIDIA DGX SPARK). To ensure deterministic and reproducible data extraction, the generation temperature was set to 0.0. A zero-shot prompting framework, combined with a custom Python-based regular expression parser, was engineered to autonomously extract predefined therapeutic metrics (e.g., cell source, Senescence-Associated Secretory Phenotype suppression) from 3,228 PubMed abstracts into a structured JSON format. To validate the architectural necessity of massive parameter scale, an offline inter-model cross-validation was conducted against a lightweight local secondary model (0.5-billion parameters).

Results:

The AI-driven pipeline achieved a 100% parsing success rate with zero data loss. Inter-model cross-validation revealed that while the 0.5-billion parameter model achieved an apparent accuracy of 92%, it yielded a Cohen κ of 0.0 due to majority class collapse on highly imbalanced biomedical datasets. This mathematically proved the necessity of the 120-billion parameter model for nuanced text extraction. As a proof-of-concept, the 120-billion parameter pipeline autonomously identified 878 Mibyo-relevant studies and stratified an adipose-derived stem cell cohort (n=182). Downstream statistical synthesis of the AI-structured data successfully identified critical translational trends, mathematically proving via logistic meta-regression that cell-free exosomes yield a significantly higher probability of systemic resilience compared with whole-cell therapies (odds ratio 4.27; P < .001), while automatically mapping unexploited delivery routes (e.g., intranasal).

Conclusions:

The proposed edge-hosted LLM infrastructure provides a secure, highly accurate, and scalable paradigm for biomedical text mining. By automating data structuration and overcoming the cognitive limitations of lightweight local models, this computational approach drastically accelerates evidence synthesis and identifies hidden translational gaps, demonstrating significant utility for metascience and next-generation medical research.