Page 26 - Ebook-Book JCMS 2025

Simultaneously, extensive longitudinal cohorts have been established, including patients with T2D,
hypertension (HT), chronic kidney disease (CKD), atrial fibrillation (AF), heart failure (HF), stroke,
dementia, migraine, and those undergoing abdominal surgery. These cohorts, containing both structured
and unstructured data (e.g. electronic health records, imaging reports), have been used to emulate RCTs,
evaluate treatment effects (e.g. SGLT2i, GLP-1RA), conduct cost-effectiveness analyses informed by
real-world data, and model complex disease trajectories using both conventional statistical methods (e.g.
Cox proportional hazards models) and machine learning (ML) algorithms.
The analytical capabilities extend to advanced techniques in unstructured data analysis, including
natural language processing (NLP). This includes the development of sophisticated tools using word
embeddings, transformers (BERT, RoBERTa, PubMedBERT), and large language models (LLMs) such as
ChatGPT and Gemini. Applications range from creating automated ICD coding systems to detecting cancer
recurrence from clinical notes and extracting structured data from unstructured reports (e.g. EKGs,
echocardiograms, MRI scans).
Finally, leveraging expertise in evidence synthesis and data science, AI-powered tools for SRs
(SR-AI) are being developed to automate study selection, data extraction, and risk-of-bias assessments.
This work utilizes both commercial (ChatGPT, Google Gemini) and open-source (e.g. Qwen, LLaMA) LLMs,
offering researchers flexible options based on their preferences and resources. The goal is to
significantly expedite and enhance the efficiency of SR-conduct. This multifaceted research program has
made significant contributions to both methodological advancements and the generation of impactful
real-world evidence.
24 Joint Conference in Medical Sciences 2025