JMIR Publications

ABSTRACT

Background:

Osteogenesis imperfecta (OI), or “brittle bone disease,” is a rare congenital, lifelong genetic disorder characterized by bone fragility, recurrent fractures, and skeletal deformities.

Objective:

We aim to systematically evaluate the effectiveness and safety of biologics in patients with OI, integrating Artificial Intelligence (AI)-assisted assessment to enhance the rigor and efficiency of evidence synthesis.

Methods:

We conducted a systematic review and meta-analysis of trials assessing denosumab, setrusumab, teriparatide, and fresolimumab. Data were retrieved from PubMed, Embase, ClinicalTrials.gov, and the Cochrane Library up to November 30, 2024. GPT-4o was integrated into the workflow to support screening and quality assessment. Biologic interventions were compared for their effects on areal bone mineral density (aBMD) and fracture incidence. GPT-4o's performance was benchmarked against human reviewers using sensitivity, specificity, and weighted Cohen’s kappa.

Results:

Eleven trials (616 participants) were included for a systematic review, nine of which contributed to meta-analysis. In children, denosumab produced the greatest increase in lumbar spine aBMD by 25.30% (95% CI: 18.47%–32.14%) at 12 months, while setrusumab achieved up to 12.80% gain at 6 months. In adults, setrusumab yielded the highest aBMD improvement ( 9.38%, 95% CI: 6.50%–12.26%) in lumbar spine at 6 months, while teriparatide and fresolimumab showed more modest increases. However, no biologic significantly reduced fracture incidence compared to bisphosphonates. Safety profiles also varied, with denosumab associated with a high risk of hypercalcemia in children, whereas setrusumab had no treatment-related serious adverse events. AI achieved high sensitivity in abstract screening (96.8%) and full-text screening (90.9%) and reduced total screening time by over 95%. Although there was substantial agreement with humans in the quality assessment (Cohen’s kappa = 0.806), it also exhibited optimism and positional biases due to reliance on probabilistic language patterns rather than structured clinical reasoning.

Conclusions:

Denosumab and setrusumab demonstrate promising efficacy in improving lumbar spine aBMD in OI, although current evidence suggests that biologics do not provide superior fracture risk reduction compared to bisphosphonates. GPT-4o supports evidence synthesis by improving screening efficiency and quality assessment, offering a scalable solution to reduce human workload. Human oversight remains essential for tasks requiring contextual understanding and clinical reasoning.