The theoretical impact of AI-based quality evaluation of short-video health information on public cognition and treatment adherence: a case study of denosumab combined with PD-1/PD-L1 therapy for lung cancer bone metastasis.

BACKGROUND

Bone metastasis occurs in 30-40% of patients with advanced non-small cell lung cancer (NSCLC), and denosumab combined with PD-1/PD-L1 inhibitors has emerged as a promising treatment strategy. However, the "algorithmic echo chamber" effect on short-video platforms may distort patient cognition and treatment decision-making.

METHODS

A cross-sectional study was conducted using a custom-developed web crawler to collect 1,369 videos from Bilibili, Douyin, and Xiaohongshu. A total of 402 videos were included after a three-tier keyword filtering process. An AI-based evaluation system built upon the doubao-seed-1.6 model was established, integrating three international standards-Global Quality Score (GQS), Journal of the American Medical Association (JAMA) benchmark criteria, and the modified DISCERN tool-to assess multidimensional information quality. Kruskal-Wallis tests and Spearman correlation analyses were performed to explore inter-platform differences and the relationship between information quality and user engagement metrics.

RESULTS

Overall video quality was substantially below professional medical standards: the mean GQS was 2.84 ± 1.06 (56.8% of the full score), JAMA was 0.34 ± 0.57 (8.5%), and modified DISCERN was 1.55 ± 0.69 (31.0%). Significant quality differences were observed across platforms (p < 0.001, Cohen's d = 0.6-0.8): Douyin ranked highest, followed by Xiaohongshu, with Bilibili lowest. Correlation between user engagement and content quality was extremely weak (R2 = 0.004, r = 0.062), indicating substantial decoupling-high engagement did not equate to high-quality content. Medical professionals accounted for only 25.6% of content creators, while patient-generated content reached 52.2%. Evidence-based treatment information comprised merely 20.0-26.7%, whereas misleading or inaccurate claims accounted for 6.7-13.3%.

CONCLUSION

From a behavioral and cognitive perspective, the low quality of immune-oncology information on short-video platforms, coupled with algorithm-driven amplification of high-engagement but low-quality content, may exacerbate cognitive bias, potentially increasing clinical safety risks such as insufficient hypocalcemia monitoring and inadequate MRONJ prevention. Establishing a professional governance and oversight system is urgently required.