Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-2172
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-22T18:13:04.930775Z
status: published
lens: g_space
sector: health
headline: Characterizing Cancer Symptom Experiences and Related Psychosocial Concerns Through Chinese Social Media Narratives: Computational Text Analysis
dek: Background: Patients with cancer frequently experience complex and multidimensional symptom burdens. Standardized patient-reported outcome measures provide structured approaches to symptom assessment but may not capture the full range of symptom-related and psychosocial concerns expressed in patients' narratives. Patient-generated social media data offer a complementary source for examining how these concerns are expressed online.
Objective: This study aimed to characterize cancer symptom-related narratives in C…
gain_title: LLM annotation and topic modeling of patient posts identified symptom patterns, coping strategies, and 13 experiential themes that may be underrepresented in conventional symptom assessment.
problem_title: (none)
trace_subject: (none)
gain_reading: LLM annotation and topic modeling of patient posts identified symptom patterns, coping strategies, and 13 experiential themes that may be underrepresented in conventional symptom assessment.
gain_evidence: BERTopic modeling identified 13 latent themes | highlighting experiential concerns that may be underrepresented in conventional symptom assessment
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers analyzed 26,086 first-person cancer patient narratives posted on Xiaohongshu and Zhihu between November 2013 and February 2026. They used GPT-4o-mini to find six symptom domains (sleep disturbance, pain, physical function impairment, anxiety, depression, and fatigue) and BERTopic to find broader themes.

The work suggests that patient-generated social media text could complement standard symptom questionnaires by revealing concerns such as existential distress, family involvement, and financial burden. Whether these data can be used in routine clinical symptom assessment has not been established and needs further evaluation.
limitation: Differences in platform composition, uneven temporal coverage, and nonrepresentativeness of social media users limit generalizability, and the clinical utility of using such data in routine symptom assessment still requires evaluation.
tag: Evidence-backed gain
key_points: Of 26,086 posts, 13,896 (53.3%) contained at least one identifiable symptom. | Pain was the most frequent symptom (22.2%), followed by physical function impairment (15.9%) and depression (15.5%). | Pain and physical function impairment was the most common co-occurring pair (n=1407). | Coping strategies appeared in 27.2% of posts, with medication most prominent for pain and sleep disturbance, and social and psychological support more common for anxiety and depression. | LLM-human agreement for symptom extraction had mean Cohen kappa values of 0.714 and 0.754 across the six domains.
rundown: The corpus was collected with cancer-related search terms. First-person narratives were identified with GPT-4o-mini, which reached an F1-score of 81.2% against the adjudicated reference standard, while human raters showed 89.2% agreement (Cohen kappa 0.784).

BERTopic produced 13 latent themes, including cancer treatment trajectories, health care system interaction, existential distress, family involvement, and financial burden. These extend beyond the six symptom domains of the conventional framework.

Platform sensitivity analysis found a similar share of posts with at least one symptom on Xiaohongshu and Zhihu (53.2% vs 54.1%; P=.35), although the distributions of individual symptoms differed between platforms.
sources:
- peer_reviewed | Journal of Medical Internet Research | https://doi.org/10.2196/102862 | 2026-10-20
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

1c82e0f20253880d332df0601781ba6dcccb20b8bdf7869a1e550929ba8f0dd1

previous

0000000000000000000000000000000000000000000000000000000000000000