Early warnings in tweets: detecting pre-retraction signals and their association with retraction timing through natural language processing and survival analysis

This study investigates the early warning signs of scientific paper retractions by analyzing X discourse before formal retraction. Leveraging natural language processing techniques—including sentiment analysis (to assess public emotional reactions) and keyword extraction (to identify topics and Red-Flag terms)—alongside survival analysis (to model time-to-retraction), we examined 1239 retracted papers indexed in PubMed (2019–2022) that received at least one tweet before their official retraction. 13 highly tweeted viral papers were detected and excluded. Co-word analysis of the tweets revealed nine distinct thematic clusters: Computational Methods & Analytical Approaches; Evidence, Journals & Research Practices; Scholarly Communication & Online Discourse; Research Practice & Misconduct Issues; Publication Ethics & Retraction Contexts; Research Issues & Conceptual Problems; Validation & Findings; Causality, Change & Methodological Patterns; and Integrity, Authorship & Reproducibility. These topics, alongside the presence of critical retraction-related terms called Red Flags, highlighted X’s role as an informal yet insightful platform for public scrutiny. The Cox proportional hazards model showed that tweet negativity and Red-Flags are significantly associated with retraction acceleration, with Red-Flags showing particularly strong effects. Tweet length was also positively associated with shorter time-to-retraction, though this effect diminished when Red-Flags were present, indicating that explicit misconduct signals outweigh discursive detail in their relationship to retraction timing. By contrast, the open-access status and journal impact factor did not exhibit significant effects once time-dependent dynamics and social media signals were taken into account. Moreover, more recent critical discussions on social media were associated with an acceleration of the retraction process, whereas isolated early tweets did not trigger prompt action. These relationships may reflect broader scrutiny patterns rather than direct causation, as multiple factors likely contribute to retraction decisions. These findings contribute to understanding how social media activity coincides with scientific self-correction processes, highlighting how digital discourse may serve as one of several indicators in research integrity monitoring. The study underscores the need for further research to examine the complex interplay between online discourse and editorial decision-making.