Social Media Listening

The foundational process of continuously harvesting publicly available posts, comments, and forum discussions from platforms like X (Twitter), Reddit, and patient community sites. Unlike structured Individual Case Safety Reports (ICSRs), social media data is highly noisy, colloquial, and lacks standardized medical terminology. Ingestion pipelines must handle diverse formats, emojis, and informal language while respecting platform API rate limits and terms of service. The goal is to create a raw, searchable corpus for downstream pharmacovigilance analytics without introducing sampling bias.

The application of specialized Medical Named Entity Recognition (NER) models to identify potential adverse event mentions within informal social media text. This task is significantly harder than processing clinical notes due to:

• Lexical Variability: 'My heart was racing like crazy' vs. 'palpitations'
• Negation and Uncertainty: 'I thought I might be having a stroke but it was just a migraine'
• Indication Confusion: Distinguishing a symptom of the treated condition from a true drug reaction Contextual embeddings from transformer models are critical for resolving these ambiguities.

The algorithmic process of linking an extracted adverse event mention to a specific medicinal product mentioned within the same social media post or thread. This requires resolving informal drug names, misspellings, and brand-generic mappings against standardized terminologies like RxNorm. A single post may mention multiple drugs and multiple events, demanding robust relationship extraction to correctly form the drug-event combinations necessary for disproportionality analysis. Co-reference resolution is essential to track which drug caused which symptom across multiple sentences.

Once product-event pairs are aggregated, statistical methods like the Proportional Reporting Ratio (PRR) and Empirical Bayes Geometric Mean (EBGM) are applied to identify combinations reported more frequently than expected. Social media data serves as a complementary channel to traditional sources like FAERS and VigiBase, potentially offering earlier signal detection due to real-time patient reporting. However, the lack of a defined denominator and inherent selection bias require careful Bayesian shrinkage to minimize false-positive signals.

An automated classification step that assesses whether a social media post describes an event meeting regulatory seriousness criteria (death, life-threatening, hospitalization, disability, congenital anomaly). This involves deep semantic analysis beyond keyword matching. For example, 'I ended up in the ER' implies hospitalization, while 'I thought I was going to die' may not meet the life-threatening threshold. Triage models prioritize posts for human review, ensuring that potential ICSRs are not missed while filtering out non-serious noise.

A critical governance layer that prevents the same adverse event narrative from being counted multiple times across different platforms or user accounts. A single patient might post the same experience on X, Reddit, and a patient forum. Probabilistic deduplication algorithms compare textual similarity, timestamps, and user metadata to identify likely duplicates. This step is essential to avoid inflating disproportionality scores and generating spurious signals from a single, widely-shared anecdote.

The systematic process of identifying new or previously unknown potential causal relationships between a drug and an adverse event from accumulated data sources. Social media listening serves as a supplementary input to traditional signal detection workflows, potentially identifying patient-reported signals earlier than structured reporting channels.

• Relies on statistical disproportionality analysis
• Social data adds real-world patient experience context
• Requires signal validation before regulatory action

The Medical Dictionary for Regulatory Activities is a clinically validated international terminology used for standardized coding of adverse event information. Social media mentions must be normalized to MedDRA terms before integration into safety databases.

• Hierarchical structure: SOC → HLGT → HLT → PT → LLT
• Enables cross-source aggregate analysis
• Essential for mapping colloquial language to clinical terminology

A formatted report containing detailed information about a single adverse event experienced by an individual patient. Social media posts that meet minimum reporting criteria—identifiable patient, identifiable reporter, suspect drug, adverse event—must be processed as ICSRs.

• Governed by E2B (R3) transmission standards
• Requires causality assessment
• Social media complicates reporter verification

Quantitative statistical methodology identifying drug-event combinations reported more frequently than expected. When applied to social media data, these methods help distinguish true safety signals from background noise in high-volume, unstructured patient conversations.

• PRR: Frequentist ratio calculation
• ROR: Odds-based disproportionality measure
• EBGM: Bayesian approach adjusting for data sparsity

The systematic evaluation of the likelihood that a suspected drug caused an observed adverse event. Social media reports often lack critical clinical detail, making causality assessment challenging without follow-up investigation.

• Evaluates temporal relationship
• Considers dechallenge/rechallenge data
• Accounts for confounding by indication
• Social posts rarely contain sufficient detail alone

Systematic global screening of published scientific and medical literature to identify ICSRs and new safety findings. Social media listening extends this concept to grey literature and real-time patient forums, capturing safety information before formal publication.

• Complements traditional database searches
• Requires source validation protocols
• EMA GVP Module VI provides regulatory guidance