Artificial Intelligence for Post-Marketing Pharmacovigilance of Herbal and Botanical Medicines: A Scoping Review and the MIAHPV Framework

Technical Summary: Artificial Intelligence for Post-Marketing Pharmacovigilance of Herbal and Botanical Medicines

Problem Statement
The paper identifies a critical surveillance gap in the pharmacovigilance of herbal and botanical medicines. While Artificial Intelligence (AI) and Machine Learning (ML) have transformed signal detection for conventional pharmaceuticals, these methods have not been systematically applied to herbal product Individual Case Safety Reports (ICSRs) as the main exposure. The primary barrier is not the resistance of herbal data to AI, but rather the severe structural deficiencies in existing reporting databases (FAERS, VigiBase, EudraVigilance, and India's PvPI).

Three specific data quality failures render current herbal ICSRs largely incoherent for algorithmic analysis:

1. Botanical Identity Inconsistency: Reports often lack Latin binomial standardization, listing plants under vernacular, commercial, or transliterated names (e.g., Hypericum perforatum appearing under 23+ distinct strings), preventing algorithmic aggregation.
2. Co-medication Disclosure Gaps: Incomplete reporting of concurrent prescription drugs prevents the differentiation of herb-drug interactions from background adverse events or herb-alone reactions.
3. Systematic Underreporting: Driven by the perception of herbal products as "natural and safe," spontaneous reporting rates are low (pooled estimate of 1.42% adverse event reporting rates across 26 studies), and reporting often lacks the granularity required for signal detection.

Consequently, despite high rates of concurrent herbal and prescription drug use globally, AI-driven signal detection has not been successfully applied to herbal ICSRs as a primary exposure category in published literature.

Methodology
This study is a scoping review conducted in accordance with PRISMA-ScR guidelines.

• Search Strategy: The author searched PubMed/MEDLINE and the Cochrane Library for studies published between January 2010 and March 2026. The search focused on the intersection of AI/ML methods and pharmacovigilance data for herbal/botanical products.
• Constraints: The search was limited to freely available sources due to the author's status as an independent undergraduate researcher without institutional access to Embase, Scopus, or Web of Science.
• Selection: From 297 initial records, 7 studies met inclusion criteria. These studies applied AI/NLP to herbal safety data but were found to focus on pre-market risk prediction, unstructured text mining (social media/clinical notes), or supplement mention extraction, rather than AI-based signal detection on structured ICSRs.
• Framework Development: The MIAHPV (Minimum Information for AI-Assisted Herbal Pharmacovigilance) framework was developed through a two-step evidence-mapping process:
  1. Identifying data-quality requirements of specific AI methods (NLP, Bayesian disproportionality, graph neural networks).
  2. Mapping these requirements against documented deficiencies in herbal ICSRs found in the reviewed literature and regulatory frameworks.

Key Contributions: The MIAHPV Framework
The paper's primary contribution is the proposal of the MIAHPV framework, a 7-point standard designed to evaluate the "AI-readiness" of herbal ICSRs. The framework distinguishes between Discriminating criteria (which differentiate processable from non-processable reports) and Threshold criteria (necessary minimum conditions).

The seven criteria are:

1. Botanical Identity Completeness (Discriminating): Must include Latin binomial name, plant part, and preparation type.
2. Co-medication Disclosure Completeness (Discriminating): All prescription medications taken within 30 days must be listed with dosing for the primary suspect drug.
3. MedDRA-Coded Adverse Outcome (Threshold): At least one adverse event term must be coded at the Preferred Term level.
4. Reporter Type and Qualification (Threshold): Reporter must be a health professional or a consumer with explicit clinical diagnosis documentation.
5. Product Authentication Documentation (Discriminating): Verified commercial name with batch number, independent lab confirmation, or GMP-certified manufacturer documentation.
6. Time-to-Onset Plausibility (Threshold): Exposure-to-onset interval must fall within a pharmacologically plausible window.
7. Causality Pre-assessment Documentation (Threshold): Presence of a narrative causality assessment or WHO-UMC terminology.

Based on these criteria, the framework proposes a three-tier classification for ICSRs:

• Tier 1 (AI-processable): Satisfies all 7 criteria (estimated 10–20% of reports).
• Tier 2 (Partially processable): Satisfies core thresholds plus at least one discriminating criterion; requires uncertainty quantification (estimated 40–50%).
• Tier 3 (Non-processable): Lacks critical criteria; AI processing would generate noise (estimated 30–45%).

Results

• Literature Gap: The review confirmed that no published study has applied AI-based signal detection specifically to herbal product ICSRs as the main exposure. Existing AI applications in this domain are limited to pre-market toxicity prediction or unstructured text mining, not structured spontaneous reporting analysis.
• Data Quality Reality: A meta-analysis of 26 studies revealed that not one of the included studies provided complete botanical ingredient details. The scale of data incompleteness (nomenclature fragmentation, mixture attribution issues, and underreporting) is the root cause of the surveillance failure.
• Regulatory Variance: A cross-jurisdictional comparison (FDA, EMA, WHO, CDSCO) revealed inconsistent requirements for botanical coding. For instance, FAERS does not require Latin binomial names, and PvPI lacks a dedicated botanical nomenclature field, resulting in low estimated MIAHPV compliance rates across all regions (ranging from ~5% in India to ~30% in the EU).

Significance and Claims
The paper explicitly frames its significance with modesty, avoiding claims of empirical validation.

• Conceptual Tool, Not Validated Instrument: The author states that MIAHPV is a "conceptual framework derived from evidence mapping," not an empirically validated instrument. It has not been tested against a labeled dataset with known ground truth.
• Defining the "Minimum": The framework's value lies in defining a specific, actionable answer to the question: "Which reports are worth analyzing, and which will contaminate the analysis?" It shifts the field from assuming herbal data is inherently unsuitable for AI to recognizing that the data is currently unsuitable due to remediable quality failures.
• Prerequisite for Progress: The paper argues that improving AI methods is insufficient without first addressing data infrastructure. MIAHPV serves as a specification to guide future data collection and a filter to prevent AI pipelines from processing noise.
• Future Agenda: The author proposes that the next steps involve prospective validation of the framework against national datasets (e.g., FAERS, PvPI) and a Delphi consensus study to refine the criteria.

In summary, the paper argues that AI-assisted pharmacovigilance for herbal medicines is currently impossible to execute reliably due to data quality failures. The MIAHPV framework provides the first structured standard to assess data readiness, serving as a necessary prerequisite for any future successful application of AI in this domain.