Pharmacovigilance NLP
Pharmacovigilance NLP is the use of natural language processing in the detection of adverse drug reactions.
The pharmaceutical sector is expected to further grow the spectrum of specialized drugs with an emphasis on rapidly developing treatments in the face of possible and real pandemics. In lockstep with a growing catalog of drugs is the need to monitor those drugs for adverse reactions in both clinical-trial and post-market settings.
Simply stated, the amount of available data—when compared to what is reported about these drugs—is so vast that tools must be created and used to proactively seek out adverse drug reactions through pharmacovigilance NLP.
Tenasol, as a supplemental offering performs adverse event detection on all types of data including:
HL7 C-CDA medical records,
HL7 FHIR messages,
HL7 ADT messages,
X12 EDI transactions,
unstructured data (PDF/TIF/TIFF/JPG/JPEG/GIF/RTF/TXT/DOCX)
Audio data (deconstructed to text with speech-to-text or diarization)
Video data (deconstructed to text with speech-to-text or diarization)
Medication Error vs ADR vs ADE vs Side Effect
The four scenarios can often be confused with each other:
Medication Error: A drug is incorrectly delivered to the patient. Liability falls upon either the clinician or the patient not following the instructions of the drug. Sometimes harmful, liability usually falls upon the provider or the patient.
Adverse Drug Event (ADE): Any harmful event reported by a pharmaceutical company or less often by a patient or provider. Can be caused by any number of reasons. Always harmful, liability varies by situation.
Side Effect (SE): A type of ADE, that is a known effect of taking a drug, regardless of dose.
Adverse Drug Reaction (ADR): When a drug has a direct and damaging effect, despite no dosage or delivery errors. Always harmful, liability usually falls on pharmaceutical organization.
Type A ADR: somewhat predictable and dose - dependent.
Type B ADR: completely unpredicted, and not typically dose-dependent.
While the volume of drugs and treatments dispensed per day is massive, the signals intelligence going into the results of those are of significant interest. Experts are tasked with managing data to separate what is an ADE, and what is more narrowly an ADR.
How Pharmacovigilance NLP Works
Traditionally, adverse events discovered by pharmacists, practitioners, and patients are offered the opportunity to report problems to MedWatch, a program run by the FDA. However, the volume of this data is small. Part of this is that there are low incentives for entities to report these occurrences outside of formal requirements to do so.
The alternative is straight up data mining of available data, both public and private. Pharmacovigilance NLP therefore has several steps:
Convert Data to Text: Data may come in the form of audio, video, images requiring OCR, or other formats and will need to be shifted to plain text for evaluation
Detection of the Target Drug: Drugs may be identified using various numerous identifiers (e.g. NDC or RxNorm), brand names, or generic names.
Pharmacovigilance NLP: Detection of adverse drug events using pharmacovigilance NLP based on context of the text or structured data where the drug is noted. Tenasol provides confidence of event presence for unstructured findings, and a boolean indicator if its a structured instance.
Tenasol chose to make use of a BERT NLP model that is pretrained such that confidence of events may be expressed. Transformer/LLM models do not offer this, and also present issues associated with cost volumes which can be particularly high for the volume of data processed by Tenasol.
Narrowing to Adverse Drug Reactions: This is performed by filtering out incorrect dosage scenarios, as well as side effects to constrain output to just ADR candidates.
Output: Tenasol typically returns results for pharmacovigilance NLP in XLSX or CSV. On special request, we can parse any source and return API results of ADR detection.
Tenasol Unstructured Pharmacovigilance NLP performance on SMM4H Pharmacovigilance NLP dataset.
The below example shows a screen capture of the Tenasol live demo of adverse event NLP detection. While only a few of the mentioned effects are mentioned hundreds of reactions are scanned for and seek drug associations during the transcript.
Live demonstration of Tenasol pharmacovigilance NLP during the pilot of House M.D.
Results of Pharmacovigilance NLP
Results of a pharmacovigilance program are turned into reports that can more quickly be sifted through my pharmacovigilance specialists. These signals, if strong enough are aggregated and further analyzed to evaluate their severity, frequency, and relevance.
ADR findings are then submitted to bodies such as the FDA (mentioned earlier) as well as EMA, or the WHO. If these are validated, drug labels changes may be made to include warnings, or contraindications. In extreme cases where a drug itself is at fault, a targeted or fill recall is possible.
Conclusion
Pharmacovigilance NLP is transforming how adverse drug reactions (ADRs) are detected, enabling faster, more accurate identification of potential safety concerns in both clinical and post-market settings. Tenasol stands out in this domain by offering robust capabilities to process structured and unstructured data, including medical records, multimedia files, and text documents. Its ability to handle diverse formats and deliver actionable insights—whether through structured outputs like CSV/XLSX files or API integrations—positions it as a versatile tool for modern pharmacovigilance efforts.
By leveraging datasets like SMM4H, Tenasol demonstrates its commitment to advancing ADR detection with precision and efficiency. As the pharmaceutical landscape evolves, tools like Tenasol will play an increasingly critical role in ensuring patient safety and fostering trust in new therapies. This synergy of technology and pharmacovigilance is not just a solution to current challenges but a foundation for a safer, more innovative future in healthcare.