Advances in Ontology-Based Mining of Adverse Drug Reactions

arXiv:2512.11452 · stat.ME · Submitted 2025-12-12 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Advances in Ontology-Based Mining of Adverse Drug Reactions".

Jane: Here's the gist of the paper. Post–marketing pharmacovigilance is essential for identifying adverse drug reactions (ADRs) that elude detection during pre–marketing clinical trials By integrating an adverse event (AE) ontology into a zero–inflated negative binomial model, this study improves ADR detection by accounting for biological similarities among correlated AEs and addressing the excess of zero counts The ontology–based model consistently outperforms classical models such as the Gamma–Poisson Shrinker (GPS) when applied to real data from the Veneto drug safety database.

Tom: Modeling Signal Detection The study introduces a novel approach that integrates an adverse event (AE) ontology into a zero–inflated negative binomial model to improve ADR detection.

Jane: This method is designed to account for biological similarities among correlated AEs and address the excess of zero counts, which are common in pharmacovigilance data.

Tom: The underlying statistical framework is based on the zero–inflated negative binomial (ZINB) model, which accounts for zero counts while using shared information among AEs within the same group.

Paper summary: Tom: So looking at the title, "Advances in Ontology-Based Mining of Adverse Drug Reactions," this paper is really focused on improving the signal detection process itself using better data organization.

Jane: It seems like the authors are saying that by structuring how we categorize adverse events using an ontology, and then feeding that structure into a sophisticated statistical model like the zero-inflated negative binomial framework, we get a more accurate picture of what's actually happening with drug safety.

Lu: The implication is that for post-marketing pharmacovigilance, which is essential because clinical trials miss things, we can move toward methods that look beyond just counting events and start looking at the underlying biological connections between those events.

Meng: So it’s about making the search smarter, not just running a bigger search. It’s about understanding how those related events cluster together in a way that matters for safety monitoring.

Tom: Right. The core message is that combining ontology with this specific model addresses both the issue of excess zero counts and the problem of correlated adverse events, which classical models often struggle with.

Jane: And they provide concrete evidence from their application on real data from the Veneto drug safety database showing that this ontology-based approach consistently outperforms classical models like Gamma-Poisson Shrinker.

Lu: So for the wider field, it shows that using structured representation learning—the ontology—is a powerful way to improve signal detection in complex biological surveillance data sets, especially when those reports are sparse or noisy.

Meng: It’s interesting because they don't just propose a new algorithm; they show that this specific structural approach provides measurable gains over established statistical frameworks under real-world conditions.

Tom: And the future work they mention is adding patient-level information, which means eventually we’ll be able to separate what's caused by the drug from other factors affecting a patient.

Jane: That moves us closer to a truly causal understanding, which is probably the ultimate goal when we are monitoring drugs in this way. This paper sets up a strong foundation for that next step.

Conclusion: Tom: So, to wrap up this look at "Advances in Ontology-Based Mining of Adverse Drug Reactions," we're talking about how using a structured system for adverse events helps us find drug problems faster.

Jane: Exactly. The authors are showing that by organizing those events into an ontology—a kind of map for the types of side effects—we can actually see patterns in the data that were previously hidden because the events looked too different.

Lu: What this means is that instead of just looking at every single event separately, you group them based on biological similarity, which makes finding a real drug signal much more reliable.

Meng: From an engineering standpoint, it's about making the search space smaller and more meaningful so we don't get overwhelmed by noise in the safety data.

Tom: Right. The main point is that this specific method, zGPSAO, beats older models because it handles those messy zero counts and those correlated events better than just counting everything up.

Jane: So for a listener who doesn't read papers, think of it like this: instead of looking at a pile of random rocks to see if one is valuable, you use a classification system to group the rocks by their shape and texture first.

Lu: That’s the representation learning part. The ontology acts as that shape and texture guide for the statistical model. It takes complex, messy data and puts it into a format where the computer can actually learn meaningful relationships between drugs and side effects.

Meng: But how much does this actually cost to run? If we use this ontology mapping, is the computational overhead for training these models going to be prohibitive when we scale up to millions of drug reports?

Tom: That’s a fair point, Meng. The paper mentions data thinning as a way to keep the model stable without losing information, which suggests they found a practical way to manage that complexity.

Jane: And the authors are also looking ahead, saying they plan to add patient information later on, like if someone has another health condition or takes another medicine while reporting an adverse event.

Lu: That's the next logical step. Moving from just drug-event associations to trying to figure out if a specific drug caused the event when you factor in everything else a patient is dealing with.

Tom: So, this paper shows us how to build a smarter detection system now, and the roadmap points toward building an even more complete picture of patient safety next.

Kenenisa Tadesse Dame, Pietro Belloni, Ugo Moretti, Fabio Scapini, Marco Tuccori, Alessandra R. Brazzale

Department of Statistical Sciences, University of Padova, Italy · Department of Diagnostics and Public Health, University of Verona, Italy

stat.ME

Submitted: 2025-12-12

Updated: 2026-10-30

Comments: 20 pages, 22 figures, 2 tables

Code: https://github.com/kenenitadesse/zGPS-AO-PSI

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 77/100

The gist: Post–marketing pharmacovigilance is essential for identifying adverse drug reactions (ADRs) that elude detection during pre–marketing clinical trials By integrating an adverse event (AE) ontology…

Key concepts

Zero-Inflated Negative Binomial (ZINB) Model
This statistical framework is used to analyze count data, which is common in pharmacovigilance. It specifically addresses two issues: the high number of zero counts and the correlation between different adverse events (AEs). The ZINB model helps accurately estimate these complex patterns in drug safety data.
Adverse Event (AE) Ontology
The ontology is a structured system used to categorize and define relationships between different adverse events. By using this, the model can recognize biological similarities among correlated AEs, allowing it to treat related events as a single, more informative group for detection.
Data Thinning Technique
This technique is a method used to partition data counts while preserving their essential structure. It was developed to ensure robust model validation by avoiding the information loss that occurs during random data splitting, maintaining the integrity of drug-AE pairs.

Terminology

Summary

Post–marketing pharmacovigilance is essential for identifying adverse drug reactions (ADRs) that elude detection during pre–marketing clinical trials By integrating an adverse event (AE) ontology into a zero–inflated negative binomial model, this study improves ADR detection by accounting for biological similarities among correlated AEs and addressing the excess of zero counts The ontology–based model consistently outperforms classical models such as the Gamma–Poisson Shrinker (GPS) when applied to real data from the Veneto drug safety database. p. 0

Where the paper says it

Brazzale1 1Department of Statistical Sciences, University of Padova, Italy 2Department of Diagnostics and Public Health, University of Verona, Italy Version: December 2025 Abstract Post–marketing pharmacovigilance is essential for identifying adverse drug reactions (ADRs) that elude detection during pre–marketing clinical trials. This study explores a novel approach that integrates an adverse event (AE) ontology into a zero–inflated negative binomial model to improve ADR detection.

Modeling Signal Detection

The study introduces a novel approach that integrates an adverse event (AE) ontology into a zero–inflated negative binomial model to improve ADR detection. This method is designed to account for biological similarities among correlated AEs and address the excess of zero counts, which are common in pharmacovigilance data. p. 0

Where the paper says it

This study explores a novel approach that integrates an adverse event (AE) ontology into a zero–inflated negative binomial model to improve ADR detection. By accounting for the biological similarities among correlated AEs and addressing the excess of zero counts, this method more effectively disentangles AE associations.

The underlying statistical framework is based on the zero–inflated negative binomial (ZINB) model, which accounts for zero counts while using shared information among AEs within the same group. p. 1
Where the paper says it

Explicitly integrating the ontology of the AEs into the estimation process would allow sharing the biological similarities among related AEs, which in turn enhances the accuracy of true AE signal detection while minimizing false positives. This challenge is addressed by the zero–inflated negative binomial (ZINB) model [22, 23], which accounts for zero counts while using shared information among AEs within the same group.

Ontology-Based Signal Detection

The analysis extends to groups of related AEs defined by the MedDRA ontology, allowing for aggregation of related events into separate contingency tables. The zero–inflated Gamma–Poisson shrinker with AE ontology (zGPS. p. 3

Where the paper says it

Each AE group is represented by a separate contingency table, allowing aggregation of related events. The zero–inflated Gamma–Poisson shrinker with AE ontology (zGPS.

AO) method is a hierarchical empirical Bayesian framework designed for analyzing AE data in drug safety surveillance. p. 3
Where the paper says it

The zero–inflated Gamma–Poisson shrinker with AE ontology (zGPS. AO, for short) method [28] is a hierarchical empirical Bayesian framework designed for analyzing AE data in drug safety surveillance, which addresses the two key challenges of excess zeros and overdispersion in count data.

This model utilizes a mixture of two negative binomial models to represent the marginal distribution of the count nij, which turns out to be a zero-inflated negative binomial (ZINB) model. p. 2
Where the paper says it

Instead of treating each λij as an independent parameter, the GPS model introduces shrinkage by assuming λij follows a two–component mixture of gamma distributions with parameter θ = (α1, β1, α2, β2, ω), λij ∼ ωg(λij ; α1, β1) + (1 − ω)g(λij ; α2, β2), (3) where ω ∈ (0, 1) is the probability that λij comes from the first component of the mixture and (αk, βk) > 0 for k = 1, 2, represent the shape and rate parameters of the two gamma components, while g(λ; α, β) = β α Γ(α) λ α−1 e −βλ denotes the gamma probability density function with mean α/β and variance α/β2 . By integrating over λij in model (2) using (3), the marginal distribution of the count nij , Pr(Nij = nij | θ, Eij ) = ωNB nij ; α1, Eij Eij + β1 + (1 − ω)NB nij ; α2, Eij Eij + β2 , (4) turns out to be a mixture of two negative binomial models, where NB(nij ; α, Eij/(Eij + β)) is the negative binomial probability evaluated at nij with shape parameter α and parameter Eij/(Eij + β).

The group–level reporting rate si is defined as E(λij), and a higher value indicates more evidence that the AE group is associated with drug i. p. 4
Where the paper says it

We are interested in the AE group–level reporting rate si = E(λij ) = (1 − pi)µi (8) for drug i. The higher the value of si , the more evidence we have that the AE group is associated with drug i.

Model Validation Strategies

To address challenges in model validation, the paper introduces a data thinning technique based on convolution–closed distributions to partition counts while preserving their underlying structure. This approach enables robust model validation that respects inherent relationships in SRS data. p. 1

Where the paper says it

To address this, we introduce a data thinning based on convolution–closed distributions, which partitions counts while preserving their underlying structure [24, 25]. This enables robust model validation that respects the inherent relationships in SRS data.

The comparison of partitioning strategies showed that stratified splitting and the proposed data thinning method preserved the full set of drug–AE pairs in both training and validation, avoiding coverage loss observed with random splitting. p. 9
Where the paper says it

With random sample splitting, certain AEs were randomly absent from the validation set, resulting in the exclusion of all corresponding drug–AE pairs and highlighting that random splitting can undermine model validation in a pharmacovigilance setting. In contrast, stratified splitting and the proposed data thinning method preserved the full set of drug– AE pairs in both training and validation, avoiding the coverage loss observed with random splitting.

Where the Gains Come From

The superior performance of the zGPS.AO model over GPS is derived from sharing information between similar AEs in addition to better handling of zero counts. p. 8

Where the paper says it

AO model clearly outperforms the GPS model, as indicated by a smaller MSE and larger AUC. This improvement derives from sharing information between similar AEs in addition to the better handling of zero counts.

The application to the Veneto pharmacovigilance database confirmed these advantages, as zGPS. p. 0
Where the paper says it

Statistical significance is evaluated using a permutation–based maximum statistic that preserves AE correlations within individual reports. Simulations and an application to real data from the Veneto drug safety database demonstrate that the ontology–based model consistently outperforms classical models such as the Gamma–Poisson Shrinker (GPS).

AO identified more signals, including all those detected by GPS, while improving discrimination of true drug–AE associations and reducing false negatives. p. 8
Where the paper says it

Based on these criteria, the zGPS. AO model identified 245 AE signals across the 12 studied drugs, which include the only 15 signals detected by the GPS model.

Furthermore, data thinning provides stable and reliable model performance that approaches the effectiveness of stratified splitting while avoiding biases introduced by random partitioning. p. 9
Where the paper says it

Data thinning approaches stratified performance in accuracy and AUC but shows declining sensitivity and F1 at higher ε, reflecting reduced detection of true positives. Random splitting provides stable sensitivity and relatively consistent precision but underperforms in specificity and overall reliability.

The results show that stratification consistently achieves the highest accuracy and AUC across different training proportions. p. 9
Where the paper says it

Figure 7 and Table 2 summarize the comparative performance of data thinning, stratified splitting and random splitting. Stratified splitting consistently achieves the highest accuracy and AUC, with strong specificity, making it the most balanced strategy.

Conclusion

The findings demonstrate that the zGPS.AO model enhances ADR signal detection by combining AE ontology with a zero–inflated negative binomial (ZINB) framework while addressing excess zeros and correlations among AEs. p. 9

Where the paper says it

7 Discussion Our findings show that the zGPS. AO model enhances ADR signal detection by combining AE ontology with a zero–inflated negative binomial (ZINB) framework while addressing excess zeros and correlations among AEs.

Future extensions plan to integrate patient–level covariates such as demographics, comorbidities, or concomitant medications to distinguish causal drug–AE relationships from confounding. p. 0
Where the paper says it

Keywords: Data Thinning, Empirical Bayes, Pharmacovigilance, Post–Selection Inference, Zero–Inflated Model 1 Motivation and background An adverse event (AE) is any untoward medical occurrence in a patient receiving a medicinal product, regardless of causality. Adverse drug reactions (ADRs) are a subset of AEs that are harmful, unintended and reasonably attributed to the drug.

The work confirms that data thinning provides stable and reliable model performance without losing any drug–AE information. p. 9
Where the paper says it

Data thinning approaches stratified performance in accuracy and AUC but shows declining sensitivity and F1 at higher ε, reflecting reduced detection of true positives. Random splitting provides stable sensitivity and relatively consistent precision but underperforms in specificity and overall reliability.

The original Veneto pharmacovigilance data used in this work are owned by the Regional Center for Pharmacovigilance and cannot be publicly shared. p. 8
Where the paper says it

This improvement derives from sharing information between similar AEs in addition to the better handling of zero counts. 3 Veneto drug safety data We then applied the statistical methods discussed previously to a dataset of adverse event reports from 2023 to 2024 provided by the Pharmacovigilance center of the Italian Veneto region (CRFV) for I = 12 drugs with hight reporting frequency.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve signal detection in post-marketing pharmacovigilance by integrating an adverse event ontology into a zero-inflated negative binomial model (zGPSAO method). This allows for sharing the biological similarities among related AEs, which enhances accuracy while minimizing false positives, as described in Section 3.

  2. Enhance model validation robustness by replacing conventional random splitting with convolution closed distributions to create independent training and validation datasets without data loss. This ensures that all drugs or AEs are absent from one subset is avoided, leading to more robust model evaluation.

  3. Improve signal identification through post-selection inference by implementing a data thinning technique based on convolution closed distributions. This method allows for the creation of independent training and validation datasets while retaining all drug adverse event pairs, which is superior to conventional random splitting in ensuring reliable and consistent model evaluation.

  4. Refine statistical significance testing by using a permutation-based maximum statistic that accounts for the interdependence structure inherent in spontaneous reporting data. This involves permuting AE sets rather than individual drug-AE pairs to preserve the association among AEs reported within each report, leading to a q-value adjusted for multiple comparisons.

  5. Increase predictive accuracy of AE level relative reporting rate estimates by using the empirical Bayes geometric mean (EBGM) as an alternative measure. This is derived from the posterior distribution and provides a measure that is on the same scale as the relative reporting rate, allowing for better assessment of drug safety issues than simple RRs.

  6. Improve model performance comparison by systematically evaluating data partitioning strategies using metrics like accuracy, AUC, sensitivity, specificity, precision, and F1 score across varying training proportions epsilon. This provides a comprehensive view of which strategy—data thinning or stratified splitting—offers the best balance across metrics.

Abstract

Post-marketing pharmacovigilance is essential for identifying adverse drug reactions (ADRs) that elude detection during pre-marketing clinical trials. This study explores a novel approach that integrates an adverse event (AE) ontology into a zero-inflated negative binomial model to improve ADR detection. By accounting for the biological similarities among correlated AEs and addressing the excess of zero counts, this method more effectively disentangles AE associations. Statistical significance is evaluated using a permutation-based maximum statistic that preserves AE correlations within individual reports. Simulations and an application to real data from the Veneto drug safety database demonstrate that the ontology-based model consistently outperforms classical models such as the Gamma-Poisson Shrinker (GPS). For post-selection inference, we furthermore explore a data thinning technique for convolution-closed families, enabling the creation of independent training and validation datasets while retaining all drug-AE pairs. This approach is compared with conventional random train/test splitting, which may leave some drugs or AEs absent from one subset, and stratified splitting, which requires expanding aggregated counts into individual instances. The data-thinning technique and stratified splitting yield very similar results, with stratified splitting showing a slight benefit, and both clearly outperform random splitting in ensuring reliable and consistent model evaluation.

Sources

Related papers