Abstract
Big longitudinal observational databases present the opportunity to extract new knowledge in a cost effective manner. Unfortunately, the ability of these databases to be used for causal inference is limited due to the passive way in which the data are collected resulting in various forms of bias. In this paper we investigate a method that can overcome these limitations and determine causal contrast set rules efficiently from big data. In particular, we present a new methodology for the purpose of identifying risk factors that increase a patients likelihood of experiencing the known rare side effect of renal failure after ingesting aminosalicylates. The results show that the methodology was able to identify previously researched risk factors such as being prescribed diuretics and highlighted that patients with a higher than average risk of renal failure may be even more susceptible to experiencing it as a side effect after ingesting aminosalicylates.
Introduction
In modern medicine, Randomized Clinical Trials (RCTs) are the gold standard for establishing causality. However, RCTs often lack the scale or duration necessary to detect rare or long-term adverse drug reactions (ADRs). While massive longitudinal observational databases (LODs), such as electronic health records, offer a potential solution, they are fraught with confounding biases. For instance, age-related susceptibility might be mistaken for a drug's side effect if older patients are more likely to be prescribed the medication. This research introduces a scalable data-mining pipeline designed to identify "causal contrast sets"—combinations of factors that significantly elevate the risk of a specific ADR while adjusting for common covariates like age and gender.
Methodology
The proposed "Causal Contrast Set Mining" methodology follows a four-step process to filter potential risk factors from massive datasets:
- Database Partitioning: Patients prescribed the target drug (e.g., 5-ASAs) are divided into two cohorts: those who experienced the side effect (renal failure) within a specific window and those who did not.
- Association Rule Mining: Using techniques similar to market basket analysis, the algorithm identifies medical events or co-prescribed drugs that appear frequently in the "side effect" group. A Support Ratio is calculated to highlight factors disproportionately present in the affected group.
- Logistic Regression Filtering: To account for confounding, a logistic regression model tests the interaction between the drug and the candidate risk factor, adjusting for age and gender.
- Ranking: Candidate factors are ranked by their p-values, prioritizing those with the strongest statistical evidence of being genuine risk enhancers.
Data Source
The study utilized The Health Improvement Network (THIN) database, a comprehensive UK-based longitudinal record containing data for over 3.5 million active patients. The dataset includes demographics, medical diagnoses (READ codes), and drug prescriptions (British National Formula codes). To mitigate "historical event dropping" bias, the researchers excluded the first year of each patient's record at any new practice.
Results and Discussion
The methodology was validated using aminosalicylates (5-ASAs), which are known to cause rare renal failure. The algorithm successfully identified several key risk factors, aligning with existing clinical knowledge and providing new insights.
As shown in the results table, Hypertensive disease emerged as the top-ranked risk factor ($p = 1.62 \times 10^{-30}$), suggesting that general susceptibility to kidney issues significantly increases the risk of drug-induced renal failure. The analysis also confirmed diuretics (specifically Furosemide) as a major trigger; patients on Furosemide were 3.7 times more likely to experience the side effect. Additionally, indicators of dehydration, such as the use of the anti-diarrheal Loperamide, were flagged.
The "filter" philosophy of this approach is crucial: it is not intended to provide definitive proof but to narrow down thousands of variables to a manageable set of high-probability candidates for further epidemiological or clinical investigation.
Conclusion
This research demonstrates that combining association rule mining with rigorous statistical filtering can effectively identify risk factors for rare ADRs in large-scale observational data. While the method currently faces challenges in distinguishing between direct and indirect factors (e.g., a drug vs. the condition it treats), it offers a powerful tool for drug safety surveillance. Future work will focus on automating threshold tuning and applying the pipeline to a broader range of medications to further validate its robustness.
Summary updated