Identifying Candidate Risk Factors for Prescription Drug Side Effects Using Causal Contrast Set Mining

Jenna Reps, Zhaoyang Guo, Haoyue Zhu, Uwe Aickelin

TL;DR

Researchers have developed a new data-mining pipeline that identifies risk factors for rare drug side effects by combining "market basket" analysis with rigorous statistical filtering. By applying this to a massive database of UK medical records, they successfully identified known triggers for renal failure in patients taking aminosalicylates, demonstrating a scalable way to complement expensive clinical trials.

What Problem Does This Paper Address?

In medicine, Randomized Clinical Trials (RCTs) are the gold standard for proving cause and effect. However, RCTs are often too small or too short-lived to catch rare, long-term side effects. While we have massive "longitudinal observational databases" (electronic health records), using them for causal inference is a nightmare. Because the data is collected passively, it is riddled with confounding—for example, if older patients are more likely to take a drug and more likely to have kidney issues, a simple analysis might wrongly blame the drug for the kidney problems.

The authors aim to bridge this gap by creating a method that can efficiently sift through millions of records to find "causal contrast sets"—combinations of factors that significantly increase the risk of a specific adverse drug reaction (ADR) while accounting for common biases like age and gender.

Methods

The authors propose a four-step "Causal Contrast Set Mining" methodology:

  1. Database Partitioning: They split patients who took the target drug (in this case, 5-ASAs) into two groups: those who experienced the side effect (renal failure) within a month and those who did not.
  2. Association Rule Mining: Using an algorithm typically used for "market basket analysis" (e.g., "people who buy milk also buy bread"), they identify sets of medical events or other drugs that appear frequently in the group that suffered the side effect. They calculate a Support Ratio to see which factors are much more common in the "side effect" group than the "safe" group.
  3. Logistic Regression Filtering: To ensure these factors aren't just appearing because of age or gender, they run a logistic regression. This model treats the side effect as the dependent variable and tests the interaction between the drug and the candidate risk factor, while adjusting for covariates.
  4. Ranking: Finally, they rank these candidate risk factors by their p-values. The lower the p-value, the more confident they are that the factor is a genuine risk enhancer.

Datasets

The study utilizes The Health Improvement Network (THIN) database.

  • Scope: A massive UK-based longitudinal database containing medical and prescription records for over 3.5 million active patients across 600 general practices.
  • Data Types: It includes demographics (age, gender), medical diagnoses (using the hierarchical READ code system), and drug prescriptions (using British National Formula codes).
  • Preprocessing: The authors notably excluded the first year of a patient's record at a new practice to avoid "historical event dropping" bias.

Key Insights and Ideas

  • The "Filter" Philosophy: The authors emphasize that this isn't a tool for definitive proof, but rather an ADR risk factor filter. It’s designed to narrow down thousands of possibilities to the most likely candidates for further clinical study.
  • Interaction is Key: The methodology doesn't just look for things that cause renal failure; it looks for factors that make the drug-side-effect link stronger.
  • Leveraging Hierarchy: By using the READ code system's tree structure, the algorithm can look at broad categories (like "Infection") or specific diagnoses (like "Tuberculosis"), allowing it to find patterns at different levels of granularity.

Key Results

The researchers tested their method on aminosalicylates (5-ASAs), drugs used for inflammatory bowel disease known to cause rare renal failure (0.17 cases per 100 patients/year).

  • Validation of Known Risks: The algorithm successfully flagged diuretics (specifically Furosemide) and dehydration (indicated by the use of Loperamide, an anti-diarrheal) as major risk factors. Patients on Furosemide were 3.7 times more likely to experience renal failure after taking 5-ASAs.
  • General Susceptibility: The top-ranked risk factor was Hypertensive disease (p-value: $1.62 \times 10^{-30}$). This suggests that patients who already have a higher baseline risk for kidney issues are significantly more vulnerable to the drug's specific side effects.
  • New Leads: The results also highlighted potential links with influenza vaccinations and certain antibiotics (Amoxicillin), providing new hypotheses for epidemiologists to investigate.

Limitations and Future Work

  • Direct vs. Indirect Factors: A major limitation is the inability to distinguish between a direct risk factor and an indirect one. For example, a drug might be flagged not because it causes the side effect, but because it is used to treat a condition that does.
  • The 5% Threshold: The study used a minimum support threshold of 0.05 (meaning a factor must appear in 5% of cases). While this reduces noise, it might miss extremely rare but potent risk factors.
  • Future Directions: The authors plan to develop ways to automatically "tune" this support threshold and apply the methodology to a wider range of prescription drugs to test its robustness across different medical domains.

Summary updated

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.
BibTeX
@inproceedings{reps2015identifying,
  author = {Jenna Reps and Zhaoyang Guo and Haoyue Zhu and Uwe Aickelin},
  title = {{Identifying Candidate Risk Factors for Prescription Drug Side Effects Using Causal Contrast Set Mining}},
  booktitle = {HIS},
  year = {2015},
  url = {https://openreview.net/forum?id=GM7PJOyCs0},
}

Loading related content...

Figure