Abstract
This proof of principle study shows the potential of text-analytics to mine clinical data warehouses to uncover 'natural experiments' such as the use of Cilostazol in CHF patients. We envision this method will have broad applications for examining difficult to test clinical hypotheses and to aid in post-marketing drug safety surveillance. Moreover, our observations argue for a prospective study to examine the validity of a drug safety warning that may be unnecessarily limiting the use of an efficacious therapy.
Introduction
Cilostazol is the primary FDA-approved treatment for intermittent claudication, a debilitating symptom of Peripheral Arterial Disease (PAD). Despite its efficacy, the drug carries a "black box" warning—the FDA’s most severe caution—advising against its use in patients with heart failure. This warning was largely based on theoretical risks extrapolated from other phosphodiesterase III inhibitors that demonstrated increased mortality. Consequently, many patients are denied a potentially beneficial therapy, and conducting randomized controlled trials (RCTs) in this population is ethically challenging.
This study seeks to provide "Practice-Based Evidence" by leveraging 18 years of unstructured clinical notes from 1.8 million patients. By analyzing real-world clinical data, researchers aimed to determine if the theoretical risks of Cilostazol manifest in actual practice, particularly among high-risk patients with congestive heart failure (CHF).
Methods: The Informatics Pipeline
The researchers developed a high-fidelity, end-to-end informatics pipeline to transform millions of unstructured clinical notes into a structured "patient-feature matrix" suitable for statistical analysis.
The pipeline involved several critical steps:
- Text Mining and Annotation: Using an optimized version of the NCBO Annotator with 22 clinical ontologies to identify diseases, drugs, and symptoms.
- NLP Refinement: Implementing NegEx and ConText algorithms to filter out negated terms (e.g., "no chest pain") and family history mentions.
- Normalization: Mapping drugs to RxNorm and clinical terms to UMLS concepts to standardize terminology (e.g., equating "heart attack" with "myocardial infarction").
- Temporal Ordering: Establishing an "index time" based on the first prescription to distinguish between baseline conditions and subsequent outcomes.
- Propensity Score Matching (PSM): To mitigate selection bias, a 1:5 matching strategy was used to create a balanced control group across 18 clinical variables.
The primary dataset was the Stanford STRIDE database (1.8 million patients), with validation performed using the Palo Alto Medical Foundation (PAMF) dataset (1.2 million patients).
Results: Challenging the Black Box Warning
The study successfully "re-discovered" known effects, such as Cilostazol's ability to reduce claudication symptoms (OR = 2.35), validating the pipeline's sensitivity to therapeutic signals.
Key safety findings include:
- No Increased Cardiovascular Risk: In the matched cohort, there was no statistically significant association between Cilostazol and major adverse cardiovascular events (MACE). Odds ratios for stroke (1.13), myocardial infarction (1.00), and death (0.86) remained near or below baseline.
- Safety in Heart Failure: Crucially, in the high-risk CHF subgroup, Cilostazol use was not associated with increased mortality or arrhythmias. This finding was consistent across both the STRIDE and PAMF datasets.
- Arrhythmia Profile: Despite theoretical concerns, the study found no increase in atrial fibrillation, ventricular tachycardia, or sudden cardiac death among users.
While Cilostazol users had higher rates of limb-related procedures (MALE), the authors suggest this reflects "confounding by severity"—the drug is typically prescribed to patients with more advanced, lifestyle-limiting PAD.
Discussion and Conclusion
This study demonstrates that large-scale text-mining of electronic health records can uncover "natural experiments" where clinical practice deviates from guidelines, providing evidence where RCTs are lacking. While the NLP pipeline was tuned for high specificity (98%) at the cost of some sensitivity (73%), the sheer volume of data allowed for robust conclusions.
The findings suggest that the FDA's black box warning for Cilostazol may be unnecessarily restrictive. These results justify a formal prospective clinical trial to re-evaluate the safety of Cilostazol in heart failure patients and highlight the potential for automated, text-based post-marketing drug safety surveillance.
Summary updated