Developing an operational framework for high‑severity potential drug-drug interactions

Potential drug-drug interactions (pDDIs) are highly prevalent, particularly in older adults exposed to polypharmacy. They are identified using heterogeneous sources, including Summary of Product Characteristics, regulatory reference documents, and interaction checkers (ICs), often embedded in clinical decision support systems. However, the clinical relevance of many pDDIs remains uncertain, and substantial discordance exists across sources in terms of pDDI inclusion and severity grading. In addition, the large number of low-priority pDDIs presented to prescribers by these tools contributes to alert fatigue and limits their use and usefulness. The generally low level of evidence supporting many pDDIs appears to be at least partially responsible for both these issues. Improving knowledge in this field therefore requires, in parallel, strengthening the evidence base for pDDIs and improving the way they are operationalized in clinical tools. The aim of this thesis was to develop a methodological framework to improve the identification and prioritization of high-severity pDDIs. Specifically, it aimed to: (1) quantify discordances across commonly used pDDI sources in terms of management recommendations; (2) develop a reliable gold standard for high-severity pDDIs based on the ANSM Thesaurus, the French reference document on pDDIs; and (3) evaluate and refine this gold standard through systematic comparison with a natural language processing (NLP)-based automated detection tool. Three studies were conducted to achieve these objectives. The first study developed a structured scale to classify pDDIs into four management-based priority categories and applied it to 218 real-world pDDIs identified in 1,923 patients. Six major pDDI sources were compared. Overall agreement between sources was moderate (AC1 = 0.44). Agreement improved when analyses focused only on the classification of the highest-priority pDDIs versus all others. Nevertheless, substantial heterogeneity in management recommendations across sources was observed. The second study developed a gold standard for high-severity pDDIs based on the ANSM Thesaurus. More than 30,000 unique drug pairs extracted from discharge prescriptions of 6,027 hospitalized patients aged ≥65 years were independently annotated by two experts using a predefined decision algorithm, with adjudication of discordant cases. Inter-rater agreement was very good (AC1 = 0.94). High-severity pDDIs were identified in 9.7% of patients and were strongly associated with the number of drugs prescribed. The gold standard was then used to evaluate an NLP-based detection tool (PRoSIT), which showed moderate sensitivity (63.1%) but high negative predictive value (96.0%). The third study examined discordances between PRoSIT and the gold standard. All discrepancies were reviewed to determine whether they were due to algorithmic limitations or to human annotation errors in the gold standard. Among false negatives, 81.0% were attributable to human-related errors. After correcting these errors, sensitivity increased from 63.1% to 89.4% and the negative predictive value from 96.0% to 99.2%. Most errors were concentrated in a small number of recurring pDDI clusters. Overall, this thesis shows that improving the identification of high-severity pDDIs requires both better references and advanced automated tools. Gold standards derived from reference documents should be considered evolving resources that require periodic revision. The concentration of high-severity pDDIs in a limited number of recurrent drug combinations also supports targeted educational strategies and focused alerting systems. Future work should combine incremental improvement of reference sources with outcome-based studies evaluating the clinical impact of high-priority pDDIs, in order to improve knowledge about pDDIs, source quality, and operational efficiency.