Professional · Issue 2 · 2026-10-08
Algorithmic Detection of Intentional Medication Discontinuations: A Step Toward Scalable Deprescribing Research
Editorially reviewed research summaries, context, and limitations.
Glenn K Goodrich, Elizabeth A Bayliss, James Lagrotteria, et al.
Summary: Goodrich and colleagues developed and validated an algorithm combining NLP-style text string searches of clinical documentation with medication order data to identify clinically intended medication discontinuations across five drug classes in adults older than 65. Using a manually reviewed gold-standard cohort of 1,588 individuals, the full algorithm achieved sensitivity ranging from 74–82% and specificity from 75–95% depending on drug class. The combined approach consistently outperformed either text search or discrete order data used alone.
Why it matters: Deprescribing research is hampered by a fundamental measurement problem: administrative claims data cannot reliably distinguish an intentional clinician-directed discontinuation from a lapse in adherence, a coverage gap, or patient non-compliance. If that distinction cannot be made at scale, observational studies of deprescribing effects are severely confounded at the exposure-definition level. This work offers a practical, EHR-based method to partially resolve that ambiguity, which could meaningfully improve the quality of real-world deprescribing evidence generated from large health system datasets.
Expert perspective: The methodological contribution is genuine and the problem being solved is real. Claims data — the workhorse of pharmacoepidemiology — is nearly blind to clinical intent. This paper takes a reasonable engineering approach: combine simple but carefully tailored text strings with structured order data, validate against manual review, and report honest performance metrics.
Several aspects deserve scrutiny. The gold-standard cohort (n=1,588) is modest in size, and the per-drug-class validation samples are small — particularly bladder antimuscarinics (n=80) and antithrombotics (n=160). Performance estimates from samples this small carry wide confidence intervals not reported in the abstract, making it difficult to judge how stable the sensitivity and specificity figures actually are. Small validation samples also increase the risk that algorithm tuning overfits to idiosyncratic documentation patterns in one health system.
The text-string approach — essentially keyword matching rather than large-language-model NLP — is deliberately simple. Simple approaches tend to be transparent, reproducible, and portable. However, drug-class-specific tailoring was required, meaning the approach is not fully generalisable without additional development work for each new medication category or clinical context. Documentation culture, EHR vendor, and clinician habits vary substantially across health systems, and external validation is the obvious next step and a meaningful current limitation.
Sensitivity values in the 74–82% range mean that roughly one in five to one in four intended discontinuations will be missed. For a study of deprescribing effects on mortality, missing 25% of exposures would attenuate effect estimates substantially. Researchers using this algorithm should think carefully about how residual misclassification would bias their particular analysis and whether sensitivity analyses or probabilistic exposure models are warranted.
From a clinical informatics and registry standpoint, the study offers a useful template: define a clinically meaningful construct that cannot be captured with structured data alone, build a gold-standard manually reviewed reference set, develop an algorithm combining structured and unstructured data, and report calibrated performance metrics. This is useful, methodologically careful, incremental progress — a credible foundation, not a solved problem.
Relevance: Clinical Informatics and Health Data Methods
Source: https://pubmed.ncbi.nlm.nih.gov/42593400/
---