Studying Deprescribing Using Routinely Collected Healthcare Data: Old Challenges and New Opportunities

Abstract

In routinely collected healthcare data, deprescribing is in most cases indistinguishable from drug discontinuation. This is not a minor technical challenge; it is a fundamental limitation in most routinely collected data sources, such as dispensing data, electronic health records and registries. Deprescribing is only a subset of discontinuation, defined by its intentionality, clinical supervision and patient perspective: ‘Deprescribing is the process of withdrawal of an inappropriate medication, supervised by a healthcare professional with the goal of managing polypharmacy and improving outcomes’ [1]. Supervision by clinicians and goal-setting with patients are defining features, yet they remain invisible in routinely collected data. Any attempt to identify deprescribing without acknowledging this constraint risks conflating clinical decision-making with other reasons for discontinuation such as patient adherence issues. This paper reflects on how creative use of routinely collected data may nevertheless further our understanding of the impact of deprescribing and, on occasion, allow distinguishing deprescribing from other forms of discontinuation.

There is a need for observational studies in deprescribing research. Large-scale deprescribing trials are unlikely, as most trials depend on pharmaceutical industry funding and deprescribing lacks commercial incentive. Hence, registry studies (routinely collected administrative health care data) remain one of the few options for studying deprescribing at scale [2]. Although this function is constrained by the inability to distinguish deprescribing from other forms of discontinuation in routinely collected data, creative use of data and design can overcome some challenges.

We acknowledge calls for stricter conceptual clarity in the use of the term deprescribing, emphasizing shared decision-making and structured follow-up [3]. This definition is typically achievable in intervention studies. However, in pharmacoepidemiologic research, such detailed process elements are rarely observable. If conceptual criteria were applied rigidly, most register-based studies would be excluded from deprescribing research. We take a more pragmatic view. When the aim is to study deprescribing, even without full access to process details, it is reasonable to use the term, provided the operational definition is clearly stated and its limitations acknowledged. This allows the research to stay conceptually focused while contributing to the broader evidence base.

The challenges of identifying deprescribing are not new to the field of pharmacoepidemiology and can be summarized as below.

The key challenges of identifying deprescribing should not discourage researchers from studying this in routinely collected data. Rather, we argue that careful use of data can move the deprescribing field forward.

An alternative to the approach of enriching data is to find examples where the bias from exposure misclassification and confounding is less prevalent.

Medication management systems can reduce exposure misclassification. An example of this comes from Sweden's ApoDos system where patients receive pre-sorted, time-stamped medication pouches every 2 weeks. As prescriptions are renewed regularly, primary non-adherence is nearly impossible, and secondary non-adherence is restricted to selective pill omission. This means that when a drug is not renewed, the decision has most certainly been made in consultation with a physician, making this a robust setting for studying deprescribing in real-world data. Medication management systems with similar features that reduce the risk of non-adherence and irregular filling patterns are also used in some US nursing homes [10]. Using such data is likely to limit exposure misclassification but does not reduce confounding by indication and can reduce generalizability.

Medication shortages, that is, where a drug is temporarily unavailable, can be used as natural experiments in pharmacoepidemiological studies [11]. In deprescribing research, medication shortages could provide causal effect estimates, where the confounding effect of the indication/reason for deprescribing is removed. However, it should be noted that drug shortages are often handled by switching to alternative treatments, which makes it harder to find realistic opportunities for deprescribing researchers.

Deprescribing remains difficult to study in routinely collected data, mainly because it cannot be reliably distinguished from other forms of discontinuation unless explicitly recorded. This limitation is fundamental but should not be paralysing. Register studies are essential for generating large-scale real-world evidence on deprescribing. By combining prescribing data with care events, using structured settings like multidose systems and refining prescriber and patient-level signals, we can improve identification. Until deprescribing is routinely documented, our best option is to triangulate intent from patterns. This is not a workaround, but a necessary foundation for studying deprescribing at scale.

The authors declare no conflicts of interest.