All the included studies (94/94, 100%) focused on subtasks related to the creation of a BPMH, whereas 2.1% (2/94) also addressed the identification of discrepancies (Table 1). In terms of BPMH creation subtasks, most studies (92/94, 97.9%) extracted medication information from textual or visual sources [30,40-130]. Only 2.1% (2/94) of the studies focused on verifying the accuracy of already compiled medication information, and both of these studies leveraged a collaborative filtering approach to identify potential omissions from pharmacy medication lists [131,132]. Of the 2 studies that addressed the identification of discrepancies, both identified medications that were consistent as well as inconsistent between medication lists. One study used a conditional random field (CRF) model to extract medication entities from free-text clinical notes and compared them to medications extracted from structured discharge prescriptions using string matching, term co-occurrence, and drug synonyms, labeling medications from the clinical notes as “matched” or “discrepant” [80]. The other study developed a web-based application that retrieved medication lists from the EHR, allowing patients to manually confirm the extracted medications and add missing medications. Differences between the AI-extracted medications and patient-entered medications were flagged by the application [84]. No studies addressed the resolution of discrepancies between medication lists. As such, the highest stage of information processing automation achieved by the included models and tools was information analysis, with most studies (92/94, 97.9%) only automating information acquisition; none automated decision selection (Figure 3).

The included studies used a variety of ML methods, with many using multiple methods (Figure 4). The most commonly used methods included CRFs (36/94, 38.3%) [30,41,43-45,48,52,55,57,67,68,70,71,73,75,78-81,94,95,99,100,103,109,115-117,119,123,126,127], recurrent neural networks (34/94, 36.2%) [30,40,41,45,48-52,59,61,67,68,70-74,76,78,81,85,91,92,94,95,101,103,109,119,123,125-127], convolutional neural networks (31/94, 33%) [40,42,45,49,52,59,64,72,74,76,77,81,82,87,90,92,95-97,101,102,108,112,114,119,120,122-124,127,130], transformers (25/94, 26.6%) [30,46,47,57,60,74,75,85-87,92-95,101,103-107,113,118,119,121,128], and support vector machines (20/94, 21.3%) [53,55,56,59,65,66,73,79,83,86,91,99,100,111,115,123,125-127,130]. Studies using free-text data (eg, notes in the EHR) typically used CRFs, recurrent neural networks, support vector machines, and transformers, whereas image-based studies mainly used convolutional neural networks. Owing to heterogeneity in task definitions (eg, differences in information extraction objectives and input modalities), performance metrics across studies were not directly comparable.

Among the studies that used free-text data sources (72/94, 76.6%), there was variation in the granularity of information extraction. Most of these studies (57/72, 79.2%) used ML-based named entity recognition techniques to extract a variety of medication-related elements, such as dosage, route of administration, frequency, and indication [30,41,43-53,55,57,58,60,61,63,67-71,73-75,78-81,85,91-95,99,100,103,104,106-109,113,115-119,121,123,125-128]. A smaller proportion of these studies (28/72, 38.9%) extended their approach to include relation extraction, identifying associations between these elements and specific medications (eg, linking a particular dosage to a specific medication) [30,41,43-45,47,50-52,63,73,78,79,81,87,91,93,99,100,106,107,115,116,123,125-128]. An even more limited subset (9/72, 12.5%) focused on identifying medication changes (eg, discontinuation or dosage adjustment) and the context of such changes (eg, the temporality of the change and the person responsible for the change) [46,47,57,86,103,104,106,118,121]. These studies typically assessed performance using standard evaluation metrics for natural language processing information extraction tasks, including precision, recall, and F₁-score.

Studies that used image data (21/94, 22.3%) were primarily focused on medication identification. These studies frequently analyzed images of pills [42,77,82,96,97,102,112,114,120,122,124,129,130] or blister packs [66] to identify medications or generate a ranked list of possible medications. Others applied optical character recognition to pill bottle images or handwritten prescriptions and used natural language processing methods to extract medication information [59,76,101]. Most pill identification studies (10/13, 76.9%) used single-pill images, although a few (3/13, 23.1%) addressed multi-pill image identification [77,96,97]. No studies using pill images explicitly reported using patient-generated images, although many (8/13, 61.5%) incorporated intentional variations in lighting, zooming, and background, among other aspects, to simulate photos taken in real-world conditions [77,82,97,112,120,122,124,130]. The image-based studies often used more general accuracy metrics.

Most included studies (67/94, 71.3%) used clinical notes in the EHR [30,41,43-54,56,57,60-63,65,67,68,70,71,73-75,78-81,83-89,91,93-95,98-100,103-106,109-111,113,115-119,121,123,125-128], and a significant proportion of studies (21/94, 22.3%) used images [40,42,59,64,66,72,76,77,82,90,96,97,101,102,112,114,120,122,124,129,130]. In most studies using images, the images featured pills (13/21, 61.9%), although 38.1% (8/21) of the studies used images of other medication-related items (eg, medication bottles and handwritten prescriptions) [40,59,64,66,72,76,90,101]. A few studies used other sources, such as pharmacy records (3/94, 3.2%) [92,131,132], patient-physician conversation transcripts (2/94, 2.1%) [58,107], structured prescription lists in the EHR (1/94, 1.1%) [80], user-generated structured data (1/94, 1.1%) [84], user-generated notes (1/94, 1.1%) [108], or a medical internet forum (1/94, 1.1%) [48].

Although many studies (33/94, 35.1%) used author-collected data [40,42,56,58,59,61,64-66,80,83,84,90,96-99,101,107,108,111-114,124,129,131,132], most studies used publicly available data from various benchmarking shared tasks, including i2b2 2009 (19/94, 20.2%) [48,53-55,61-63,74,79,88,89,93,95,100,110,115-117,119], n2c2 2018 (16/94, 17%) [30,41,45,49,52,71,73,85,87,91,93-95,123,127,128], n2c2 2022 (12/94, 12.8%) [46,47,57,60,61,75,103-106,118,121], MADE 1.0 2018 (7/94, 7.4%) [43,44,50,51,81,125,126], and the National Institutes of Health National Library of Medicine Pill Image Recognition (4/94, 4.3%) datasets [77,120,122,130]. The i2b2 2009 and n2c2 2018 datasets consisted solely of discharge summaries [100,134], whereas the n2c2 2022 and MADE 1.0 datasets were not restricted to a single note type [135,136]. The n2c2 2018 and 2022 datasets were annotated by at least 2 independent expert annotators per note. While MADE 1.0 also used 2 annotators, they were not independent. In contrast, the i2b2 2009 dataset was released without annotations, requiring researchers to perform their own labeling. The National Library of Medicine Pill Image Recognition dataset includes both standardized, high-quality pill images and more variable consumer-quality images to reflect real-world user-generated image conditions [137].

Most of the included studies (93/94, 98.9%) were limited to model development, with only 1.1% (1/94) of the studies advancing to a more mature stage of implementation [84].

This scoping review aimed to characterize how researchers have used AI to facilitate MedRec tasks and subtasks. Notably, all the included studies (94/94, 100%) addressed subtasks related to the creation of a BPMH, but only a handful (2/94, 2.1%) developed models or tools that facilitated discrepancy identification, and none facilitated discrepancy resolution. While there is growing interest in applying AI to support MedRec tasks, most research remains focused on demonstrating technical feasibility rather than evaluating clinical integration or real-world impact. As such, efforts to automate MedRec tasks using AI are still preliminary, with current work largely focused on a limited set of subtasks, whereas other subtasks remain unexamined.

The creation of a BPMH is a critical component of MedRec. One study of inpatient MedRec found that most unintentional medication discrepancies that had the potential to cause harm were the result of medication history errors [138], highlighting the importance of this task. In addition, BPMH creation can be time-consuming [139], emphasizing the value of automating aspects of this process.

Although all studies included in this review (94/94, 100%) developed models or tools that focused on subtasks of BPMH creation—such as medication extraction or verification—only 1.1% (1/94) developed a method that could potentially be applied to automate BPMH creation in an end-to-end manner [80]. This gap suggests that the end-to-end automation of MedRec tasks is limited not only by the state of AI but also by infrastructural barriers such as limitations to the quick exchange of medication data between EHRs and pharmacies or data incompleteness [140]. Previous work examining barriers to effective MedRec has identified the inability to access reliable medication information as a recurring challenge as users often lack access to records from external settings and the information present in EHRs may be incomplete or outdated [141-143]. These challenges may explain why so few studies (1/94, 1.1%) had progressed beyond model development to more advanced stages of implementation. Given that many models target information acquisition tasks, future research should examine what information is available to clinicians at the point of care, how these tools can augment that information, and the extent to which their use reduces cognitive and operational burden. A well-integrated medication repository that consolidates comprehensive drug histories across care settings could mitigate infrastructural challenges by improving data completeness and interoperability [144]. While such systems are not yet widely adopted or fully integrated across care settings, their expansion and improvement represent an important opportunity for future MedRec work. Efforts to update and make efficient use of health care infrastructure are key to facilitating the research, development, and implementation of tools that can scale the automation of MedRec.

The BPMH subtasks that the included studies addressed were relatively broad in scope (ie, collecting medication information from various sources and verifying the accuracy of this information). While BPMH creation is central to MedRec, its subtasks overlap functionally with other medication-related processes. This functional overlap highlights the potential for cross-task model reuse and suggests that automation efforts need not be strictly siloed within this first MedRec stage. Many of these models have broader applications beyond MedRec, including pharmacovigilance, adherence monitoring, and medication review (ie, evaluation and optimization of a patient’s medications [6]). More broadly, automating the creation of a complete medication list may enable downstream analytical use cases such as risk stratification (eg, age-related medication risks and fall risk from sedating medications), detection of potentially inappropriate prescribing or polypharmacy, and identification of deprescribing opportunities [145-147]. This broader relevance may partially explain the sustained research interest in developing these methods, particularly as public benchmarking initiatives such as the n2c2 2018 shared task on ADE and medication extraction [134] have incentivized the development of tools that support medication data processing across use cases.

The extent to which the included studies used data sources that reflect actual medication use is unclear, representing a key limitation in current BPMH automation research. The prevalence of patient nonadherence reported in the literature is often high [148], highlighting the importance of using sources of medication information that reflect actual use when developing a BPMH. The n2c2 2022 shared task on contextualized medication event extraction included identification of the “actor” responsible for medication changes (including patient-initiated changes) [135], suggesting that clinical notes may contain discernible indicators of actual use. Other data sources that may provide insights into actual medication use are patient-physician and patient-nurse conversation transcripts. As with clinical notes, such tools would require contextualized approaches to distinguish between mentions of prescribed regimens and actual use. Similarly, although pill image–based models could provide a closer approximation of actual use than health care professional-documented sources, studies have yet to validate them on patient-submitted images. Future research should focus on incorporating data sources that reflect actual medication use (eg, patient self-reported data) to improve the accuracy and completeness of the BPMH.

Some tools developed may also have relevance in virtual or remote care contexts. For instance, image-based methods for medication identification could be integrated into virtual care video visits to facilitate the creation of a BPMH (eg, by acting as an additional method to verify the medications taken by patients). This functionality could serve as a supplemental verification method, which may be useful for patients with communication barriers or low health literacy or managing complex medication regimens. As health care systems continue to expand telehealth offerings, these tools may become increasingly relevant for extending MedRec capabilities outside of traditional clinical environments.

Beyond gathering medication data to create a BPMH, later MedRec tasks often require tailored approaches to identify and resolve discrepancies. This review found one study that compared medications from clinical notes and discharge prescriptions and another that flagged discrepancies between AI-extracted and patient-entered medication lists. Only one of these studies incorporated patient-entered data as a source of medication information, reflecting actual medication use [84]. The other study that supported this task compared structured discharge prescription lists and medications extracted from free-text clinical notes [80], although it is unclear whether such notes reflect prescribed regimens, actual use, or both. The limited adoption of AI for this task may be partially related to the existence of non–AI-based electronic MedRec tools that facilitate discrepancy identification. For example, the Twinlist MedRec tool uses rule-based processing to detect discrepancies, using visual elements such as spatial grouping and highlighting to guide users in addressing these discrepancies [22]. Other tools have adopted similar strategies [149,150]. However, AI models hold promise for automating information analysis more comprehensively by enabling more robust discrepancy identification (eg, in cases in which medication information is not standardized in structure) or handling more complex medication regimens (eg, dose tapering). Given that existing tools for discrepancy identification may already be addressed in some settings, AI and health system researchers should engage with clinicians to identify what MedRec tasks or subtasks would benefit most from automation, ensuring that the tools meaningfully complement clinical workflows.

Discrepancy resolution represents a significant opportunity for AI innovation in MedRec as it requires interpreting and integrating contextual information that current tools do not address. Inconsistent resolution practices may contribute to the lack of clear safety benefits from MedRec, highlighting the need for standardized approaches to discrepancy resolution [151]. This process often involves determining whether a discrepancy is intentional or unintentional, a subtask that may require additional details such as the timing of a medication change. The recent n2c2 2022 shared task on contextualized medication event extraction suggests that these considerations are receiving increasing attention [135]. For example, in addition to helping discern actual medication use from clinical notes, automated contextualized extraction could facilitate the identification of health care professional-initiated medication changes as well as the nature of such changes (eg, stopping a medication), which may aid decision-making in light of discrepancies. Still, curating datasets to train models for discrepancy resolution may be particularly challenging, requiring granular annotations of clinical intent that are rarely captured in real-world records. Furthermore, other organizational barriers to MedRec, such as ambiguity in roles and lack of standardized workflows [152,153], may be less amenable to technological solutions, emphasizing the importance of evaluating tools in real-world contexts where such factors are relevant.

The automation of decision selection in MedRec requires improved data collection practices and focused model development. To mitigate this gap without overburdening clinicians, one potential strategy is to focus structured data entry efforts on high-risk or high-variability medications where the clinical stakes are greatest, such as high-alert medications [140]. For instance, given the risks of severe hemorrhage associated with improper use of antithrombic medications [154], this medication group could be a key target for AI-assisted MedRec. Focusing on high-alert medication groups such as these would allow for more targeted model training while preserving overall workflow efficiency. Moreover, to automate this MedRec task effectively, future models must be capable of analyzing multiple data types—including clinical notes, structured orders, medication timelines, and even health care professional communication patterns—to infer intent and recommend appropriate actions. This multimodal challenge, which extends beyond collecting the correct information, remains an underexplored but essential area for advancing toward fully integrated, AI-assisted MedRec workflows.