Medication reconciliation remains a demanding patient safety process, especially when medication information must be gathered, compared and corrected across care settings. A scoping analysis published in the Journal of Medical Internet Research maps how artificial intelligence has been applied to medication reconciliation tasks and how far automation has progressed. The analysis searched MEDLINE, Embase, Web of Science, IEEE Xplore and Compendex in June 2024, then used backward citation searching to identify additional records. Resource constraints limited eligibility to English-language reports. After screening database and citation records, 94 studies met the inclusion criteria. The overall picture is narrow but useful for healthcare leaders: AI work in this area concentrates heavily on creating a best possible medication history, while discrepancy identification receives little attention and discrepancy resolution remains unaddressed.
Automation Centres on Medication History
Medication reconciliation includes three core tasks: creating a best possible medication history, identifying discrepancies between medication lists and resolving those discrepancies. Current AI applications focus almost entirely on the first stage. All 94 included entries addressed subtasks related to creating a best possible medication history, mainly by extracting or verifying medication information. Only two applications also addressed discrepancy identification, and none addressed discrepancy resolution. As a result, the highest automation stage reached was information analysis, while almost all applications automated only information acquisition. The distribution leaves later reconciliation decisions outside the main focus of current automation work.
Must Read: Electronic Medication Systems Bring Mixed Safety Gains
This concentration reflects the nature of available inputs and the maturity of current AI applications. Medication histories often require information from clinical notes, pharmacy records, structured lists, images and other inputs. AI tools can support part of that process by extracting medication names and associated details, but end-to-end automation remains rare. Only one method had potential relevance to automating best possible medication history creation in a broader manner. Limited exchange of medication data between electronic health records and pharmacies, incomplete records and outdated information can restrict what AI tools are able to use at the point of care.
Data Inputs and Methods
Clinical notes in electronic health records were the dominant input, used in 67 of the 94 entries. Image-based data also played a notable role, appearing in 21 entries. These images included pills, blister packs, medication bottles and handwritten prescriptions. Other inputs appeared much less often, including pharmacy records, patient-physician conversation transcripts, structured prescription lists in the electronic health record, user-generated structured data, user-generated notes and a medical internet forum. Pill image work generally did not explicitly report validation on patient-submitted images, although several applications incorporated variations in lighting, zoom and background to simulate real-world conditions.
The included applications used several machine learning methods, often in combination. Conditional random fields, recurrent neural networks, convolutional neural networks, transformers and support vector machines were among the most common approaches. Text-based applications typically used conditional random fields, recurrent neural networks, support vector machines and transformers. Image-based work mainly used convolutional neural networks. Because task definitions, input types and evaluation methods differed, performance results were not directly comparable. Many text-based applications used named entity recognition to identify medication details such as dosage, route of administration, frequency and indication. A smaller group linked these details to specific medications, while a more limited subset focused on medication changes and their context. Evaluation metrics often followed the modality, with text tasks using precision, recall and F1-score.
Development Remains Separate from Practice
Research maturity remains limited. Ninety-three of the 94 entries were retrospective model development work, and only one had progressed to a usability stage. Most also relied on publicly available benchmarking datasets, including i2b2 2009, n2c2 2018, n2c2 2022, MADE 1.0 2018 and National Institutes of Health National Library of Medicine Pill Image Recognition datasets. These datasets helped drive medication information extraction work, but they also reflect a development environment that remains separate from routine medication reconciliation workflows. For healthcare decision-makers, this distinction matters: a model that performs well on a benchmark does not necessarily demonstrate usability, interoperability, safety or operational value in a live clinical setting. Health system deployment therefore remains largely untested within the available evidence base.
Discrepancy identification and resolution remain the least developed areas. The two applications that addressed discrepancy identification took different routes. One compared medications extracted from free-text clinical notes with medications extracted from structured discharge prescriptions and labelled items as matched or discrepant. Another used a web-based application that retrieved electronic health record medication lists, allowed patients to confirm or add medications and flagged differences between AI-extracted and patient-entered medication lists. No included application addressed discrepancy resolution, which may require information about clinical intent, timing and the nature of medication changes. Multiple data types may be needed, including clinical notes, structured orders, medication timelines and professional communication patterns.
AI-based medication reconciliation remains at an early stage. Current work mainly automates the collection of medication information for best possible medication history creation, while discrepancy identification is uncommon and discrepancy resolution is absent. The next phase requires more than model refinement. Usable, well-integrated tools need evaluation in clinical environments, with attention to data completeness, interoperability, workflow fit and trust. Implementation evidence therefore remains central to adoption decisions. Automation is most relevant when it supports clinicians within complex medication reconciliation processes rather than replacing the judgement required to resolve discrepancies safely.
Source: Journal of Medical Internet Research
Image Credit: iStock
References:
Tabja Bortesi JP, Becerra MP, Ranisau J et al. (2026) AI-Based Automation for Medication Reconciliation: Scoping Review. J Med Internet Res;28:e86760.
