AI Learns to Read Pills: Systematic Review Maps the Future of Medication Safety

Every year, medication errors and adverse drug events harm countless patients in hospitals, pharmacies, and homes around the world. A pill picked up in dim light, an imprint too faint to read, a blister pack stripped of its label — these small failures of recognition can cascade into serious clinical consequences. A new systematic review published in Frontiers in Digital Health now offers the most comprehensive picture yet of how artificial intelligence is being trained to see, read, and verify medications, and it arrives at a striking conclusion: the technology is advancing rapidly, but the evidence base behind it still lags far behind its promise.

The review, conducted in accordance with the PRISMA 2020 reporting guidelines, set out to map the entire landscape of AI-based medication recognition. The research team searched four major bibliographic databases — Scopus, Web of Science, IEEE Xplore, and PubMed/MEDLINE — for studies published between 1 January 2016 and 4 September 2026. Their search strategy combined two conceptual blocks, one covering medication recognition tasks and the other covering artificial intelligence methods, ensuring that no relevant corner of the field was overlooked. From 895 retrieved records, the reviewers removed 136 duplicates and screened 759 records, ultimately including 91 studies in the final synthesis.

The rigor of the screening process is itself notable. Two independent reviewers assessed titles and abstracts, achieving 84.72 percent observed agreement with a Cohen’s kappa of 0.527 — a moderate level of concordance typical for early-stage screening, where borderline decisions are common. The methodological quality assessment was far more precise. Both reviewers independently applied eight assessment domains, labeled D1 through D8, to all 91 included studies before resolving disagreements through consensus. Across 728 domain-level assessments, observed agreement reached 97.66 percent, with an aggregate Cohen’s kappa of 0.959, and domain-specific kappa values ranged from 0.857 to a perfect 1.000. In plain terms, the two reviewers almost never disagreed about the quality of the evidence, which lends considerable weight to the review’s conclusions.

What did those 91 studies reveal? The field has organized itself around several core computational tasks. Object detection systems locate medications within images — a pill on a countertop, a vial in a cluttered drawer, a blister pack in a patient’s hand. Classification systems then assign those detected objects to specific medications, distinguishing a lisinopril tablet from a levothyroxine one by shape, color, and embossing. Optical character recognition, or OCR, extracts text from imprints, labels, and packaging, a task made fiendishly difficult by the tiny fonts, curved surfaces, and partial obstructions characteristic of real-world medication photography. Increasingly, the most capable systems are hybrid and multimodal architectures that fuse visual features with textual and contextual signals, mirroring how a pharmacist might combine what a pill looks like with what is printed on it and what the patient says it is.

Perhaps the most forward-looking trend the review documents is the move away from brute-force deep learning toward data-efficient paradigms. Traditional convolutional networks demand thousands of labeled images per medication class, an untenable requirement when formularies contain tens of thousands of distinct products. The included studies therefore increasingly employ few-shot, low-shot, and zero-shot learning, in which models recognize medications they have never explicitly been trained on by leveraging pretrained representations or semantic relationships. Incremental learning approaches allow systems to absorb newly approved drugs without catastrophic forgetting of older ones — a critical property in a pharmacopeia that changes every year. These techniques transform medication recognition from a closed-world classification problem into an open-world recognition challenge, which is precisely how the clinical reality behaves.

The applications described in the literature extend well beyond the pharmacy counter. Assistive technologies aim to help visually impaired and older adults identify their medications independently, converting a camera image into a spoken medication name. Mobile applications put recognition in the pocket of any smartphone user, while Internet-of-Things-enabled systems embed recognition into smart pill dispensers and connected medicine cabinets that can verify the right drug at the moment of dispensing. In clinical settings, medication recognition feeds directly into verification workflows and clinical decision support, cross-checking what a clinician holds against what the electronic record says should be there. Together, these use cases position AI recognition not as a laboratory curiosity but as an emerging digital informatics component of medication safety infrastructure.

Yet the review’s most sobering findings concern evaluation and validation. The 91 studies exhibit striking heterogeneity in datasets, imaging conditions, evaluation protocols, and validation strategies, which makes meaningful comparison between systems nearly impossible. One team may test on carefully staged photographs taken under studio lighting; another on grainy images captured by patients in their kitchens. Benchmark datasets differ in medication coverage, class balance, and annotation quality, so a model reporting 98 percent accuracy in one paper may be incomparable to a model reporting 90 percent in another. Worse, the review found that external validation, cross-dataset testing, and evaluation on unseen medications remain limited across the field. Most systems are validated on data drawn from the same distribution as their training sets, leaving open the question of whether they would survive contact with a different hospital, a different camera, or a different country’s drug formulations.

This gap between internal performance and external generalizability is not a trivial academic concern. A medication recognition system that fails on an unseen generic formulation, or that degrades under the lighting conditions of a real ward, could misidentify a drug with life-threatening consequences. The review’s authors argue that prospective deployment studies and clinical testing — evaluations conducted in actual care settings before widespread adoption — are rare in the current literature. Without them, reported accuracy figures function more as engineering milestones than as evidence of clinical safety. The field, in other words, has been optimizing the metric it can measure rather than the outcome that matters.

The review’s conclusions chart a clear research agenda for closing that gap. Future work, the authors argue, should prioritize diverse and representative datasets that capture the true variability of medication appearance across manufacturers, lighting conditions, and patient environments. Leakage control — ensuring that near-duplicate images of the same medication do not appear in both training and test sets — must become standard practice, because silent leakage inflates reported performance. Independent multicenter validation and explicit unseen-medication evaluation would establish whether systems generalize beyond their birthplace. The authors also call for explainable AI, so that clinicians and patients can understand why a system made a particular identification, and for privacy-preserving approaches that protect sensitive health images. Interoperability with existing health information systems and resource-efficient deployment — models that run on ordinary smartphones rather than data-center GPUs — round out the technical priorities.

Finally, the review reframes what success should look like. Evaluation, the authors contend, must extend beyond accuracy metrics to encompass clinical utility, safety, usability, trust, and patient-centered outcomes. A recognition system is only valuable if it prevents errors in the messy conditions of real care, if clinicians and patients actually use it, and if its failures are graceful rather than catastrophic. As AI-based medication recognition evolves from a computer vision task into a component of digital health infrastructure, the measure of progress will shift from how well models perform on curated benchmarks to how reliably they protect patients. This systematic review provides both a map of how far the field has come and a candid accounting of the distance still to travel — and it suggests that the next decade of medication safety will be decided not by cleverer algorithms alone, but by the discipline to prove, in clinics and in homes, that those algorithms deserve our trust.

Subject of Research: Artificial intelligence-based medication recognition for medication safety

Article Title: Artificial intelligence for medication recognition and safety: a systematic review of detection, classification, OCR, and digital health applications

Article References: Sai Sachin, A., & Karthikeyan, J. (2026). Artificial intelligence for medication recognition and safety: a systematic review of detection, classification, OCR, and digital health applications. Frontiers in Digital Health, 8, Article 1963607. https://doi.org/10.3389/fdgth.2026.1963607

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1963607

Keywords: artificial intelligence, medication safety, medication errors, computer vision, optical character recognition, systematic review, digital health, few-shot learning, clinical decision support, drug verification, deep learning, patient safety