JMIR Publications

ABSTRACT

Background:

Medication errors account for a large proportion of all medical errors. In most homes, patients take a variety of medications for a long period. However, medication errors frequently occur since patients often throw away the container of the medications.

Objective:

We proposed a deep learning-based system for reducing medication errors by accurately identifying prescription pills. Given the pill images, our system locates the pills in the respective pill databases in Korea and the United States.

Methods:

We constructed the system with a pill recognition step and a pill retrieval step. In particular, the pill recognition step consists of modules that recognize the three features of pills and their imprints separately and correct the recognized imprint to fit the actual data. Our system adopts a language model as an imprint correction module to correct imprint characters. We identify the pill using similarity scores of pill characteristics with those in the database.

Results:

The experimental results demonstrate that our system achieves top-1 accuracy levels of 85.6% and 74.5% for unknown pills in two different databases. Furthermore, our system achieves 78.0% top-1 accuracy with consumer images by training only one image per pill. The results demonstrated that our system could identify and retrieve new pills without additional model updates.

Conclusions:

Our study suggests the possibility of reducing medical errors by proposing that the introduction of AI can identify numerous pills with high precision in real-time. Our study suggests that the proposed system can reduce patients’ misuse of medications as well as help medical staff focus on higher-level tasks by alleviating time-consuming lower-level tasks.