Comparing Traditional Counting Tray With Ai Image Recognition Application In Prescription Preparing

• By: LIN, Chih kang (Cheng Ching General Hospital, China Taiwan)
• Co-author(s): Mr Chih kang Lin (Cheng Ching Hospital, Taichung City, China Taiwan)
  Mrs Hsing-Yu Hsu (China medical university hospital, Taichung City, China Taiwan)
  Mrs June-Wan Chang (Cheng Ching Hospital, Taichung City, China Taiwan)
  
• Abstract:

  Background
  Fully automated pill-dispensing machines can improve the medication dispensing process. However, such equipment is expensive and difficult for small regional hospitals or clinics to afford. In recent years, the development of artificial intelligence (AI) image recognition systems has flourished. To reduce the burden on pharmacists, it is important to find alternative solutions that are affordable, easy to obtain, and easy to operate.
  Purpose
  This study uses a mobile app that provides AI image recognition for counting medications.Comparing the time and error rate of medication dispensing by pharmacists using traditional medication counting trays versus the AI image recognition app.
  Method
  This study conducted in February 2023 in regional hospital. Five pharmacists were involved in the study, and they compared the traditional manual counting method (control) with an AI image recognition app (intervention) for counting medications. The medications have three groups, 28, 56, and 112 pills, each consisting of three types of pills: round tablets, capsules, and irregular tablets. The study compared the time required and the rate of counting errors between the two methods. The statistical analysis used a paired t-test to compare whether there was a significant difference.
  Results
  This study included 45 sets of data for each method. The average time required for the control group was 43.51 (±32.12) seconds, and the average time required for the intervention group was 33.40 (±21.04) seconds. The results showed a significant difference of 10.11 (95% CI= 3.91~16.31; p=0.002) .
  Subgroup analysis was conducted for the three groups ,28, 56, and 112 pills. The average time required for the control group was 17.33 (±2.97) seconds, 30.67 (±10.61) seconds, and 82.53 (±24.07) seconds. The average time required for the intervention group was 20.67 (±4.03) seconds, 26.67 (±9.35) seconds, and 52.87 (±25.69) seconds. The results of the tests were -3.33 (95% CI= -6.04~-0.62; p=0.019), 4.0 (95% CI= -3.20~11.20; p=0.253), and 29.67 (95% CI= 11.05~48.29; p=0.002), respectively.
  The control group had a 0% error rate, while the intervention group had a 4.4% (2/45) error rate. The cause of the errors in the intervention group was misidentification of the image due to unclear separation between the reflection and the edge of the pill.
  Conclusion
  In this study , when the pill count is less than 56, the traditional counting trays are more advantageous than the AI app, while when it reaches 112 pills, the AI app can improve the dispensing speed more than traditional counting trays. Although the error rate is higher than that of traditional methods, errors can be avoided if the causes of errors are well understood. In addition, such apps have no restrictions on drug shape and counting tray quantity, fast recognition speed, and the ability to serve as an alternative solution for fully automated pill-dispensing machines. After the end of this study, several pharmacists at our hospital downloaded and used this AI image recognition app, the potential of AI app in dispensing is undoubtable.