Improving Quality of ICD-10 Coding Using AI: Protocol for a Crossover Randomized Controlled Trial

Gastrointestinal Diseases

The goal of this randomised trial is to learn about the role of AI in clinical coding practice. The main question it aims to answer is: Can the AI-based CAC system reduce the burden of clinical coding and also improve the quality of such coding? Participants will be asked to code clinical texts both while they use our CAC system and while they do not.

Once participants are recruited, they are randomly allocated to 2 groups without allocation concealment. Allocation concealment will not be relevant for clinical coders since it is known whether a participant is assisted or not, and we will not develop a placebo coding assistant. We will, however, conceal the allocation of subjects for the analyses. In total, participants will code 20 clinical notes, where each note belongs to a single patient. The participants are asked to complete the experiment in 1 sitting without interruptions, and they cannot revisit or go back to previous notes. In the event that participants are interrupted, they are asked to exit the experiment, and any incomplete records are discarded as invalid. The user study process can be summarized in the following steps: 1. Study participants are randomly allocated to group 1 and group 2. 2. To prepare participants for the experiment, a short video tutorial is played after the consent form is signed and right before the clinical coding task commences. 3. In period 1 with 10 clinical notes, group 1 uses the control interface, while group 2 uses the intervention interface. 4. Data are logged in the background using button presses (eg. time, assigned codes, and comments). 5. Then, there is an immediate crossover to period 2 for the last 10 clinical notes. 6. Data continue to be logged in the background using button presses. 7. At the end, participants in both groups will complete the system usability scale.

Norway