AI-assisted decision making – an audit study

n/a

We use an online audit experiment to study the impact of AI-assisted advice under various conditions (independent variables) on the quality of healthcare (outcome) on the “Chunyu Doctor” platform, where online physician consultations are common and the availability of AI consultation for patients has grown recently. Additionally, attending physicians constitute the foundation of the platform’s online consultation service.

The unit of analysis is the physician who conducts the online consultation on the platform. We use the Internet to send a number of standard patients to physicians, completing the consultation and collecting behavioral data on physicians’ responses to various treatments. The experiment implements a 2x2 design that adjusts the accuracy of AI-assisted diagnosis (accuracy-treatment) and whether or not patients’ screenshots of their chat histories with the AI are shared to physicians (transparency-treatment). In addition, there is a control group that does not get any intervention. Three more questions are added after the experiment, so we can get physicians’ attitudes about using the AI tool. Due to attrition, the additional dataset will probably be relatively small.

2023-09-21

2023-10-15

Diagnosis: whether the diagnosis of physician is correct or not.
Effort: including the number of recommended items, the number of important items, and the IRT (Item Response Theory) score.

The IRT score reflects the comprehensive quality of items, and its construction can be seen in Das et al. (2016).

Treatment (if the dataset is available): whether the treatment given by physicians is correct or unnecessary/harmful.
Subjective trust (if the dataset is available): physicians’ subjective trust towards the AI.

We run a field experiment in the online healthcare market to study the effect of various AI-assisted diagnostic information on physicians’ behavior and the quality of their healthcare.

1. Case
As a common chronic disease, unstable angina is used as the study’s case since it is appropriate for the online consultation setting. And several audit studies (Sylvia et al., 2015; Das et al., 2016; Si et al., 2023) have used it as a common case. According to these studies and guidelines for the diagnosis and management of unstable angina, we construct the professional script for standard patients and the checklist with both the recommend items and important items for diagnosing this disease.

2. Sample
The experiment will be conducted in September 2023 with a sample of 250 physicians (if all physicians accept orders sent from standard patients). If we assume that the proportion of physicians with correct diagnosis in the control group is 15% (Das et al., 2016) and α=0.05, and the number of samples across all treatments is the same, we should need a sample of 165 to have an estimated power of 0.9. Therefore, the statistical power of the experiment should be sufficient.

3. Treatments
Five treatments in the experiment are shown as below.
Table 1. Treatments
Without transparency Transparency
Low accuracy T1 T2
High accuracy T3 T4
Control（T0） Without the AI-assisted advice

The AI-assisted diagnosis is angina for treatment with high accuracy, and it is cardiovascular disease for treatment with low accuracy. It is clear that angina is a more accurate diagnostic than cardiovascular illness because it only represents around 2% of cardiovascular disease in China. The handling with and without transparency depends on whether or not the physicians are given the chat screenshot with AI from patients. We implement these interventions by sending consultation orders with different information, and various treatments’ contents are displayed as follows.

Control (T0): Doctor, I'm a little tired, and my chest hurts.
Treatment 1 (T1): Doctor, I'm a little tired, and my chest hurts. Previous AI consultation with Chunyu Huiwen indicated that I may have cardiovascular disease. Could you help me see what is wrong with me? And how should it to be treated?
Treatment 2 (T2): Doctor, I'm a little tired, and my chest hurts. As can be seen in the screenshot below, previous AI consultation with Chunyu Huiwen indicated that I may have cardiovascular disease. Could you help me see what is wrong with me? And how should it to be treated?
Treatment 3 (T3): Doctor, I'm a little tired, and my chest hurts. Previous AI consultation with Chunyu Huiwen indicated that I may have the angina. Could you help me see what is wrong with me? And how should it to be treated?
Treatment 4 (T4): Doctor, I'm a little tired, and my chest hurts. As can be seen in the screenshot below, previous AI consultation with Chunyu Huiwen indicated that I may have the angina. Could you help me see what is wrong with me? And how should it to be treated?

We also add several questions after the experiments, which enable us to get physicians’ attitudes about using the AI tool. These questions are displayed as follows.
1. Doctor, I’m not sure whether to use the AI consultation tool or not, do you think the prediction of AI-assisted advice is reliable or not?
2. Doctor, I received a diagnosis with confidence (probability) from some AI tools. What is the confidence (probability) that I need to go to the hospital for a checkup?
3. Doctor, would you reduce the use of AI tool because of patients’ distrust towards it even if you believed that the AI tool was useful?

We made a stratified randomization based on physicians’ department and title.

The unit of randomization is at the individual physician level.

No

n/a

250 physicians.

250/5.

The statistical power of the experiment should be sufficient, please refer to the sample section above for an explanation.