The Effect of Human-AI Uncertainty Calibration vs. AI Uncertainty Alone on the Diagnostic Accuracy of Human Experts for Skin Lesions - a Randomized Controlled Trial.
试验速览
- 阶段
- 不适用
- 状态
- 尚未招募
- 发起方
- 入组人数
- 50
- 主要终点
- Accuracy
研究概览
简要总结
The goal of this randomized controlled study is to compare the effect of a new, personalized uncertainty-aware decision model (FDM) to a standard image recognition model in improving the diagnostic accuracy while reducing diagnostic uncertainty in experienced dermatologists tasked with differentiating between melanomas, moles and other benign skin lesions. The main question it aims to answer: Is the FDM a feasible method for an improved human AI partnership in which trust is build, misdiagnoses are avoided, and uncertainty is duly introduced or reduced.
The investigators expect to see only a slight increase in collective diagnostic accuracy for both interventions as the the human participants are skilled dermatologist and thus have high accuracies pre-intervention.
The investigators expect to see a higher increase in diagnostic certainty for the FDM intervention compared to the diagnostic certainty in the Base Model intervention.
The investigators expect to see a higher amount of diagnosis changes from incorrect to correct in the FDM group compared to the Base Model group.
The investigators do not expect any learning effect during the study.
Participants will start by answering a series of training cases consisting of images of skin lesions. These are used to train their individual FDM (only for the FDM-intervention group). From here, the participants will be randomized into two arms determining which of the two interventions they are exposed to. The participants will solve each case withouth any intervention first, and this reply will act as a control.
详细描述
A detailed description of the FDM is presented in the references.
研究设计
- 研究类型
- Interventional
- 分配方式
- Randomized
- 干预模型
- Parallel
- 主要目的
- Diagnostic
- 盲法
- None
入排标准
- 性别
- All
- 接受健康志愿者
- 是
入选标准
- •Board certified dermatologists with clinical experience in dermoscopic diagnosis.
排除标准
- •Doctors who have not yet finished their specialization and dermatologists.
- •Dermatologists without clinical experience in dermoscopic diagnosis.
研究组 & 干预措施
Base Model
The study participant is presented with a patient case including patient demographics (gender, age, placement of lesion) and two lesion images: 1 overview image, and 1 dermoscopic image. They are asked first to indicate an initial diagnosis along with their self-perceived uncertainty for this specific case before they receive Intervention 1. This initial diagnosis will act as the control. Intervention 1 is AI-generated multi-class probabilities (from a model trained on a large dataset of dermoscopic and overview images similar to the ones used for testing) and only the most likely diagnosis is presented accompanied by uncertainty estimates in percent.
After the AI input, the study participant is given the chance to change their diagnosis and indicate any potential shift in uncertainty.
干预措施: Base Model (Other)
FDM
The initial diagnosis and indication of self-perceived uncertainty follows the same procedure as for Intervention 1. Intervention 2 is the most likely diagnosis accompanied by a calibrated uncertainty generated by the FDM model (i.e. trained on the study participants previous answers + the crowd annotations on the training data + the base model prediction).
After the AI input, the study participant is given the chance to change their diagnosis and indicate any potential shift in uncertainty.
干预措施: FDM (Other)
结局指标
主要结局
Accuracy
时间窗: Immediately after the intervention.
Diagnostic accuracy in differentiating between melanoma, nevus, and benign keratosis. Defined as the percentage of correct diagnoses. Ground truth is based on histopathologically verified diagnoses.
次要结局
- Uncertainty(Immediately after the intervention.)
- Cut-off uncertainty(Immediately after the intervention.)
研究者
Julie Renata Bjerremand
Principal Investigator
Copenhagen Academy for Medical Education and Simulation