Artificial Intelligence Augmented Training in Skin Cancer Diagnostics for General Practitioners

Statistics and Results Using data from the E-app, scores on the MCQs and the diagnostic accuracy of the GP's a dose/response relationship is calculated, in order to answer the research questions posed above.

Aim The aim of this project is to investigate the dose/response effect on GPs' proficiency from AI augmented training and clinical feedback in skin cancer diagnostics. This project will examine how much training is needed before the GPs' ability to diagnose and correctly refer skin cancer is affected.

Research Questions In a population of Danish GPs what is the dose/response effect of hours spent with an educational platform that offers AI augmented training and clinical feedback on their diagnostic accuracy and accurate clinical management (treatment, dismissal, referral)?

Does access to an educational platform that offers AI augmented training and clinical feedback increase the number of malignant skin lesions referred by Danish general practitioners without simultaneously increasing the number of incorrect benign referrals? Can the participating GPs clinical accuracy be predicted from the MCQ-score by comparing their quiz answers and diagnostic accuracy on their registered lesions with their score on the MCQ?

Design This stratified superiority RCT will include 90 GPs in Denmark. The study period is five months. Three months of active intervention, then two months of post intervention observation period.

Participating GPs must give signed consent before receiving a short introductory course regarding the use of hard- and software. They are then asked to complete a short questionnaire regarding the use of the platform as well as completing a proficiency test using a multiple choice questionnaire (MCQ) at baseline and again after 1 and 3. The MCQ consists of 12 skin lesion cases randomly sampled from a test library developed and validated by our research group.

Stratification and randomization:

Eligible participants will be stratified based on gender, results on the pre-test, age, type of clinic and self-reported years of experience diagnosing skin lesions, before being randomized (allocation ratio 3:1) into either the intervention or control group. Intervention All participating doctors will be asked to register all skin lesions examined due to a suspicion for non-melanoma or melanoma skin cancer, raised by the GP or patient, throughout the study period (Figure 2). Skin lesions will be registered using the R-app that enables photo acquisition and registration of: patient social security number (CPR), location of lesion, tentative diagnosis and chosen clinical action (referral, excision, monitoring, none) ect.

The intervention consists of AI augmented training and clinical feedback through an educational smartphone app (E-app), whilst the control group registers the lesions but otherwise continues its standard clinical practice.

The E-app has two modules:

The training module includes AI enhanced case training on a library of 10,000+ benign and malignant skin lesion cases each coupled to written learning modules. Clinical feedback module gives the user diagnostic feedback on all cases registered in the registration module. Feedback during the trial will be based on either histopathology or the consensus agreement of domain experts (if no biopsy is taken). Feedback on referred or dismissed skin lesions will be provided by independent experts in skin cancer diagnostics ( >10 years of experience) through a web-based review system developed by our group. Statistics The mean improvement in clinical and MCQ-tested diagnostic accuracy measured will be compared between intervention and control groups using independent t-tests. Clinical accuracy will be determined using histopathology and/or expert consensus as gold standard. Both per-protocol and Intention To Treat analysis will be calculated in order to elucidate potentially missed differences and to allow readers to interpret the effect of the intervention.

Power and Sample Size To our knowledge there is no data within the literature that describes the typical number of skin lesions seen by a Danish GP. Based on estimates from clinical experience roughly eight patients are seen by each GP per week with skin lesions suspected of skin cancer. With 90 GPs registering 90% of all lesions this amounts to 8.417 skin lesions registered during the three months. However, the ratio of benign to malignant skin lesions (BMR) referred or removed by GPs is typically somewhere between 10-29:1. Specialised centers in the UK have reported that it is possible to achieve accurate melanoma diagnostics with a BMR of 1.74-6.3:1, based on clinical evaluation aided by dermatoscopy alone. Based on these numbers an effect size of Cohen's d = 0.5-0.8 is realistic. This would require around 25-40 GPs in each group to achieve a minimum 80% power.

Ethical considerations Patient participation contains no immediate strain or discomfort for the patient, and no change to current clinical practise, as dermoscopic evaluation is part of the clinical examination of skin lesions. The images captured are stored safely and anonymously with no risk for the patient.

Participating patients receive current standard of practice or an additional teledermatological evaluation of their skin lesions, giving them an expert opinion rather than only the opinion of their GP, in cases where patients are not referred to a dermatologist. A setup with minimal discomfort for the patient and the evaluation of their skin lesions by an expert justifies the participation of the patients.

With the current low diagnostic accuracy of GP's the educational nature of the intervention justifies the study for the sake of all future patients with skin lesions that are less likely to be misdiagnosed.

Educational interventions on doctors do not require approval by The National Committee on Health Research Ethics in Denmark. However, a waiver has been obtained from the The Committee on Health Research Ethics in the Capital Region of Denmark (case number H-20059977) and ethical considerations have been made and the project is in concordance with the Helsinki Declaration II.