Practical Randomized Controlled Trial of Artificial Intelligence for Melanoma Diagnosis (PRACTA-MEL) - PROJECT SUMMARY/ABSTRACT To reduce patient distress, combat rising healthcare costs, and improve overall patient outcomes, there is pressing need to minimize number needed to biopsy (NNB) and maintain our melanoma detection rate. The potential for curing melanoma through surgery alone, particularly at its earliest stages, underscores the importance of early detection to reduce morbidity and mortality. However, the clinical diagnosis of melanoma is challenging, resulting in dozens of unnecessary skin biopsies performed for every melanoma identified. Algorithms have shown the potential to outperform and enhance the competence of expert dermatologists in classifying dermoscopy photos by lesion diagnosis in controlled settings and pilot studies. Preliminary data from single-center studies at Memorial Sloan Kettering Cancer Center (MSK) and Stanford University (SU) indicate that utilization of artificial intelligence (AI) as a second opinion in lesions indicated for biopsy by the clinician can improve the NNB; however, prior to implementing AI in clinical practice, deployment to the clinical setting must be rigorously tested in order to assess and address potential risks of AI implicit biases. This proposal aims to close this evidence gap and build real-world data through 2 specific aims: (1) determine the potential benefits and barriers to technology adoption by performing a mixed qualitative and quantitative methods analysis of clinician surveys, followed by implementation of a mitigation framework to overcome potential barriers; and (2) determine the impact on NNB and melanoma detection rate with real time AI feedback by executing a dual center phase 2 randomized controlled trial (RCT) at MSK and SU, which will be the first RCT of dermoscopy-based AI in clinical practice. This trial will also maintain its focus on addressing potential AI biases by emphasizing diverse recruitment, measuring how the trial team is doing both in performance and in recruitment, and oversight through a steering committee of experts in AI bias with patient advocate engagement. The RCT will compare standard of care dermatology visits for skin checks of lesions suspicious for melanoma (control arm) to visits that are the same in every regard except for the addition of AI-assistance, representing an enhanced standard of care (intervention arm). Through our partnership with Canfield Scientific, the AI algorithm on study will be seamlessly integrated into clinics’ imaging process, enabling our research team to feasibly analyze a total of 18,000 lesions on the RCT, giving our study significant power and generalizability. We will also make all images acquired in the study public for future benchmarking. Our AI tool will be assessed for bias against Fitzpatrick Skin Type, which will be metadata that is also made public to the International Skin Imaging Collaboration (ISIC) Archive for future work. This project will lay the groundwork for the design and implementation of phase 3 multi-center clinical trials to strengthen the evidence base for this cutting-edge technology in a move toward widespread clinical adoption.
