Implementation of Teledermoscopy and Artificial Intelligence

Study objective: 1\. Is teledermoscopy an equally safe method as conventional care for skin cancer patients? 2. How should teledermoscopy be performed and by whom? 2. How does teledermoscopy affect current care and organization of skin diseases? 3. Can deep learning AI improve diagnostic accuracy by being a diagnostic support to dermatologists? 4. How should AI be implemented in care pathways for optimal interaction between assessors and AI? Material and methods: 1. Study setting Patients scheduled for a skin examination for a suspected skin lesion can be recruited, both in dermatology clinics and participating primary care clinics or digitally after visiting primary health clinics. 2. Data acquisition Patient data and images will be collected with the Dermicus® application. Participating PCP and dermatologists will also fill in a questionnaire about their assessment of the patient. Additional informaiton will be collected from medical records, i.e. histopathological diagnoses, and from national registries. Questionnaires will be entered in a digital data base using REDCap®. 4\. Statistical analyses 4.1. In part 1, the number of consultations send to dermatologists and for pathological analyses before introduction of teledermoscopy and during the first and second year if using teledermoscopy will be analysed. Descriptive statistics will be presented and differences between the different time periods will be tested with t-tests and paired t-tests. Additionally, other aspects of teledermatology will be examined, such as the impact on diagnostic accuracy of performing double-reader instead of singe-reader assessments as well as the propensity in primary care physicians to perform total body skin examinations when using teledermoscopy instead of traditional referals to dermatologists for the assessment of skin lesions. 4.2. Also part 2 has several substudies which focus on the impact of diagnostic accuracy and clinical management when introducing AI that perform different tasks in the teledermocopy work flow. For example, one AI will be trained to categorise all skin lesions referred through teledermoscopy in eight diagnostic categories and another will be looking at finer-grained diagnostic classes, such as subclassification of basal cell carcinoma. Measures of diagnostic accuracy will be estimated comparing dermatologists with and without access to diagnostic algorithm support. Measures reported include sensitivity, specificity, and Area under ROC-curve (AUROC). Part 2 The study will also report in detail the impact the AI has on the willingness to change a diagnosis or a management plan (in all directions). Generalized linear mixed-effects models (GLMMs) with a binomial distribution and logit link function will be used to assess the impact of AI assistance on diagnostic accuracy and management decisions. The models will include AI assistance, patient age, sex, and diagnostic confidence as fixed effects, with random intercepts for individual assessors to account for clustering of observations within assessors. 5\. Power set to 0.8 and significance to 0.05. 10% censures. 5.1. Patients recruited to this study can be used in several of the sub-studies. The aim of the study is to collect 8000 patients in total in this study. 5.2. To detect a difference in "unimaged skin cancers" between teledermoscopy and conventional care of patients 1200 cases and 2400 controls need to be included. 5.3. To detect a 10% difference in sensitivity/ specificity of diagnostic ability in PCPs before and after working with teledermoscopy 3400 patients need to be included. 5.4. To investigate how artificial intelligence should be implemented in clinical care the investigators have calculated that 6000 patients are needed to detect a 10% difference in sensitivity and specificity in the subgroups. Ethical considerations and data management: Data will be collected using Dermicus®, a CE-certified digital platform and mobile application. With the application downloaded on iPhones®, locked for any other uses, the history of the patients are registered. Then, by connecting the iPhone to a dermoscope, macroscopic and dermoscopic images are captured. All data will be stored on the servers of the health care region of Skåne, where the studies are conducted. Once a case has been created and sent to the data base all information will be deleted from the iPhone®. Additional data will also be retrieved from relevant medical records, e.g. histopathological diagnosis, and manually registered in an electronic database at a highly secure location (LUSEC/ REDCap provided by Lund University/ ShareFile provided by Region Skåne) . Data collected from PCP and dermatologists by questionnaires will also be registered in this data base by means of electronic surveys (REDCap). Information from primary care on total number of visits, referrals to dermatologists and referrals to pathology regarding skin lesions will be extracted from patient administrative systems. Age- and sex matched controls will be used for the study investigating missed skin cancer. These controls will be randomly selected from patients that was referred to a skin clinic by paper referral during the same period as the teledermoscopically referred patients were gathered. Algorithms for skin cancer diagnosis will be implemented in the web platform of Dermicus for the studies of introduction of artificial intelligence. Teledermoscopic assessors will be instructed on when and how to use these different tools. Every month the newly entered data will be checked for completeness, and in the case of missing data, reminders to participating investigators will be send. When the data sets are complete, identifiers (such as personal identification number) will be replaced by a code kept secure at a different location than the data set. Data will thereafter be extracted from the data base to perform statistical analysis. The study is approved by the Swedish Ethical Review Authority and all relevant approvals for data extraction and data storage has been obtained.