The Problem
Access to dermatology care is limited and with increasing gap between supply and demand1. According to a survey study of dermatologists, the mean ± standard deviation (SD) waiting time was 33 ± 32 days. Sixty-four percent of the appointments exceeded the criterion cutoff of 3 weeks. The waiting time ± SD for an established patient appointment was 32 ± 30 days. Sixty-three percent of the appointments exceeded the 2-week criterion cutoff for established patients. At Duke, the current lead time is 34.3 days [data obtained from PDC], and most recent calculated wait time, defined as referral to appointment, was close to 70 days. The concern is even more significant for our growing Medicare population, expected to account for 1 in 5 patients by 2030.2 In our current dermatology practice at Duke, Medicare population accounts for 50% of the charges.
In the realm of care, skin cancer accounts for a significant disease burden. North Carolina has the 5th highest incidence of melanoma and access to care can affect outcomes in cases of skin cancer especially melanoma.3,4The access times described above are far from ideal for skin cancer patients, causing significant clinical and emotional morbidity.
Given the expected rise in baby boomers, with significantly increased risk of skin cancer as North Carolinians, there is an urgent need to equip primary care providers to help screen and risk stratify patients in real time, high quality and cost-conscious fashion. Many primary care providers have difficulty effectively evaluating dermatological concerns because of variable experience and training in dermatology. A consistent decision support system can mitigate this variability, and create a powerful risk stratification tool, leveraging our frontline network of providers to enhance access to quality and valuable care.
Recently, several studies have demonstrated the promise of deep learning as convolutional neural networks (CNN) in the field of dermatology in terms of classification of skin lesions.5,6 This suggests that when properly developed, studied and operationalized, deep learning has the promise to be a helpful clinical diagnostic aid in dermatology. With ability to better classify skin lesions, an effective risk stratification strategy can be developed to streamline dermatology access by ensuring high risk patients are seen quickly and offering decision support to manage low risk conditions without referral. This approach can address issues of dermatology access times and quality of care at the highest standards with careful oversight of and access to expert dermatologists.
However, while the promise exists in theory, there is lack of prospective data demonstrating the clinical utility, cost effectiveness, accuracy and consistency of CNNs in clinical practice. With the continued evidence of deep learning in medicine, our multidisciplinary team at Duke is a significant advantage to advance clinical data to enhance real time practice of medicine.
Our Solution
Images from Duke dermatology archives (27,000) and open access International Skin Imaging Collaboration (ISIC; 23,000) with a well characterized ground truth (histopathological diagnoses) have been utilized to train a classification CNN model that can evaluate clinical and dermoscopy images for the 7 most common benign and malignant skin lesions. Additionally, the model is able to detect blurry images, poor light quality, and able to auto segment lesion of concern from photographs taken from an iPhone camera, further democratizing the process.
The current state of the model is dependent on retrospective clinical image data and it is not currently a part of clinical workflow. With assistance from the DIHI grant, we will deploy this model into clinical practice to be used in real time. Primary care providers will be able to take images at point of care, with the goal of stratifying the acuity of the condition, and providing clinical decision support to the PCP. Incorporating this technology as an EHR-based clinical decision support tool, there’s an opportunity for system-wide intervention to present real-time information (such as image quality and proposed diagnosis) in the routine care workflow to improve diagnostic decisions. To enable rapid identification of relevant dermatology diagnosis, our proposed pilot will develop and validate the algorithms to automatically classify lesions and improved risk stratification of patients which will lead to better use of dermatology resources and lower wait times/more expedited care for patients.
