A machine learning-based, decision support, mobile phone application for diagnosis of common dermatological diseases.

BACKGROUND

The integration of machine learning algorithms in decision support tools for physicians is gaining popularity. These tools can tackle the disparities in healthcare access as the technology can be implemented on smartphones. We present the first, large-scale study on patients with skin of colour, in which the feasibility of a novel mobile health application (mHealth app) was investigated in actual clinical workflows.

OBJECTIVE

To develop a mHealth app to diagnose 40 common skin diseases and test it in clinical settings.

METHODS

A convolutional neural network-based algorithm was trained with clinical images of 40 skin diseases. A smartphone app was generated and validated on 5014 patients, attending rural and urban outpatient dermatology departments in India. The results of this mHealth app were compared against the dermatologists' diagnoses.

RESULTS

The machine-learning model, in an in silico validation study, demonstrated an overall top-1 accuracy of 76.93 ± 0.88% and mean area-under-curve of 0.95 ± 0.02 on a set of clinical images. In the clinical study, on patients with skin of colour, the app achieved an overall top-1 accuracy of 75.07% (95% CI = 73.75-76.36), top-3 accuracy of 89.62% (95% CI = 88.67-90.52) and mean area-under-curve of 0.90 ± 0.07.

CONCLUSION

This study underscores the utility of artificial intelligence-driven smartphone applications as a point-of-care, clinical decision support tool for dermatological diagnosis for a wide spectrum of skin diseases in patients of the skin of colour.