Embracing machine learning and digital health technology for precision dermatology

The promise of precision medicine in dermatology to deliver the “right treatment at the right time” is within reach through the current revolution in digital health technology and machine learning (ML). Although ML in medicine has led to concern that computers will replace clinicians, we offer our viewpoint on how ML in dermatology can provide valuable tools for delivering care and improving patient outcomes.

ML combined with connected health devices offers the potential to further detection capabilities and services for automated preliminary screening. If an individual is concerned about skin cancer for a lesion, photos could be uploaded to a web or smartphone application to obtain an initial evaluation through a deep learning algorithm, such as a convolutional neural network, which is composed of multiple layers of “neurons” that extract features to use in classifying the image as a malignant or benign lesion [1-2].Although the application is still in an experimental phase, with additional research, it is possible that patients could access such services at home on their own devices. Additionally, triaging to prioritize the order of patient scheduling based upon level of concern through an initial ML enabled screening can help provide services to patients in most urgent need of care.

Digital health technology can also enable the patient to play a more active role in providing accurate, quantitative data to inform clinical decision-making. As shown in Figure 1, since the majority of patients’ lives are outside of the healthcare setting, clinicians often depend upon information that represents only snapshots in time. With digital technology, precise and accurate records of symptoms and potential triggers with contextual information could be captured. Combining this longitudinal data with ML, causal inference, and data visualization tools, can allow for the understanding of relationships between behaviors/exposures and symptoms. Longitudinal data offers the benefit that rather than relying on a single timepoint, repeated data points are collected. With data visualization tools, both clinicians and patients can view trends in behaviors/exposures and outcomes of interest. With causal inference, it may be possible to identify causal connections between behaviors/exposures and outcomes. For example, for eczema treatment/management, patients could capture potential triggers, medication usage, and photos. ML for image analysis combined with longitudinal data and causal inference could provide insights into how certain behaviors/exposures (e.g. stress, dietary habits) impact disease severity or resolution.

Figure 1: Example of General Machine Learning (ML) Workflow for Delivering Precision Dermatology.

The general ML workflow begins with data collection. These data can consist of both data from the healthcare setting, which represent only a small fraction of the patient’s life, and data from continuous, remote monitoring through digital health technology. Afterwards, data cleaning (e.g. data normalization, scaling, handling of missing values) is performed to prepare the training data, which is used by the machine learning algorithm during model development. Next, model evaluation involves calculating performance measures such as accuracy, sensitivity, specificity, and the area under the curve (AUC). Once the model is deployed/implemented, it can be used to obtain a prediction on a new observation. This prediction can then be used to inform clinical decision-making (e.g. treatment optimization, flare predictions, cancer screening, differential diagnosis).

Although the hype surrounding ML has resulted in some initial concern that “robots [are] coming for dermatologists’ jobs,” [3] rather than replacing dermatologists, ML technology will be essential in realizing the promise of precision dermatology. Nevertheless, many questions remain surrounding how to safely and effectively integrate these tools into clinical practice and patient care. Rather than fearing the emergence of ML, it is essential that dermatologists guide the development of this technology to ensure that these ML tools are clinically useful and augment our abilities to improve patient outcomes.