Artificial Intelligence and Psoriatic Disease

Psoriatic disease is a systemic, immune-mediated condition affecting skin and/or joints. It is associated with multiple comorbidities such as cardiovascular disease (CVD), diabetes, hypertension, metabolic syndrome, anxiety, and depression.  Despite access to numerous advanced treatments, including biologics and small molecules, which have transformed psoriasis management, many unmet needs remain.  AI, which refers to a machine’s ability to communicate, reason, and operate independently in both familiar and novel scenarios in a similar manner to a human, may offer a solution to diagnostics and management of psoriatic disease.¹

It should be noted that although the terms AI and machine learning (ML) are often used interchangeably, ML is a subset of AI related to teaching machines to learn tasks from data through pattern recognition and inference.  Indeed, algorithms can also be learned via deep learning (DL), which can be performed without labelled data sets, via a neural network with multiple layers of “neurons” that have adjustable weights (or mathematical functions), with ML to train or test data across its network for improved accuracy and performance.²

In this paper, the authors conducted a literature review on the use of AI in psoriatic disease, finding 38 relevant papers, including a range of observational, interventional, and descriptive studies across many populations and datasets.³

AI can support the diagnosis of psoriasis. Indeed, the Google AI tool, which permits users to upload three well-lit images of the skin, hair, or nail of concern and which is underpinned by a DL system formed from a data set containing over 16,000 pictures of skin disorders, was shown to be noninferior to 6 board-certified US dermatologists and superior to 6 primary care doctors.⁴ Investigating whether psoriasis can also be differentiated from other inflammatory disorders, Zhao et al classified 8,021 images of 9 common skin conditions, including psoriasis, using convolutional neural networks from a cohort of patients from a Chinese hospital.⁵ Their algorithm was superior to 25 Chinese dermatologists when tested on 100 new images. They reported a misdiagnosis rate of 3% compared to 27% by dermatologists.⁵

Clinical assessment, such as Psoriasis Area and Severity Index (PASI), is widely used by dermatologists but presents an additional challenge for AI, such as textural changes, thickness, and the proportion of affected body surface area features in the scoring. Huang et al used a database of 14,096 images from a cohort of 2,367 Chinese patients with psoriasis to estimate PASI by processing the images with a convolutional neural network to extract specific features.⁶ The DL method thus developed was comparable to PASI calculated by 43 dermatologists and has been successfully used in 18 different sites via the use of an app, demonstrating a significant strength of AI in reaching greater numbers of patients – particularly those in remote and resource-limited settings.⁶

Predicting the right treatment for the right patient remains challenging, especially as there remains a lack of robust data on the utility of biomarkers in predicting response to therapies. However, using AI to combine genotypic and phenotypic characteristics of patients with psoriasis to identify the most appropriate treatment is starting to emerge. Promisingly, Emam et al, analyzed data from 681 patients with psoriasis from the Danish national registry using six different ML techniques to identify patterns from demographic and clinical data: generalized linear model, support vector machine, decision tree, random forest, gradient-boosted trees, and DL.⁷ Treatment outcomes were predicted with high accuracy and less than 18% classification error, with data that are routinely available to clinicians.⁷ Subsequently, Nielsen et al used the same registry to retrospectively predict the most suitable biological therapy for patients with psoriasis. They found that gradient-boosted decision trees, a specific type of ML, performed significantly better than logistic regression for predicting specific biologic therapy.⁸ This technique could predict discontinuation of a given biologic within the first year of treatment with an accuracy of 62.9% to 67.6%. Importantly for dermatologists, healthcare data is well-suited to interpretation via ML, inferring patterns that may deepen our understanding of treatment trajectories and pathophysiology, taking psoriasis care one step closer to personalized medicine.⁸

Indeed, genome-wide association studies, which provide enormous amounts of complex data, can be swiftly interpreted via ML techniques to predict comorbidities. Patrick et al used ML (including random forest, conditional inference forest, shrinkage discriminant analysis, and elastic net regression) to identify nine novel loci for psoriasis following evaluation of more than 7,000 genotyped patients with psoriasis provided by genome-wide association studies.⁹ Love et al also analyzed data on 2,318 patients with psoriatic arthritis to identify 31 psoriatic arthritis-related predictors.¹⁰

In summary, there is significant potential in using AI for the diagnosis and management of psoriatic disease. AI could contribute to timely diagnostics, especially in areas with limited access to a dermatologist. Moreover, it could potentially predict: i) the right medication for the right patient, ii) the development of adverse events to a specific drug, and iii) the development/progression of comorbidities, which could bring us one step closer to personalized medicine. However, there are challenges in applying AI in everyday practice, including accountability, generalizability, and patient acceptance. Furthermore, assessment of disease severity or application in people with diverse skin tones remains challenging, and it is important that further steps are taken to improve its accuracy.¹¹ Nevertheless, although we will likely see an increase in the application of AI, which will translate to greater benefits to our patients, there will also be continued human-human, face-to-face interaction.

A Brief Note from the Author (Dr. Maria-Angeliki Gkini):

• AI can enhance the diagnosis of psoriasis by accurately analyzing skin images, aiding in the early detection of the disease and potential early intervention.
• AI models could contribute to predicting the appropriate treatment for the appropriate patient and tracking treatment outcomes.
• AI-powered apps can enable patients to self-monitor their disease, promoting proactive management.
• AI is here, it is present, and it is not just looking to the future. There are still challenges and limitations in its use in dermatology, and there is a need for improvisation and monitoring under an appropriate legal framework.
• Physicians must familiarize themselves with using AI, understand its limitations, embrace it, and use it responsibly to support our decision-making for optimal management of our patients.