Automated machine learning predicts necessary upcoming therapy changes in patients with psoriasis vulgaris et arthritis and uncovers new influences on disease progression: retrospective study (Preprint)

Abstract

Psoriasis vulgaris (PsV) and Psoriatic arthritis (PsA) are intertwined multifactorial diseases with significant impact on health and quality of life, which can be debilitating due to chronicity and treatment complexity. Predicting treatment response and disease progression in these conditions is challenging, but crucial for optimising therapeutic interventions. The emerging technology of automated machine learning (AutoML) offers a promising approach to rapidly build highly accurate predictive models based on patient characteristics and treatment data.

The study aimed to develop highly accurate ML models using AutoML to address key clinical questions in PsV and PsA patients, including predicting therapy changes and identifying reasons for therapy changes, factors influencing skin lesion progression or factors associated with an abnormal BASDAI score.

After extensive dataset preparation of clinical study data from PsV and PsA patients, a secondary dataset was created and ultimately analysed using AutoML to build a variety of predictive models and select the most accurate one for each variable of interest.

"Therapy change at 24 weeks follow-up" was modelled using the eXtreme Gradient Boosted Trees Classifier with Early Stopping model (AUC of 0.9078 and LogLoss of 0.3955 for the holdout partition) to gain insight into the factors influencing therapy change, such as the initial systemic therapeutic agent, the score achieved in the CASPAR classification criteria at baseline, and changes in quality of life. An AVG blender of 3 models (Gradient Boosted Trees Classifier, ExtraTrees Classifier, Eureqa Generalised Additive Model Classifier) with an AUC of 0.8750 and a LogLoss of 0.4603 was used to predict therapy changes on two hypothetical patients to highlight the importance of such influencing factors. Notably, treatments such as MTX or specific biologicals showed a lower propensity for change. A further AVG Blender of RandomForest Classifier, eXtreme Gradient Boosted Trees Classifier and Eureqa Classifier (AUC of 0.9241 and LogLoss of 0.4498) was then used to estimate "PASI change after 24 weeks" with the primary predictors being the initial PASI score, change in pruritus and change in therapy. A lower initial PASI score, and consistently low pruritus were associated with better outcomes. Finally, "BASDAI classification at baseline" was analysed using an AVG Blender of Eureqa Generalised Additive Model Classifier, eXtreme Gradient Boosted Trees Classifier with Early Stopping and Dropout Additive Regression Trees Classifier with an AUC of 0.8274 and LogLoss of 0.5037. Factors influencing BASDAI scores included initial pain, disease activity and HADS scores for depression and anxiety. Increased pain, disease activity and psychological distress were generally likely to lead to higher BASDAI scores.

The practical implications of these models for clinical decision making in PsV and PsA have the potential to guide early investigation and treatment, contributing to improved patient outcomes.