MACHINE LEARNING CAN PREDICT GIANT CELL ARTERITIS RELAPSE AFTER GLUCOCORTICOID TAPERING
Abstract: AB0380
Authors: V. Venerito et al.
To date reliable biomarkers and risk factors for relapsing giant cell arteritis (GCA) after glucocorticoid (GC) tapering are still lacking. Machine learning (ML) is emerging as a promising tool for the implementation of complex multi-parametric decision algorithms. A ML approach allows to handle complex non-linear relationships between patient attributes that are hard to model with traditional statistical methods, merging them to output a forecast or a probability for a given outcome.
Key content:
GC-naïve GCA patients who presented to 4 tertiary care centres between January 2015 and January 2019, who underwent GC therapy and regular follow up visits for at least 12 months were retrospectively analyzed and used for training and validation (through 10-fold cross-validation) of n.2 ML algorithms, namely Decision Trees (DT) and Random Forest (RF). The training and validation dataset consisted of 85 GCA patients (59 female, 69.4%) with mean age 73.8 (±8.7) years at presentation. They were treated with 27.1 (±17.4) mg prednisone (PDN) equivalent at first visit. During GC tapering 34 of them (40%) experienced a disease relapse within 12 months. The test dataset consisted of 22 patients (14 female, 63.4%) with mean age 75.5 (±8.7) years at presentation, who underwent GC induction therapy with a mean dose of 30.3 (±17.3) mg PDN equivalent. Nine of them (40.9%) had a GCA flare during GC tapering, within 12 months. Accuracy of DT and RF in predicting the outcome of interest on the training dataset was 68.3% and 73.4% respectively. On testing datasets DT and RF accuracy was 57.1 and 72.4%, respectively.
Relevance:
The authors state: «RF algorithm can predict GCA relapse after GC tapering with fairly good accuracy». However, the most important patient attributes for RF forecast were found to be CRP and ESR baseline levels as well as age and symptom duration (months) at first visit. Thus, at the end of the day, the variables identified by ML appear to be the ones taken into account by a practising rheumatologist?!
