187 Recurrent neural networks to predict biologic treatment outcomes in psoriasis

Evidence-based personalized treatment strategies do not exist to guide biologic of psoriasis. As a result, trial-and-error approach is adopted. Recurrent neural networks(RNNs) are deep learning artificial intelligence models that make predictions on longitudinal data, such as within the British Association Dermatologists Biologic and Immunomodulators Register(BADBIR). We aimed develop RNNs predict 1) drug survival; 2) achievement absolute psoriasis severity index (PASI)≤2, equivalent 90% reduction in baseline(PASI 90) at 6 12 months initiating using real-world clinical data. BADBIR data were engineered enable RNN readability. Model 1 included all cohort patients with follow-up 2 naïve prescribed adalimumab, etanercept or ustekinumab baseline 12-month PASI. Binary classification random forest(RF) developed for comparison. Baseline (excluding PASI model inputted. (n=10,932) outperformed RF survival parameters except recall (precision 0.90 vs 0.79, 0.65 0.75, area under receiver operating characteristic (AUROC) 0.83 AUC 0.50). Excluding improved AUROC (0.86). K-fold cross-validation was performed biologic-naïve patient (n=10,642): 0.81 (SD 0.01). predicted PASI≤2 RF. (6 months, n=3,849): precision 0.59 0.56, 0.93 0.55, 0.71 0.66. Precision (0.65 0.44 (RF), n=6,506). This first study employing outcomes large dataset. RFs previously published predictive models. Further evaluation including assessment early time windows external validation required understand relevant factors driving their application.