THU0006 Application of machine learning methods for prediction modelling of psoriatic arthritis in patients with psoriasis. (12th June 2018)

• Record Type: Journal Article
• Authors: Jalali-najafabadi, F.; Dand, N.; Ho, P.; Smith, C. H.; Barker, J. N.; McHugh, N.; Warren, R. B.; Barton, A.; Bowes, J.
• Is Part Of: Annals of the rheumatic diseases. Volume 77(2018)Supplement 2
• Page Start: 232
• Page End: 233
• Publication Date: 2018-06-12
• Subjects: Rheumatism -- Periodicals; 616.723005
• DOI: 10.1136/annrheumdis-2018-eular.4430
• Languages: English
• ISSNs: 0003-4967

Abstract

Abstract : Background: Approximately 30% of patients with psoriasis develop a chronic inflammatory arthritis referred to as psoriatic arthritis (PsA). The ability to accurately predict which psoriasis patients will develop PsA would enable early intervention and help prevent disability. Both psoriasis and PsA have a substantial genetic risk component, however the utility of using genetic risk factors for the prediction of PsA is currently unknown. Alleles of the human leukocyte antigen (HLA) genes represent the largest genetic effects observed for both psoriasis and PsA (HLA-C*0602 and HLA-B*27 respectively); these genes are highly polymorphic with extensive linkage disequilibrium (LD) which will make variable (feature) selection using statistical models very challenging. Machine learning methods, such as information theoretic criteria, are well suited to this challenge and will find a subset of the original variables that enable more accurate prediction. Objectives: To apply machine learning methods for feature selection of HLA alleles and evaluate the accuracy of these feature for the prediction of PsA. Methods: Feature selection was performed using information theoretic criteria methods which are classifier independent methods that provide a ranking of genetic features that differentiate PsA from cutaneous-only psoriasis. Multiple methods were tested; mutual information maximisation (MIM), joint mutual information (JMI), minimal-Redundancy-Maximal-Relevance (mRMR) and conditional mutual information maximisation (CMIM). Two principal components (population stratification) and age of psoriasis onset were included as potential confounders. The Bagged Trees method was used for classification and the performance of the predictive models were assessed using area under the receiver operating characteristic curve. These methods were applied to a dataset of 1462 PsA cases and 1132 cutaneous-only psoriasis cases using 2-digit and 4-digit classical HLA alleles imputed using the SNP2HLA algorithm. Results: The single most important features based on rank were identified as HLA-B*27 (2-digit) and HLA-B*2705 (4-digit) by the four different feature selection techniques; this is consistent with previous analyses of this data using regression based methods. However, the predictive accuracy of these single features was found to be poor (AUC 0.55 HLA-B*27). Sequentially adding additional HLA features based on rank substantially improved the performance of the classification model where 20 2-digit features selected by JMI demonstrated an average AUC of 0.84 based on 10 cross-fold validation (figure 1). Conclusions: The results demonstrate that classification models constructed from multiple HLA alleles substantially outperform classification based solely on the previously reported PsA risk allele (HLA-B*27). Importantly, the study demonstrates that this additional information is efficiently captured using information theoretic criteria methods which capture correlations between markers. Disclosure of Interest: None declared