USING MATHEMATICAL TECHNIQUES IN MACHINE LEARNING TO PREDICT TREATMENT OUTCOMES IN PATIENTS WITH SEVERE BULLOUS DERMATOSES.

Introduction

Artificial intelligence and, in particular, machine learning (ML) methods have recently found broader applicability in many medical fields, including dermatology. They can be used to predict patient`s response to therapy, severity and activity of various diseases, including dermatologic conditions. ML particularly has been used for stratification of dermatoscopic features in dermatooncology (to classify malignant and benign skin neoplasms), as well as in patients with immune-related dermatoses (psoriasis, atopic dermatitis, acne, etc.). It is important to note that direct immunofluorescence reaction image analysis systems were most often used for patients with severe bullous dermatoses (BD) and showed high sensitivity in recognizing diagnostic data. However, it should be noted that simultaneous analysis of imaging, genetic, and immunologic data using ML techniques has not been performed.

The goal of the study

The study aimed to design a prognostic model to predict resistance to systemic glucocorticosteroids (CSs) in patients with BD, and to classify them into steroid-resistant and steroid-sensitive subgroups based on the obtained laboratory data on genomic (HLA-DRB1 and HLA-DQB1, A3669G polymorphism of the β-isoform of the glucocorticoid receptor (GR), expression of α- and β-isoforms of GR) and non-genomic (TNF- α, IL4, IL15, IL10, CXCL8, CLL11, granulysin) mechanisms, using ML methods to improve the accuracy of early diagnosis of these diseases and to ensure timely appointment of personalized therapy with immunosuppressive, cytostatic and biological drugs.

Materials and Methods

The study included 150 patients based on the evaluation of HLA-DRB1 and HLA-DQB1 alleles, A3669G polymorphism of the β-isoform of GR, and expression of α- and β-isoforms of GR by PCR. Of these, 107 patients were pemphigus (AP), 86 (57.3%) - pemphigus vulgaris (PV), 13 (8.67%) - pemphigus foliaceus (PF), 6 (4%) - paraneoplastic pemphigus (PNP), and 2 (1.4%) - pemphigus vegetans (Pveg); 29 (19.3%) with bullous pemphigoid (BP); 14 (9.33%) from the severe bullous allergic reaction group - Stevens-Johnson syndrome (SJS)/toxic epidermal necrolysis (TEN); and 92 healthy controls. 67 BD patients were included in the group in which cytokine, chemokine profiles, and granulysin levels in serum and blister fluid were detected and evaluated by ELISA. Of these, 43 patients with pemphigus, 11 - with BP, 13 - with SJS/TEN and 43 healthy controls. The following ML methods were used: logistic regression, support vector machine, “decision tree”, “random forest”, “gradient boosting”, operational analysis (AUC, ROC) and metrics “recall” and “precision”.

Results

The logistic regression method showed the highest metrics: “Recall” (1.000), “Precision” (0.938) and “ROC-AUC” (0.992), indicating that this model was quite accurate in distinguishing between classes and a very high probability of the algorithm correctly determining whether the person was patient with BD or healthy. We derived the formula P(Y=1)=1/(1+e^(-(-4.0479+0.1102∙Alfa) ) ), which indicates that when the expression value of α-isoform of GR exceeded 36.7 units, the probability of developing BD was more than 50% (95% CI: 1.084-1.150; significance: 1.116). The support vector machine highlighted the most significant features predictive for SR: DRB1 alleles in high resolution, such as DRB1*13:02; DQB1*05:02, DQB1*02:01, DQB1*05:01 and DQB1*03:01, and the presence of A3669G polymorphism of the β-isoform of GR. Conversely, alleles that increased the likelihood of lack of SR in BD patients included DQB1*06:02, DQB1*01:01, DQB*04:03, and DQB1*06:03. “CatBoost” and “random forest” methods most accurately predicted the likelihood of SR in BD patients when the following most significant features were present: high levels of IL15, IL4 and CXCL8 in blister fluid, high serum granulysin levels, diagnosis and severity of AP, administration of azathioprine (F1-metric: 0.538721, ROC AUC: 0.879167, Recall: 0.541667, Precision: 0.857143) and presence of type 2 diabetes mellitus.

Conclusion

The obtained formula enables determination tendency of the feature impact (high expression of α-isoform of GR) on the probability of BD using the odds ratio and to risk-stratify patients with a very high degree of accuracy. The random forest algorithm and the support vector machine enable us to classify BD patients into steroid-resistant and steroid-sensitive subgroups based on the level of HLA class II alleles and the presence of the A3669G polymorphism of the β-isoform of GR. Blister fluid can be used for diagnosis, differential diagnosis, prediction of the severity of BD, and evaluation of response to CSs therapy. In addition, blister fluid analysis will be able to predict patient`s response to conventional CSs therapy at early stages of the disease, allowing timely correction of treatment and thereby reducing the risk of potential complications and progression of BD.