Exploring immune-inflammation markers in psoriasis prediction using advanced machine learning algorithms

Abstract

Background: Psoriasis is a chronic immune-mediated inflammatory skin disorder characterized by multifactorial pathogenesis. Recent studies have extensively highlighted the strong associations between psoriasis and various inflammatory markers, which are considered novel predictive tools for evaluating systemic inflammation.

Methods: Cross-sectional data from the NHANES were analyzed in this study. To assess model performance and generalizability, the dataset was randomly divided into 70% for training and 30% for validation. To address class imbalance in the training data, a hybrid resampling technique (SMOTEENN) was applied. Subsequently, nine classification algorithms were developed using the processed training set, including random forest, neural networks, XGBoost, k-nearest neighbors, gradient boosting, logistic regression, naïve Bayes, AdaBoost, and SVMs. The final gradient boosting was implemented via the gbm package in R, with hyperparameters selected from the default tuning grid of the caret framework. Inflammatory biomarkers with the highest classification utility were identified based on the predictions of the best-performing model.

Results: A total of 22,908 participants were included in the final analysis. Gradient boosting (AUC: 0.629, 95% CI: 0.588-0.669) demonstrated the highest performance, followed closely by logistic regression (AUC: 0.627, 95% CI: 0.588-0.666). Among all the inflammatory markers, MLR exhibited the best classification performance, with an AUC value of 0.662 (95% CI: 0.640-0.683), followed by NLMR, with an AUC value of 0.661 (95% CI: 0.640-0.683). Other markers, including the NLR, dNLR, SII, SIRI, and PLR, had AUC values ranging from 0.658 to 0.661. The MLR had the highest relative importance score, demonstrating its critical role in the model's predictive performance for psoriasis classification. The NLR ranked second, followed by the SII and SIRI, which had moderate contributions, whereas the PLR contributed the least.

Conclusions: Among all the tested algorithms, the gradient boosting model achieved the best performance. Not only does it achieve the highest predictive accuracy, but it also excels in classification efficacy and feature importance analysis, highlighting key inflammatory markers such as the MLR, SII, and NLR. These markers are significant as reliable indicators for evaluating systemic inflammation and predicting the development of psoriasis, emphasizing their potential clinical applications.

Keywords: machine learning algorithms; monocyte-to-lymphocyte ratio; national health and nutrition examination survey; neutrophil-to-monocyte ratio; psoriasis.

Figure 1

Flowchart illustrating the participant selection process in NHANES cycles.

Figure 2

Relationships between the NLR, dNLR, SII, SIRI, MLR, PLR, NLMR, and AISI and psoriasis status. Each variable’s contribution and interaction within the data are analyzed to provide insights into their predictive capabilities and potential correlations. Symbols *, **, and *** indicate statistical significance levels: * p < 0.05, ** p < 0.01, *** p < 0.001.

Figure 3

Correlation coefficient plot illustrating the relationships among various inflammatory markers. The coefficients are represented by a color gradient, with darker hues indicating stronger correlations.

Figure 4

Receiver operating characteristic (ROC) curves for various models. The true positive rate (TPR) is plotted against the false positive rate (FPR) for each model. The curves highlight the area under the curve (AUC) values, providing a comparative measure of model performance.

Figure 5

(A) Weighted ROC curves for the metrics with covariates. Each curve represents the classification effectiveness of the corresponding metric for predicting the target variable, adjusted for covariates such as age, sex, ethnicity, marital status, the family income–to–poverty ratio (PIR), BMI, waist circumference, glucose levels, lipid levels, alcohol consumption, smoking status, and hypertension. (B) Feature importance of the variables in the gradient boosting machine model.