A Taylor-Series Engineering Approach to Machine Learning-Based Early Detection of Systemic Lupus Erythematosus
Harrison Baek
21/07/2026
This study addresses the challenge of early detection of systemic lupus erythematosus (SLE), an autoimmune disease that is often diagnosed late due to its heterogeneous clinical presentation. I investigate whether Taylor-series–derived features can enhance machine learning (ML) models by capturing nonlinear relationships in gene expression data, thereby improving classification performance. I hypothesize that Taylor-series–derived features, which represent local curvature and gene–gene interactions, will significantly improve model accuracy compared to using original gene expression features alone. To evaluate this, I analyze gene expression data from 26 SLE patients and 46 controls using Logistic Regression, Random Forest, and Support Vector Machine models. I expand selected informative genes into second-degree polynomial features and apply Lasso regularization for feature selection before retraining the models. The Taylor-enhanced models increase average classification accuracy from 93.9% to 98.5%, with the largest improvement observed in the SVM model. Approximately 40% of the selected features involve gene–gene interactions associated with biologically relevant pathways, including interferon signaling and B-cell activation. However, these improvements are not statistically significant according to McNemar’s test, likely due to the small sample size. Overall, the findings suggest that nonlinear Taylor-series–derived features can capture meaningful biological relationships and improve predictive performance, warranting further validation in larger datasets before any clinical application. Although performance improvements were observed, the lack of statistical significance indicates that these differences may be attributable to limited statistical power and should be interpreted cautiously.