01 Data collection

Three stage-specific GEO cohorts were used so each model matched a distinct clinical question: diagnosis, next-visit activity, and treatment response (Barrett et al., 2013).

• Diagnosis: GSE72509 whole-blood RNA-seq; 99 SLE and 18 control samples (Hung et al., 2015).
• Progression: GSE65391 for training and GSE49454 for external validation; activity defined as SLEDAI ≥ 6 (Banchereau et al., 2016; Chiche et al., 2014).
• Treatment: GSE224705 first-visit SLE samples; response derived from SRI-4 labels after removing healthy controls (López-Domínguez et al., 2024).

02 Pre-processing and EDA

Expression and metadata were cleaned before modelling. Diagnosis RPKM values were transformed as log2(RPKM + 1); treatment microarray values were filtered for annotated, non-control probes and collapsed to one probe per gene (Bedre, 2023).

• Checked sample-level expression distributions and outcome balance.
• Derived longitudinal t-to-t+1 labels for progression.
• Standardised or imputed model inputs inside training workflows where required (NIST, n.d.).

03 Feature selection

Feature selection was performed before final modelling to reduce high-dimensional expression matrices to interpretable panels and to avoid leakage from external or held-out data.

• Diagnosis and treatment pre-filtered to highly variable expressed genes, then combined RF-Gini ranking, PCA checks, Boruta, and biological curation (Breiman, 2001; Kursa & Rudnicki, 2010; Liaw & Wiener, 2002).
• Progression selected expression probes from the training cohort only, removed direct leakage variables, then combined immune, clinical, treatment, temporal, engineered, and gene features.
• Final progression features were ranked with random-forest importance on training data only.

04 Modelling

Candidate models were compared within each stage rather than forcing one algorithm across all tasks. The deployed app reads the selected R model objects and sends the frontend CSV payload to the backend for inference; probability displays are summarised in Predicted probability bands.

Diagnosis compared limma signature, random forest, and LASSO. Progression compared elastic net, random forest, and GBM. Treatment response compared limma signature, random forest, LASSO, elastic net, and linear SVM. See Validation metrics below.

• Limma signature models use linear modelling with empirical-Bayes variance moderation for expression data (Smyth, 2004; Ritchie et al., 2015).
• Random forests aggregate many decision trees trained on bootstrap samples and random feature subsets; the R implementation follows this classification and regression workflow (Breiman, 2001; Liaw & Wiener, 2002).
• LASSO and elastic net are penalised regression models that shrink coefficients for feature selection, with elastic net combining L1 and L2 penalties (Tibshirani, 1996; Zou & Hastie, 2005; Friedman et al., 2010).
• GBM uses gradient boosting: sequential trees are added to improve a loss function, while linear SVM uses a maximum-margin classifier (Friedman, 2001; Cortes & Vapnik, 1995).

05 Performance evaluation

Evaluation prioritised metrics that are more informative than raw accuracy for imbalanced clinical cohorts. Diagnosis and treatment used stratified 5-fold cross-validation; progression used patient-level cross-validation for tuning and one independent external test on GSE49454 (Fawcett, 2006; Brodersen et al., 2010). See Validation metrics and Metric notes below.