Predicting EQ-5D full health state in systemic lupus erythematosus using machine learning algorithms

João Botto; Nursen Cetrez; Dionysis Nikolopoulos; Malin Regardt; Emelie Heintz; Julius Lindblom; Ioannis Parodis

doi:10.1093/rap/rkaf032

Predicting EQ-5D full health state in systemic lupus erythematosus using machine learning algorithms

Rheumatol Adv Pract. 2025 Apr 18;9(2):rkaf032. doi: 10.1093/rap/rkaf032. eCollection 2025.

Authors

João Botto^{1

2}, Nursen Cetrez^{1

2}, Dionysis Nikolopoulos^{1

2}, Malin Regardt^{3

4}, Emelie Heintz⁵, Julius Lindblom^{1

2}, Ioannis Parodis^{1

2

6}

Affiliations

¹ Division of Rheumatology, Department of Medicine Solna, Karolinska Institutet, Stockholm, Sweden.
² Department of Gastroenterology, Dermatology and Rheumatology, Karolinska University Hospital, Stockholm, Sweden.
³ Department of Neurobiology, Care Sciences and Society, Division of Occupational Therapy, Karolinska Institutet, Stockholm, Sweden.
⁴ Women's Health and Allied Health Professionals Theme, Medical Unit Occupational Therapy and Physiotherapy, Karolinska University Hospital, Stockholm, Sweden.
⁵ Department of Learning, Informatics, Management and Ethics (LIME), Karolinska Institutet, Stockholm, Sweden.
⁶ Department of Rheumatology, Faculty of Medicine and Health, Örebro University, Örebro, Sweden.

Abstract

Objectives: To determine factors associated with reports of EuroQol 5-Dimensions (EQ-5D) full health state (FHS) before and after a trial intervention in patients with SLE, resorting to machine learning algorithms.

Methods: We conducted a post hoc analysis of two phase 3 clinical trials of belimumab (BLISS-52, BLISS-76). Demographic, laboratory and clinical features were retrieved and the Monte Carlo Feature Selection algorithm was employed, then further refined upon consideration of collinearity and clinical relevance. We used support vector machine with radial basis function kernel (SVMRadial), least absolute shrinkage and selection operator (LASSO), neural network (NNet) and logistic regression (LR) to capture both linear and non-linear relationships while ensuring interpretability and robustness.

Results: Among 1642 SLE patients, 12.9% reported FHS at baseline and 23.1% at week 52. Selected features were age, sex, Asian ancestry, baseline cSLEDAI-2K, SELENA-SLEDAI PGA, and urine protein:creatinine ratio (UPCR) and baseline EQ-5D 3-Levels (EQ-5D-3L) index score (week 52 models only). The models predicting FHS demonstrated comparable performance at baseline and week 52. A maximum area under the curve of 0.73 was seen for the baseline LASSO and LR models and a maximum of 0.77 for the week 52 LASSO and NNet models. Negative predictive values were high for all models (0.88-0.94). Calibration showed marginal improvement in week 52 models.

Conclusion: Machine learning identified older age, female sex, non-Asian ancestry, high disease activity and low UPCR to be associated with a lack of FHS experience in SLE patients at baseline and week 52. High baseline EQ-5D-3L index scores constituted the strongest predictor of FHS at week 52.

Trial registration: The BLISS-52 and BLISS-76 trials are registered at www.clinicaltrials.gov (NCT00424476 and NCT00410384, respectively).

Keywords: EQ-5D; machine learning; patient-reported outcomes; quality of life; systemic lupus erythematosus.

Associated data

ClinicalTrials.gov/NCT00410384
ClinicalTrials.gov/NCT00424476