Predicting Work Disability in Systemic Lupus Erythematosus Patients: A Machine Learning Approach to Guide Early Clinical Intervention

Objectives Systemic lupus erythematosus (SLE) is a chronic autoimmune disease characterized by unpredictable flares and multiorgan involvement, often leading to progressive functional decline. These debilitating symptoms frequently impair work capacity, resulting in premature workforce exit affecting 20-40% of SLE patients. The loss of employment carries severe socioeconomic and psychological consequences, including financial instability, reduced quality of life, and increased healthcare dependency. Predicting work disability (WD) for SLE patients is critical to preserving independence and quality of life. This study aims to develop: 1-A prediction model for long-term adverse outcomes like WD and SLE-mortality; and 2-A machine learning algorithm to pre-emptively flag patients at high-risk of WD. Methods We analyzed longitudinal data from 1,044 employed SLE patients at the Toronto Lupus Clinic (1996–2024). Patients reported their employment category at every clinical visit and were categorized into 3 exclusive outcomes: WD (permanent or prolonged sick leave), SLE-mortality, and stable employment (actively employed or retiring at/after Canadian retirement age). A Random Forest classifier selected visit-to-visit disease activity measures, treatment regimens, and organ damage features to be used in a Long Short-Term Memory (LSTM) model. Each visit’s data was processed with prior history to create an updated patient health hidden state. The final hidden state was used to classify patients into their most likely final employment outcome. Risk scores for WD were calculated at each visit and early warning alerts were triggered if scores exceeded an optimized threshold. Prediction accuracy and lead time for early detection assessed model effectiveness. Results 129 patients eventually reported WD, these patients showed more frequent visits to the clinic over longer follow-ups, with higher doses of glucocorticoids and antimalarials, higher disease activity, irreversible damage, and more frequent flares (all p < 0.001) (Table 1). The prediction model achieved 91% balanced accuracy in predicting the final employment states. More significantly, the early warning machine learning algorithm identified 89% of future disability cases, providing a median 28.6-month (IQR: 14.3–42.1) warning before first reporting WD. Table 1. Characteristics of the Patients by Employment Group at last visit* and Model Results Conclusion Predicting workforce exit for SLE patients is critical and our early warning algorithm flagged patients at a risk of disability 2–3 years in advance. Implementing this system would enable timely interventions to help prevent work disability in SLE. By transforming reactive care into pre-emptive action, this prediction model closes a critical gap in SLE management, empowering health care teams to intervene before functional decline becomes permanent WD.