CO:05:1 | Predicting one-year disease activity in systemic lupus erythematosus using machine learning and explainable AI

Pier Giacomo Cerasuolo¹, Augusta Ortolan¹, Livia Lilli², Laura Antenucci², Carlotta Maschiocchi², Jacopo Lenkowicz², Lucia Lanzo¹, Luca Petricca¹, Maria Rita Gigante¹, Vivianaantonella Pacucci¹, Pierluigi Rizzuti¹, Rebecca Ventura¹, Silvia Piunno¹, Cesare Gavotti¹, Marco Gorini³, Gabriella Castellino³, Alfredo Cesario², Stefano Patarnello², Silvia Laura Bosello¹, Maria Antonietta D'Agostino¹. | ¹UOC di Reumatologia ed Immunologia Clinica, Fondazione Policlinico A. Gemelli IRCCS, Università Cattolica del Sacro Cuore, Roma; ²Gemelli Generator, Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Roma; ³AstraZeneca Italy, MIND, Milano, Italy

Abstract

Background: Systemic lupus erythematosus (SLE) is a chronic, multisystem autoimmune disease characterized by marked heterogeneity in clinical presentation, disease course, and response to therapy. This variability makes it particularly challenging to predict disease flare-ups in clinical practice, despite their profound implications on prognosis and quality of life.

Objectives: This study aimed to develop a machine learning model to predict the risk of disease activity events in SLE patients within a 12-month time frame, based on clinical status, recent history, and long-term disease trajectory. Methods: We built a structured dataset (SLE Data Mart) from electronic health records using data mining and NLP techniques. It includes demographic, therapeutic, clinical, and laboratory information from SLE patients seen at our University Hospital. Patients had at least one outpatient visit and one hospitalization coded as ICD9 710.0 or 695.4. The outcome—an activity event within 12 months—was defined as any of the following: neurological, renal, or vascular symptoms; new organ domain involvement; or hospitalization for SLE. Input features described current contact, past 12 months, and prior clinical history, expressed as binary or continuous variables. Univariable selection (Chi-Square, Mann-Whitney U tests) identified features significantly associated with the outcome (p < 0.05). Selected variables were used to train multiple models (Logistic Regression, SVC, Decision Tree, Random Forest, XGBoost), evaluated by 5-fold cross-validation. The model with the highest AUC was selected and interpreted using SHAP to identify key predictors.

Results: The dataset included 262 SLE patients and 5,952 contacts between 2012–2020. Most patients were female (88%), with a median age of 43 years and median follow-up of 6 years. Patients had a median of 16 contacts each. The cohort was split into training (70%) and testing (30%) sets. Overall, 48% of contacts were associated with an activity event. Univariable selection retained 125 of 192 variables (Table 1). Significant predictors included younger age at contact (median 44 years, p < 0.0001) and abnormal labs (e.g., proteinuria, albuminuria). The best-performing model (AUC = 0.70) identified renal involvement—especially proteinuria >=0.5 g/24h and albuminuria—as the most impactful predictors. SHAP analysis also highlighted the protective role of therapy step-down and of the current hematological or cutaneous involvement (Figure 1).

Conclusion: Renal abnormalities strongly predict future SLE activity, but other factors, including hematological and cutaneous involvement, prior therapy changes, and overall disease history, significantly contribute to to determine the risk for an activity event at 12 months.

Acknowledgements. This study was conducted with the unrestricted support of AstraZeneca.

Keywords

• Lupus
• Machine learning
• Artificial intelligence