PO:14:204 | Risk stratification in systemic lupus erythematosus patients: improving the performances of a prediction model

Lucia Lanzo¹, Silvia Laura Bosello¹, Livia Lilli², Laura Antenucci², Carlotta Masciocchi², Marco Gorini³, Gabriella Castellino³, Luca Petricca¹, Maria Rita Gigante¹, Vivianaantonella Pacucci¹, Pier Giacomo Cerasuolo¹, Silvia Piunno¹, Francesco Cristiano¹, Cesare Gavotti¹, Pier Luigi Rizzuti¹, Rebecca Ventura¹, Jacopo Lenkowicz², Alfredo Cesario², Stefano Patarnello², Maria Antonietta D'Agostino¹. | ¹UOC di Reumatologia e Immunologia clinica - Fondazione Policlinico Universitario A. Gemelli IRCSS, UCSC, Roma, Italy; ²Fondazione Policlinico Universitario Agostino Gemelli IRCSS, Roma, Italy; ³AstraZeneca MIND, Milano, Italy.

Abstract

Background. Systemic lupus erythematosus (SLE) has an unpredictable course, with alternating periods of remission and activity. Predicting disease activity in the medium term could allow timely therapeutic adjustments. To support this, a Random Forest machine learning model was developed to predict the likelihood of an activity event within 12 months, using features from the patient’s current contact, previous visit, and overall disease history. Objectives: · To evaluate the performance of the main model; · To analyze characteristics of patients with low prediction reliability; · To demonstrate improved prediction through a cascade model.

Methods. The main model generates a predicted probability (PP) of an activity event within 12 months. A threshold of 0.45 was used to classify patients as high (PP > 0.45) or low risk (PP < 0.45). Cases close to the threshold were flagged as uncertain and analyzed further. Reliability boundaries were established based on distributions of true positives and negatives (Figure 1). Patients with uncertain predictions (“weak” cases) were compared to those with strong predictions using Mann-Whitney or Chi-squared tests (Table 1). To improve reliability for weak cases, a decision tree model (cascade model) was developed using a reduced feature set (e.g., demographic and treatment variables not used in the main model). This model provided rule-based predictions for cases where the main model’s predictions lacked confidence.

Results. The study included 262 SLE patients with 5,952 recorded contacts from 2012 to 2020. Seventy percent of the data was used for model training and 30% (1,845 contacts) for evaluation. In the test set, 590 contacts were classified as weak predictions, while 1,255 were confidently predicted. Weak cases were associated with older age, male gender, higher serosal, vascular, and cutaneous involvement, greater corticosteroid use, and lower antimalarial use. Use of Rituximab was linked to improved reliability, suggesting better disease control. The main model achieved an AUC of 74.2% after excluding weak predictions. The integration of the cascade model enabled more accurate predictions for weak cases, slightly increasing the overall AUC to 74.3%.

Conclusions. Combining the main model with a cascade decision tree extends the range of reliable predictions while maintaining accuracy. This integrated system enhances personalized care in SLE by supporting timely and individualized treatment decisions, potentially improving outcomes through better anticipation of disease activity. This work was supported by an unrestricted grant from AstraZeneca

Keywords

• Machine learning and AI
• LES
• Predictive models