PO:31:168 | Risk factors for cancer in systemic sclerosis, impact on disease phenotype and prognosis, and proposal of machine learning-based personalized screening strategies: insights from an EUSTAR study

Antonio Tonutti^1|2, Francesca Motta^1|2, Stefano Erba^1|2, Britta Maurer³, Cristiana Sieiro Santos³, Duygu Temiz Karadag³, Elena Rezus³, Elisabetta Zanatta³, Fabiola Atzeni³, Francesco Benvenuti³, Gabor Kumanovics³, Gabriella Szucs³, Gianluca Moroncini³, Gonçalo Boleto³, Leila Caillault³, Lilian Maria Lopez Nunez³, Masataka Kuwana³, Massimiliano Limonta³, Maurizio Cutolo³, Oliver Distler³, Radim Bevcar³, Roberto Giacomelli³, Rossella De Angelis³, Serena Guiducci³, Stefano Stano³, Valeria Riccieri³, Yasser El Miedany³, Sophie Blaise³, Carlo Selmi^1|2, Maria De Santis^1|2 | ¹Department of Biomedical Sciences, Humanitas University Pieve Emanuele, Milan, Italy; ²Rheumatology and Clinical Immunology, IRCCS Humanitas Research Hospital Rozzano, Milan, Italy; ³EUSTAR collaborators EUSTAR * OTHER

Background. Cancer may precede, follow, or coincide with systemic sclerosis (SSc) onset. Identification of risk factors has yielded heterogeneous

Results. This study aimed to evaluate risk factors for cancer in SSc, the impact of immunosuppressants, disease trajectories, and machine learning for tailored screening.

Methods. This was a 1:1 case-control study, including SSc patients with/without cancer. Cancers were defined as interceptable if they were synchronous (diagnosed within 3 years before or after SSc) or subsequent (>3 years after SSc); otherwise, cancers were categorized as previous (>3 years prior to SSc). Comparisons, multivariable logistic regression, mixed effect modeling, and survival analysis were performed to identify the features associated with the different outcomes. Machine learning models (Random Forest and XGBoost) were tuned to predict interceptable cancers.

Results. The study included 588 patients: 89% females; median age 67 (IQR 58-75), disease duration 12 years (6-19); 34% anti-Scl70+, 43% ACA+. Diffuse SSc was found in 26%, digital ulcers in 24%, ILD in 35%, PAH in 8%, esophageal involvement in 57%. Of 295 cancer cases, these were synchronous in 82 patients, previous in 55 (median 9 years before SSc), subsequent in 158 (median 11 years after SSc). Breast cancer was the most common (32%), followed by lung (16%), gynecological (8%), colorectal (7.5%), and hematological (7%). Risk factors for synchronous cancers included smoke (OR 2.0; 95%CI 1.1-3.5), diffuse SSc (OR 1.9; CI 1.1-3.4), anti-POLR3 (OR 2.6; CI 1.2-5.7), anti-PM/Scl (OR 4.7; CI 1.2-19), anti-RNP (OR 6.8; CI 1.6-29), and high CRP (OR 1.9; CI 1.1-3.5) while digital ulcers were protective (OR 0.5; CI 0.2-0.9). Risk factors for interceptable cancers were similar to those of synchronous cancers. Patients with previous cancer showed no difference compared to cancer-free patients, but higher frequency of anti-POLR3 (16% vs 6.8%) and less ulcers (15% vs 28%; both p<0.05). Treatment with cyclophosphamide was associated with subsequent cancers (OR 2.2; CI 1.1-4.8), mycophenolate was protective (OR 0.5; CI 0.3-0.9). After adjusting for age and disease duration, cancer increased the risk of mortality (HR 2.4; CI 1.1-5.0): interceptable (HR 3.3; CI 1.5-7) and subsequent (HR 4.2; CI 1.9-9.5) cancers had the greatest impact. Longitudinal follow-up revealed a higher incidence of muscle atrophy in cancer patients (OR 1.95). Random Forest and XGBoost demonstrated good performance in identifying interceptable cancers (AUCROC 0.88 both). Random Forest had higher sensitivity (98% vs 90%) and precision (94.5% vs 77%) but lower specificity (44% vs 66%). Key predictors included ILD, digital ulcers, esophageal involvement, telangiectasia, and high CRP.

Conclusions. Emerging risk factors for cancer in SSc include autoantibodies, disease features, and immunosuppressants, such as cyclophosphamide, but not mycophenolate. Malignancy influences SSc survival, and interceptable cancers exert the greatest impact. Machine learning approaches hold promises to improve early cancer detection.