Abstract
IMPORTANCE: Cancer is a life-threatening complication of dermatomyositis (DM) and contributes significantly to mortality. A validated association-based tool is urgently needed to optimize early cancer detection and reduce diagnostic delays. OBJECTIVE: To develop and validate a practical prediction model for cancer-associated likelihood in adult patients with DM. DESIGN, SETTING, AND PARTICIPANTS: A retrospective multicenter cohort study including adults with DM or clinically amyopathic DM was carried out. Participants were recruited from the Department of Dermatology at Ruijin Hospital (training cohort) and the Department of Rheumatology at Renji Hospital (validation cohort) over the period from 2015 to 2022. Multivariate logistic regression and machine learning techniques were employed for model development and validation. The analysis took place in 2024. EXPOSURE: DM and clinically amyopathic DM. MAIN OUTCOMES AND MEASURES: The primary outcome was the occurrence of cancer in patients with DM. Model performance was assessed using the area under the curve to evaluate predictive accuracy. RESULTS: A total of 546 adults with DM or clinically amyopathic DM were included, with a mean (SD) age of 49.8 (14.2) years, comprising 166 male individuals (30.4%) and 380 female individuals (69.6%). Five factors significantly associated with concomitant cancers in patients with DM were used to construct the TIP-CA model: anti-transcriptional intermediary factor 1-γ (TIF1-γ) antibody (positive scored as 1; negative scored as 0), interstitial lung disease (present scored as -1; absent scored as 0), poikiloderma (present scored as 1; absent scored as 0), DM subtypes (DM scored as 1; clinically amyopathic DM scored as 0), and anemia (present scored as 1; absent scored as 0). The model demonstrated good discriminatory capability, achieving an area under the curve of 0.809 and 0.808 in the derivation and validation cohorts, respectively. CONCLUSIONS AND RELEVANCE: This cohort study found that the TIP-CA model effectively stratified cancer-associated likelihood in patients with DM using routinely available clinical data. By using data from multidisciplinary patient cohorts and incorporating machine learning techniques, the model minimized referral bias. This proposed model may have the potential to guide clinicians in implementing targeted cancer screening strategies and improve patient outcomes.