Identifying sex-specific symptom-based predictors of lung cancer risk using explainable machine learning.

Abstract

e22575 Background: Lung cancer presents with a broad range of symptoms, many of which are nonspecific or atypical, contributing to delayed diagnosis. Symptom presentation and clinical recognition may differ between men and women, with important implications for diagnostic pathways. However, most symptom-based lung cancer risk models include sex only as a covariate, implicitly assuming similar symptom predictors across sexes. We evaluated sex-specific symptom predictors of lung cancer risk using explainable machine learning. Methods: We analyzed a publicly available lung cancer dataset comprising demographic characteristics and self-reported symptoms, including smoking status, respiratory symptoms, systemic features, and pain-related complaints (Biswas & Nath, 2024). Patients were stratified by sex, and separate gradient-boosted decision tree models were trained for males and females. Model performance was evaluated using 5-fold cross-validated area under the receiver operating characteristic curve (AUC). SHapley Additive exPlanations (SHAP) were used to quantify and compare symptom-level contributions to lung cancer risk within sex-specific models. Results: Sex-stratified models demonstrated comparable discrimination for lung cancer risk prediction. Explainable feature attribution revealed marked differences in symptom importance between sexes. Among females, age was the most influential predictor, followed by a heterogeneous mix of respiratory and non-respiratory features, including cough, chronic disease burden, chest pain, fatigue, and swallowing difficulty; smoking contributed relatively less to overall risk prediction. In contrast, male risk prediction was dominated by respiratory symptoms—wheezing, cough, and shortness of breath—along with behavior-associated factors such as smoking, alcohol use, and peer pressure. Both the ranking and magnitude of symptom contributions differed substantially between male and female models. Conclusions: Symptom-level predictors of lung cancer risk differ meaningfully between men and women. Explainable machine learning identified a more heterogeneous, non-respiratory symptom profile among women, compared with predominantly respiratory-driven risk patterns in men. Incorporating sex-specific symptom models may improve early risk stratification and support timelier recognition of lung cancer, particularly in patients presenting with non-classic symptoms.