Which self-disclosure cues are associated with stronger visible community responses to posts by family caregivers of patients with breast cancer? An interpretable machine learning study.
AI interpretation is pending for this paper.
Open original publication →What the AI sees
Not AI summarized yet.
Research significance
Pending deeper interpretation.
Source abstract
OBJECTIVE: To identify key self-disclosure cues associated with stronger visible community responses in posts by family caregivers of breast cancer patients. METHODS: We conducted a retrospective analysis of posts from a breast cancer-related online community on Baidu Tieba. After cleaning and screening for caregiver authorship, cues were coded across content, emotion, motivation, and form. Following Lasso feature selection, logistic regression, random forest, and XGBoost were compared. Model interpretability was examined using SHAP and accumulated local effects (ALE). RESULTS: The final sample included 3,730 posts, mostly by adult children (70.8%). Lasso retained 32 of 33 features. XGBoost achieved the highest AUC (0.6653), while logistic regression had the highest recall. Robust features were identified by integrating both models. Children, Partner, Happiness, Economic Status, and History and Symptoms showed ORs of 1.46-2.26 (all P < 0.05) and ranked in the top 15 across both models. Examination and Diagnosis, Disease-related Images, and cues for seeking emotional, informational, and instrumental support were also positively associated (OR = 1.33-1.44; all P < 0.05 except informational support, P = 0.057), although their cross-model ranking consistency was generally weaker than that of the core features. SHAP and ALE analyses further suggested that Examination and Diagnosis, Disease-related Images, and emotional and informational support-seeking cues were positively associated with stronger visible community responses, whereas instrumental support-seeking showed a weaker and more heterogeneous pattern. Text Length showed a threshold-like nonlinear pattern. CONCLUSIONS: Children, Partner, Happiness, Economic Status, and History and Symptoms were the most robust positive features. Overall, stronger visible community responses were associated less with disclosure frequency than with contextual clarity, identifiable needs, and clear response entry points. These findings may inform response guidance and platform design for family caregivers.