Towards ordinal suicide ideation detection on social media

Traditional suicide ideation detection approaches treat the problem as binary classification, overlooking the nuanced spectrum of suicidal thoughts and risk levels. We propose an ordinal approach to suicide ideation detection that recognizes different severity levels of suicidal ideation, from mild thoughts to immediate risk. Our method leverages ordinal regression techniques to capture the inherent ordering in suicide risk levels, providing more granular and clinically meaningful predictions. The approach uses deep learning architectures specifically designed for ordinal outcomes, incorporating both textual content and user behavioral patterns. By modeling suicide ideation as an ordinal problem, our system can provide more nuanced risk assessments that better align with clinical practice and enable more appropriate interventions. We evaluate our approach on social media datasets with expert annotations of risk levels and demonstrate improved performance over binary classification methods, particularly in distinguishing between different levels of suicide risk.