JMIR Publications

ABSTRACT

Background:

Depression frequently co-occurs with attention-deficit/hyperactivity disorder (ADHD) and autism spectrum disorder (ASD). However, population-level differences in how depressive symptoms are expressed between these groups remains underexplored.

Objective:

This study examined whether social media users with ADHD and ASD differ in how they express Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) depressive symptoms in their tweets, focusing on the relative prominence of the nine symptoms at population level and testing whether observed differences persist across varying levels of depressive-content filtering.

Methods:

We analysed 1,282,437 tweets from 792 users (622 ADHD; 170 ASD) from a diagnosis-disclosure Twitter dataset. Tweets were pre-filtered for depressive relevance using zero-shot NLI, then classified into nine DSM-5 depressive symptoms using MentalRoBERTa fine-tuned on ReDSM5 (an expert annotated dataset). Nine-symptom profiles were created and mean-centered per user. We applied L1-penalised logistic regression with 5-fold cross-validation to distinguish between ADHD and ASD users, complemented by Pearson correlations to assess symptom co-occurrence. We tested the robustness of our approach across five filtering thresholds (0.45–0.65) using 1,000-resample bootstrapping and cross-threshold sign consistency.

Results:

For multi-label depression symptom classification, MentalRoBERTa achieved macro-F1 of 0.901 on a held-out set (>= 0.85 F1 on 8/9 symptoms), substantially outperforming the original ReDSM5 benchmark. ADHD vs ASD classification using fitted logistic regression yielded stable but modest performance (cross-validated ROC-AUC 0.645–0.653 across thresholds). Coefficient analysis revealed that cognitive issues, sleep issues, appetite change, and fatigue consistently leaned toward ADHD, while suicidal ideation and anhedonia consistently leaned toward ASD (bootstrap selection >= 0.90 across all thresholds). Psychomotor disturbance showed the same ASD-leaning directional effect with slightly lower stability. Correlation analysis revealed a largely shared symptom co-occurrence structure between groups (17 of 36 pairs individually bootstrap-robust in both groups and in the same direction); no pair met our pre-specified criterion for a robust disorder-specific difference

Conclusions:

Population-level differences in depression-related language between self-reporting social media users with ADHD and ASD were consistently observed across multiple analytic thresholds, indicating distinct patterns of depressive symptom expression. These differences reflect population-level reproducibility rather than clinical validity and should be interpreted as exploratory rather than as evidence of differing depressive phenomenology at the individual level.