Population-Level Profiling of DSM-5 Depressive Symptoms Among Self-Reported ADHD and ASD Users on Twitter: An Exploratory Study Using Advanced NLP and Statistical Analysis

Technical Summary: Population-Level Profiling of DSM-5 Depressive Symptoms in ADHD and ASD Twitter Users

Problem Statement
Depression frequently co-occurs with attention-deficit/hyperactivity disorder (ADHD) and autism spectrum disorder (ASD). While epidemiological data confirms elevated rates of depressive symptoms in both populations, quantitative evidence regarding how the phenomenology of depression differs between them at a population level remains scarce. Specifically, there is a lack of large-scale studies that simultaneously profile all nine DSM-5 major depressive episode symptoms to determine if self-reported ADHD and ASD users express these symptoms differently in naturalistic language. Existing literature often focuses on lexical or topical differences rather than formally defined symptom dimensions, and clinical accounts suggesting differences in executive dysfunction, burnout, or reward sensitivity lack population-scale validation.

Methodology
The study employed a two-stage Natural Language Processing (NLP) pipeline to analyze 1,282,437 tweets from 792 self-reported users (622 ADHD, 170 ASD) sourced from a public diagnosis-disclosure dataset.

1. Data Pre-processing and Filtering:

   • Stage 1 (Zero-Shot Screening): A zero-shot Natural Language Inference (NLI) model (cross-encoder/nli-deberta-v3-large) screened tweets for depressive relevance. The model judged whether a tweet entailed the hypothesis "DSM-5 depressive symptom present." To ensure robustness, the authors did not rely on a single cutoff but swept the probability threshold across five values (0.45 to 0.65). This stage prioritized recall to avoid permanently discarding relevant data.
   • Stage 2 (Symptom Classification): Tweets passing the pre-filter were classified into the nine DSM-5 symptoms (depressed mood, anhedonia, appetite change, sleep issue, psychomotor disturbance, fatigue, worthlessness, cognitive issues, suicidal ideation) using a fine-tuned MentalRoBERTa model. The model was trained on the ReDSM5 corpus, an expert-annotated dataset of Reddit sentences labeled by a licensed psychologist.
   • Calibration: The Stage 2 classifier utilized a per-label decision threshold calibration (ranging from 0.150 for suicidal thoughts to 0.825 for depressed mood) to maximize binary F1 scores for each symptom independently, addressing class imbalance.

2. User-Level Profiling:

   • Per-user symptom profiles were constructed by averaging the raw sigmoid scores of gated tweets (tweets passing both the zero-shot filter and the per-label symptom threshold).
   • Users with fewer than 30 qualifying tweets were excluded.
   • User-Level Centering: To isolate relative symptom emphasis from overall depressive language intensity, each user's profile was centered by subtracting their own mean across the nine symptoms.

3. Statistical Analysis and Robustness:

   • Discriminative Analysis: L1-penalized (LASSO) logistic regression was used to distinguish ADHD from ASD users based on centered symptom profiles. The model was evaluated using 5-fold cross-validation.
   • Robustness Criteria: To ensure findings were not artifacts of specific threshold choices, the authors applied a graded robustness scheme. Findings required Level 1 status (identical coefficient sign across all five thresholds AND bootstrap selection frequency ≥0.90 at every threshold) to be considered primary findings. Level 2 status required consistent signs but allowed for lower selection frequencies or intervals including zero.
   • Co-occurrence Analysis: Pearson correlations were computed between all 36 pairs of symptoms for each group. Bootstrap resampling (1,000 iterations) was used to assess the stability of correlation coefficients and the significance of differences between groups.

Key Contributions

1. Population-Level Characterization: The study provides an exploratory, population-level profile of how the nine DSM-5 depressive symptoms are differentially emphasized between self-reported ADHD and ASD communities on Twitter.
2. Advanced Classification Pipeline: It introduces a two-stage pipeline coupling zero-shot NLI pre-filtering with a MentalRoBERTa multi-label classifier fine-tuned on ReDSM5, demonstrating that domain-adapted transformers can operationalize DSM-5 criteria with high sentence-level performance (Macro-F1 = 0.901).
3. Graded Robustness Scheme: The paper contributes a methodological framework that explicitly separates findings supported by both bootstrap stability and cross-threshold consistency from those that are unstable, moving beyond single-threshold analyses.
4. Co-occurrence Analysis: It adds a user-level analysis testing whether the structural organization (pairwise co-occurrence) of depressive symptoms differs between the two groups.

Results

• Classifier Performance: The fine-tuned MentalRoBERTa model achieved a Macro-F1 of 0.901 on the held-out test set, substantially outperforming previous benchmarks on the ReDSM5 corpus.
• Discriminative Profiles: The logistic regression yielded modest but stable discrimination (ROC-AUC 0.645–0.653). Six symptoms met Level 1 (Robust) criteria:
  • ADHD-leaning: Cognitive issues, sleep disturbance, appetite change, and fatigue.
  • ASD-leaning: Suicidal ideation and anhedonia.
  • Level 2 (Directionally Consistent): Psychomotor disturbance showed an ASD-leaning trend but with slightly lower stability.
  • Symptoms like worthlessness and depressed mood did not meet robustness criteria due to unstable signs or low selection frequencies across thresholds.
• Co-occurrence Structure: The analysis revealed a largely shared symptom co-occurrence backbone between the groups. A somatic cluster (fatigue, sleep, psychomotor, appetite) and a mood trade-off structure were observed in both. Out of 36 symptom pairs, 17 were individually bootstrap-robust in both groups with no significant between-group difference. No pair met the strict criterion for a disorder-specific co-occurrence pattern (sign-stable difference with 95% CI excluding zero at all thresholds).

Significance and Claims
The authors frame this work strictly as exploratory and population-level. They explicitly state that the observed differences reflect population-level reproducibility rather than clinical validity or individual diagnostic capability. The modest AUC scores indicate that while the signal is stable across analytic thresholds, it is not sufficient for individual screening or diagnosis.

The significance of the findings lies in identifying systematic, reproducible patterns in how depressive experiences are communicated across neurodevelopmental conditions. The results suggest that while the core structure of symptom co-occurrence is shared, the relative emphasis on specific dimensions (e.g., cognitive/somatic issues in ADHD vs. anhedonia/suicidal ideation in ASD) differs. These findings are intended to serve as corroborating observational sources for clinical hypotheses regarding phenomenological differences between ADHD and ASD, rather than as evidence of distinct depressive phenomenology at the individual level. The study acknowledges limitations including reliance on self-reported diagnoses, potential selection bias, and the lack of a neurotypical control group.