Cognitive-congestion language profiling in ADHD: lexical signal auditing and negative-manifold specificity testing

Abstract

High machine-learning accuracy in classifying ADHD-related writing does not necessarily establish an ADHD-specific linguistic phenotype, as predictions may reflect diagnostic vocabulary, medication references, or community-specific topics rather than cognitive organization. This study evaluated an interpretable cognitive-congestion language profile (LCCI) and a negative-signature framework for detecting disorder-specific linguistic distinctiveness. A dataset of 25,000 Reddit-derived posts from ADHD, OCD, depression, PTSD, and Asperger/autism-related communities was analyzed using condition-name masking and author-grouped cross-validation. Twenty-five features measuring discourse structure, uncertainty, urgency, lexical diversity, and coherence were compared with TF-IDF classification. The equal-weight LCCI composite performed near chance (ROC-AUC 0.506), while the multivariable profile achieved ROC-AUC 0.654, increasing to 0.681 with learned weights. Individual feature differences were small. TF-IDF substantially outperformed LCCI (ROC-AUC 0.940), with negligible improvement from combining them. An audit classified 59 of 60 high-weight TF-IDF coefficients (98.3%) as topic-related, without quantifying their contribution to predictive performance. To test specificity independently of supervised classification, the negative-signature framework excluded each psychiatric community in turn, constructed a reference language manifold from the remaining communities, and measured residual distinctiveness using PCA reconstruction, shrinkage Mahalanobis distance, and isolation forest. ADHD exhibited no positive residual signal (ROC-AUC 0.451, 0.419, and 0.429, respectively), while OCD showed the greatest residual separability. These findings distinguish predictive discrimination, interpretable linguistic association, and disorder specificity. The principal contribution is a reproducible, generalizable falsification framework demonstrating that strong supervised classification does not establish an ADHD-specific linguistic phenotype and providing a means to evaluate candidate linguistic biomarkers before clinical translation.