Mental Health & Clinical NLP

Overview

CVX enables a fundamentally different approach to mental health detection from social media: instead of treating each post as an independent feature vector, it tracks how a user’s language evolves over time — velocity, drift direction, change points, and proximity to clinical symptom anchors.

This is the most mature application domain for CVX, validated on the eRisk shared task dataset (1.36M Reddit posts, 2,285 users) with MentalRoBERTa embeddings (D=768).

Key Findings

Embedding Anisotropy Correction (30x Signal Improvement)

MentalRoBERTa embeddings occupy a narrow cone in high-dimensional space — all pairwise cosine similarities are ~0.96 regardless of content. This makes raw anchor projections useless.

Centering (subtracting the global mean vector) amplifies the discriminative signal 30x:

Metric	Before centering	After centering
Depression user → depressed_mood anchor	cosine sim 0.975	cosine sim 0.42
Control user → depressed_mood anchor	cosine sim 0.964	cosine sim 0.09
Discriminative gap	0.011	0.33

This correction benefits ALL downstream CVX operations. See RFC-012 Part B for the academic references (Ethayarajh 2019, Su et al. 2021).

DSM-5 Symptom Proximity Profiling (B3)

9 DSM-5 symptom anchors + 1 healthy baseline, encoded as MentalRoBERTa centroids of representative clinical phrases. After centering, population-level profiles clearly separate depression from control:

Depression users: highest proximity to depressed_mood (0.35), worthlessness (0.31), anhedonia (0.28)
Control users: uniformly low proximity across all symptoms (0.05-0.09)
Drift direction: depression users show increasing proximity to symptoms over time; controls remain stable

Hierarchical Semantic Clustering (B1)

HNSW’s natural hierarchy produces semantic regions with strong clinical meaning:

Level 2 regions show clusters with depression ratios from 0.15 to 0.85
High-depression clusters contain posts about hopelessness, isolation, sleep disruption
Low-depression clusters contain posts about social activities, hobbies, future plans
This demonstrates unsupervised specialization — the graph structure separates clinical from non-clinical content without labels

Classification Results (B2)

Anchor-projected temporal features on a proper temporal split (train 2017+2018 → test 2022):

Model	F1	AUC	Precision	Recall
B1 Baseline (absolute features)	0.600	0.639	0.590	0.614
B2 Combined (anchor + polarization + velocity)	0.744	0.886	0.739	0.750
Early detection (10% of posts)	0.673	—	—	—

CVX Features Used

Feature	Purpose
`project_to_anchors()`	DSM-5 symptom proximity trajectories
`anchor_summary()`	Per-user mean, min, trend per symptom
`drift()` / `velocity()`	Rate and direction of linguistic change
`detect_changepoints()`	Onset/escalation event detection (PELT)
`regions()` / `region_assignments()`	Unsupervised semantic clustering via HNSW hierarchy
`hurst_exponent()`	Long-memory estimation (persistent vs antipersistent drift)
`path_signature()`	Trajectory shape classification
Centering (manual, RFC-012 pending)	Anisotropy correction for 30x signal improvement

Notebooks

Notebook	Focus	Status
B1_interactive_explorer	HNSW hierarchy visualization, depression ratio per cluster	Best: cluster visualization
B2_clinical_anchoring	Anchor projection pipeline, classification benchmarks	Complete
B3_clinical_dashboard	DSM-5 radar, symptom drift direction, clinical timeline	Best: population profiles, drift tracking
B1_erisk_rigorous	Rigorous eRisk evaluation (SVC+PCA, no data balancing)	Complete