Clinical Anchoring — DSM-5 Anchor Projection for Depression Detection

Abstract

Standard sentence embeddings compress clinical signals into opaque high-dimensional spaces where syntactic structure masks symptom evolution. We introduce anchor projection — a coordinate system transformation implemented natively in ChronosVector (CVX) — that re-expresses user trajectories relative to DSM-5 symptom reference vectors. On the eRisk depression detection task with a proper temporal split (train: 2017+2018, test: 2022), anchor-projected features achieve F1=0.744 and AUC=0.886, compared to F1=0.600 with absolute temporal features alone. Early detection at 10% of post history yields F1=0.673.

Related Work

• eRisk shared task (Losada et al., 2017-2022): Early risk detection of depression from social media posts. Best systems use transformer-based classifiers on user-level features.
• Concept-based explanations (Kim et al., TCAV, 2018): Testing with Concept Activation Vectors measures model sensitivity to human-defined concepts. Our anchor projection applies a similar idea at the data level.
• Clinical NLP for mental health (Coppersmith et al., 2018; Harrigian et al., 2020): Feature engineering from social media text for mental health detection. Most approaches treat users as static feature vectors.
• Temporal dynamics in depression (De Choudhury et al., 2013): Pioneering work on temporal patterns in social media for depression, but without vector trajectory analysis.

Methodology

Data

• eRisk dataset: 1.36M Reddit posts from 2,285 users
• Embeddings: MentalRoBERTa (D=768), pre-trained on mental health corpora
• Subset: 466 balanced users (233 depression, 233 control), 225,962 posts

CVX Pipeline

1. Index Construction: bulk_insert 225K vectors with save/load caching (avoids 500s rebuild)
2. DSM-5 Anchor Vectors: 9 symptom anchors (depressed mood, anhedonia, sleep disturbance, fatigue, worthlessness, concentration, suicidal ideation, appetite, psychomotor) + 1 healthy baseline, each encoded as MentalRoBERTa centroid of 3-5 representative phrases
3. Anchor Projection: cvx.project_to_anchors(trajectory, anchors, metric='cosine') → trajectory in ℝ¹⁰
4. Feature Extraction: cvx.anchor_summary() (mean, min, trend per anchor), cvx.hurst_exponent() on projected trajectory, cvx.velocity() in anchor space, topic polarization (dispersion), velocity differential (cvx.drift() for consecutive posts)
5. Classification: Logistic Regression with class_weight='balanced', both 5-fold CV and temporal split

Temporal Split (No Contamination)

• Train: eRisk 2017+2018 editions (226 users)
• Test: eRisk 2022 edition (236 users) — completely unseen

Key Results

Feature Ablation (5-Fold CV)

Model	F1	AUC	Precision	Recall
B1 Baseline (absolute features)	0.600	0.639	0.590	0.614
Anchor Only	0.746	0.849	0.739	0.759
Polarization Only	0.599	0.665	0.653	0.556
Velocity Only	0.547	0.554	0.520	0.582
Combined (B2)	0.781	0.863	0.775	0.789

Temporal Split (Train 2017+2018 → Test 2022)

Metric	Value
F1	0.744
Precision	0.659
Recall	0.856
AUC	0.886

Early Detection

% of Posts	F1	AUC
10%	0.673	0.788
20%	0.694	0.811
30%	0.719	0.829
50%	0.729	0.858
100%	0.753	0.895

CVX Functions Used

Function	Role
project_to_anchors(metric='cosine')	Transform ℝ⁷⁶⁸ → ℝ¹⁰ symptom coordinates
anchor_summary()	Mean, min, trend, last distance per anchor
hurst_exponent(projected)	Persistence of approach to depression
velocity(projected)	Rate of change in symptom space
drift(post_t, post_t+1)	Consecutive cosine displacement
save() / load()	Index persistence (avoid 500s rebuild)

Running the Notebook

conda activate cvx
cd crates/cvx-python && maturin develop --release && cd ../..
jupyter notebook notebooks/B2_clinical_anchoring.ipynb

Interactive Visualizations

See the full tutorial with Plotly plots.

---

← Back to White Paper