Non-profit · Free for educators · Open methodology

Discovering what
words don't say

Understanding students  ·  Empowering teachers

AI that reads psycholinguistic signals in student writing to flag mental health distress early — so the teacher can respond first.

90.5%

Accuracy on real student text

40

Psycholinguistic features

24.7K

Student essays audited for fairness

0 data

Stored from your students

The Problem

Every classroom has students
silently struggling

Mental health distress often surfaces in how students write — word choice, sentence patterns, emotional tone — long before they seek help. Teachers see these words every day and miss the signal buried inside.

Invisible Signals

A student's essay about summer might contain death-word density 4× above baseline. Without computational analysis, no teacher can catch this pattern across 30 students.

30 Students, 1 Teacher

A typical BC teacher reads hundreds of pieces of student writing each week. Systematic early detection is humanly impossible without a tool.

Late Intervention

Without early signals, teachers and counselors only respond after a crisis. UEdu aims to shift the window — from reaction to prevention.

How It Works

Three layers. No black box.

UEdu uses a deterministic, auditable pipeline. Every score is explainable — no opaque AI guesswork.

Step 1

Psychology Discovers

40 psycholinguistic features extracted deterministically — death-word density, self-focus pronoun ratio, cognitive distortion patterns, temporal urgency, and more.

Affective Social Cognitive Structural
Step 2

Code Measures

XGBoost classifier trained on 232,000 Reddit mental health posts scores each feature vector. The model is static, auditable, and never updates on student data.

90.47% accuracy · 0.964 AUC
Step 3

AI Translates

Claude (Claude AI) translates the signal into actionable, empathetic teacher guidance — constrained by MHFA and BC ERASE Bullying knowledge bases. The teacher decides.

Grounded in MHFA · BC ERASE

Design principle: "Psychology discovers. Code measures. AI translates. The teacher decides."  —  UEdu is a first-response companion, not a diagnostic tool. All flags require human review.

Research Results

Five experiments.
Rigorous methodology.

Validated on 232,000 Reddit posts, 24,700 ASAP student essays, and 200 real-world samples.

Exp 1

Ablation Study

232,000 Reddit posts · 5-fold CV

Best: 90.47%
M6: Hybrid XGBoost 90.47% · AUC 0.964
M5: Psycho-only XGBoost 87.26% · AUC 0.941
Exp 2

Domain Transfer

256 synthetic student texts

F1 0.987
Hybrid v3 (GPT + XGB)
F1 0.987 FPR 3.1%
GPT-4o-mini zero-shot
F1 0.967 FPR 18.8%
M5 (XGBoost + Psycho)
F1 0.900 FPR 43.8%
Exp 4

SHAP Feature Importance

Top signals driving detection

A8: Death word ratio SHAP 1.097
SO6: Self vs. other ratio SHAP 0.369
LP1: Function word ratio SHAP 0.317

Social pronoun patterns are 2× more diagnostic than emotional vocabulary alone.

Exp 5

Head-to-Head Evaluation

200 real Kaggle samples

Hybrid v3 ★
Recall 95% FPR 14%
M6 XGBoost
Recall 89% FPR 6%
GPT-4o-mini
Recall 99% FPR 56%

Hybrid v3 recommended for schools: misses almost no real distress (5% FNR) while keeping false alarms manageable.

Exp 3 · Fairness Audit

Built for
every student —
especially ESL learners

We audited UEdu against 24,728 real ASAP student essays across five demographic dimensions. Our hypothesis was that ESL students might get higher false-positive rates due to language differences.

The result surprised us. ESL students had a lower false positive rate than native English speakers (0.72×, p = 0.022). The model detects psychological distress signals, not writing ability.

ELL students: FPR 1.74% vs. non-ELL 2.40% (M5)
H3 (ESL bias hypothesis) rejected — statistically significant
Race/ethnicity gaps detected (p < 0.001) — active monitoring required

False Positive Rate by Group (M5)

ELL (English Language Learner) 1.74%
Non-ELL 2.40%
Male students 2.64%
Female students 2.02%

Lower FPR = fewer false alarms. All rates < 3%, well within acceptable clinical thresholds.

Ongoing commitment: Fairness audits will be re-run with each model update. UEdu will never ship a model with statistically significant demographic bias above predefined thresholds.

Interactive Demo

Try it yourself

Paste a writing sample and see UEdu analyze psycholinguistic features in real time.

Live Demo — Hybrid v3 vs GPT-4o-mini

Paste any student writing sample and see UEdu's Hybrid model and GPT-4o-mini analyze it side by side, with psycholinguistic feature breakdowns and teacher guidance.

Open Interactive Demo

No text is stored or used for training. All analysis is ephemeral.

About the Project

Uplifting every school

Student-Led, Public Good

Built by a high school student in BC, UEdu is offered free to all educators. No venture funding, no data harvesting — just a genuine effort to help schools support student wellbeing.

Open Methodology

All 40 features are documented and auditable. The training dataset (232K Reddit posts) is publicly available on Kaggle. No black-box AI.

Privacy First

Student text is analyzed locally — never stored, never sent to cloud services for training. PIPEDA-compliant by design.

Ready to support
your students?

UEdu is built for BC teachers. Join our early access list to be the first to pilot the tool in your classroom.

No spam. Unsubscribe anytime.