AI Transparency Statement
Model card and AI system documentation — last updated 5 April 2026
This application uses AI in two ways. We have assessed the risk level of each component under the EU AI Act risk framework:
| Component | Function | Risk Level |
|---|---|---|
| Rule-based scoring engine | Calculates wellbeing scores from questionnaire responses using fixed clinical thresholds | Limited |
| GAM/EBM interpretable model | Identifies which factors contribute most to distress (feature attributions via SHAP) | Limited |
| Qdrant vector search | Finds anonymised similar profiles from UK health survey data | Limited |
| OpenAI GPT-4o-mini | Generates plain-English explanations of results (optional, user-initiated) | Limited |
Classification rationale:This app does not diagnose, treat, or make clinical decisions. All outputs are informational and explicitly labelled as "not medical advice." The scoring uses fixed clinical thresholds from UK guidelines (NICE, CMO, SACN), not predictive ML models. The interpretable model provides explanatory attributions only — it does not determine risk levels. For these reasons, we classify the system as limited risk under the EU AI Act, subject to transparency obligations under Article 50.
If the system were used to make or influence clinical decisions, it would be reclassified as high-risk and require a conformity assessment.
Following Mitchell et al. (2019) model card format
Model Details
- Model: ExplainableBoostingClassifier (GAM/EBM) from InterpretML
- Version: 1.0 (April 2026)
- Task: Binary classification — predict GHQ-12 caseness (psychological distress threshold ≥4)
- Features: 14 (age, BMI, 12 GHQ-12 items, PA guidelines, IMD quintile)
- Output: Feature importance scores and SHAP values for explanation — NOT used for risk classification
Training Data
- Source: Health Survey for England 2022 (HSE) — UK Data Service Study 9469
- Population: 1,624 women aged 16-54 (from 9,122 total respondents)
- Train/test split: 80/20 stratified (853 train, 214 test)
- Distress prevalence: 22.7% (caseness ≥4)
- Licence: UK Data Service End User Licence
Performance
| Metric | Value |
|---|---|
| Accuracy | 1.000 |
| F1 (weighted) | 1.000 |
| Test set size | 214 |
Note:The high accuracy is expected because GHQ-12 items directly determine the caseness score (it is a deterministic relationship). The model's value is in SHAP feature attributions — identifying which items contribute most to distress — not in prediction.
Top Contributing Features (SHAP)
- Unhappiness (GHQUNHAP) — 0.044
- Sleep loss (GHQSLEEP) — 0.043
- Confidence (GHQCONFI) — 0.041
- Concentration (GHQCONC) — 0.040
- Overcoming difficulties (GHQOVER) — 0.038
Intended Use & Limitations
- Intended for: General wellbeing information for women of reproductive age in the UK
- NOT intended for: Clinical diagnosis, treatment decisions, or replacing professional medical advice
- Limitation: Training data is cross-sectional (single time point), not longitudinal
- Limitation: Age is banded (5-year groups), not continuous, due to HSE data anonymisation
- Limitation: No pregnancy-specific sub-model (only 41 pregnant women in HSE 2022)
- Limitation: GHQ-12 is a screening tool, not a diagnostic instrument
Ethical Considerations
- No individual in the training data can be re-identified from the model or its outputs
- SHAP values are population-level averages, not individual predictions about specific users
- The app includes crisis resources (Samaritans 116 123) when high distress is detected
- Under-13s are blocked; 13-17 users see enhanced privacy protections
- No PII is sent to OpenAI — only anonymised scores
- Source: National Diet and Nutrition Survey (NDNS) Years 12-15 (2019-2023) — UK Data Service Study 6533
- Population: 755 women aged 16-54
- Nutrients tracked: Protein, iron, folate, fibre, vitamin D, calcium
- Thresholds: UK Reference Nutrient Intakes (RNI) from SACN, adjusted by life stage
Key finding: Most UK women fall below RNI for iron (mean 8.5mg vs 14.8mg target), folate (mean 184mcg vs 400mcg preconception target), fibre (mean 15.2g vs 30g target), and vitamin D (mean 2.3mcg vs 10mcg target).
The app optionally uses OpenAI GPT-4o-mini to generate plain-English explanations of assessment results. This feature:
- Is opt-in only — requires explicit user consent and a button press
- Sends only anonymised scores — no name, email, date of birth, or other PII
- Uses a strict 10-rule system prompt that prohibits speculation, diagnosis, and fabrication
- Operates in zero-retention mode — OpenAI does not store or train on the data
- All outputs are clearly labelledas "AI-generated" with a disclosure box
- Temperature is set to 0.3 (low creativity) to minimise hallucination risk
Users are advised to verify AI explanations against the data-based explanation always shown above it.
- Method: Cosine similarity on 7-dimensional standardised vectors
- Features: Age, BMI, GHQ-12 scores, physical activity, nutrition score, IMD deprivation
- Database: 1,624 anonymised profiles from HSE 2022
- Privacy:Only age bands (e.g. "30-34") and category labels are shown — no individual can be identified
- Consent: Opt-in only (optional consent checkbox)
The following safeguards ensure this system does not operate without human oversight:
- All risk levels and recommendations are determined by fixed clinical thresholds from UK guidelines, not by ML model predictions
- Every result page includes a "Why this result?" section explaining the exact calculation
- Users are directed to consult their GP, midwife, or health visitor for clinical guidance
- High-risk results include crisis contact numbers (Samaritans 116 123, NHS 111)
- The app never auto-refers, auto-prescribes, or auto-diagnoses
- Users can withdraw consent, export data, and delete their account at any time
If you believe this system has produced an incorrect, misleading, or harmful output, please report it via our Report Issue page or contact us at safety@mumvera.co.uk.