AI Transparency Policy

1. Our Approach to AI

Alva Health is committed to using AI responsibly, transparently, and in a way that genuinely benefits users.

• Transparency: we tell you when AI generates content or recommendations
• Human oversight: consequential AI outputs are subject to human review
• Privacy: your health data is not sold or shared for AI training by third parties
• Non-discrimination: we actively monitor EVA for demographic bias
• Wellness only: EVA is a wellness tool and does not make medical decisions

2. EVA and EPIC 22

EVA is powered by EPIC 22, Alva Health’s proprietary metabolic intelligence core. EVA analyses CGM data, sleep, activity, dietary, and heart rate data to deliver personalised metabolic insights and predictions.

EVA does not: diagnose conditions · recommend or adjust medications · make clinically or legally significant decisions · share identifiable data with third parties for model training.

3. How EVA Makes Decisions

EVA uses a deterministic, pattern-based intelligence pipeline — not a black-box large language model.

• EVA’s reasoning is structured around validated metabolic relationships
• Recommendations are generated from your personal data patterns, not population averages alone
• The pipeline includes safety guardrails preventing recommendations outside wellness parameters
• Any natural language output from EVA is generated by a controlled, medically-reviewed output layer, not an unconstrained LLM

4. Data Used to Train and Operate EVA

5. Human Oversight and Model Validation

5.1 Human oversight

• EVA’s recommendation boundaries are set and reviewed by qualified medical and scientific advisors
• Automated monitoring flags unusual outputs for human review
• Users can always escalate concerns about an EVA recommendation to our support team
• EVA is programmed to recommend professional consultation in defined high-risk scenarios

5.2 Model validation

EVA’s intelligence model is subject to the following ongoing validation processes:

• Internal accuracy benchmarking: prediction accuracy is measured against anonymised historical outcomes on a quarterly basis
• External validation: we are building a clinical validation programme with academic partners (see Section 11)
• Regression testing: all model updates are regression-tested before deployment to ensure existing accuracy is maintained or improved
• User feedback loop: EVA’s outputs are continuously refined based on anonymised user outcome data

6. Bias Monitoring

We are committed to identifying and mitigating bias in EVA’s outputs. Our bias monitoring programme includes:

• Demographic analysis: we monitor EVA’s prediction accuracy across age groups, biological sex, and geographic regions
• Data representativeness review: we assess whether our training data is representative of the populations EVA serves
• Output auditing: we conduct quarterly audits of EVA’s recommendations to identify systematic patterns that may indicate bias
• Remediation: where bias is identified, we update the model and document the remediation in our AI incident register

7. Confidence Scoring

EVA applies confidence scoring to its predictions to signal reliability to users. The confidence score reflects:

• The amount of personal data EVA has collected from you (more data → higher confidence)
• The consistency of the pattern EVA has detected
• The historical accuracy of similar predictions for similar metabolic profiles

Confidence scores are visible within the platform where relevant and are factored into the language EVA uses to communicate predictions (e.g. ‘likely’ vs ‘you will’ vs ‘possibly’).

8. AI Incident Reporting

We maintain an AI incident register that records:

• Instances where EVA generates outputs outside its defined safety parameters
• User-reported concerns about EVA’s accuracy or appropriateness
• Identified model biases and remediation actions
• Any third-party audit findings relating to EVA’s performance

AI incidents are reviewed by our technical and medical advisory team. Significant incidents are escalated to senior leadership and, where appropriate, reported to relevant regulatory authorities.

To report an AI incident or concern: ai@alva.health

9. Accuracy and Limitations

Known limitations:

• Accuracy may be reduced during illness, significant life events, or major lifestyle changes
• CGM sensor accuracy affects EVA’s input data quality
• EVA is not trained for clinical populations and is not appropriate without clinical oversight for Type 1 diabetes, pregnancy, or active metabolic disease
• EVA’s current model is primarily trained on EU and US adult population data

10. EU AI Act Compliance

We assess EVA as a limited-risk AI system under the EU Artificial Intelligence Act (Regulation 2024/1689). EVA does not constitute a high-risk AI system under Annex III as it does not make legally significant decisions in regulated domains.

Our transparency obligations include informing users when they are interacting with an AI system. This policy forms part of those obligations.

11. Clinical Validation Roadmap

We are committed to building a robust clinical evidence base for EVA’s predictions and recommendations. Our roadmap:

12. Your Rights Regarding AI

• The right to be informed when AI generates recommendations affecting you
• The right to request human review of any EVA recommendation you dispute
• The right to opt out of anonymised data use for model improvement
• The right to report AI concerns at ai@alva.health
• The right to understand the general logic behind EVA’s recommendations

13. Trademarks

Trademarks

ALVA® and EVA® are registered trademarks of Alva Health GmbH. All rights reserved. Unauthorised use of these trademarks, including in domain names, product names, or marketing materials, is strictly prohibited without the prior written consent of Alva Health GmbH.

All other trademarks, service marks, and trade names referenced on the platform are the property of their respective owners.

14. Contact

Legal enquiries: legal@alva.health

Data protection / DPO: dpo@alva.health

General support: support@alva.health

Security issues: security@alva.health