AI model development and certification for cardiovascular medicine

The Client

KARDI.AI is a cardiovascular health technology company building AI- powered tools for early cardiac event detection. Their flagship product captures ECG signals from a consumer wearable device and needs to identify clinically significant events including atrial fibrillation with the accuracy and auditability required for medical certification. Their product roadmap depended on solving a signal processing challenge that no existing commercial AI tool had addressed.
 

The Challenge

Consumer-grade wearable devices capture ECG signals under conditions that clinical monitors are not designed for. The patient is moving. The sensor contact is imperfect. Ambient noise is higher. The resulting signal has a different noise profile and lower fidelity than the data on which most cardiac AI models are trained which means standard models fail when applied to wearable data without significant adaptation.

KARDI.AI needed a model that could reliably detect atrial fibrillation and additional cardiac KPIs from this lower-fidelity signal at clinical accuracy standards. There was no off-the-shelf solution. Existing cardiac AI models were trained on hospital-grade ECG data and did not generalise to wearable signal characteristics. Designing and building from scratch was the only viable path. The regulatory context added further constraint.

A model used to support clinical decision-making in cardiovascular medicine must meet medical device software standards. That means rigorous validation methodology, comprehensive documentation, and an audit trail through every stage of development not just a high accuracy score on a test set.

The timeline pressure was real. KARDI.AI's product launch schedule depended on having a certified, production-ready model. There was no room for a long exploratory research phase.
 

What We Did

The Discovery Phase confirmed that no existing model architecture could be adapted cost-effectively to the wearable signal environment. The noise profile was sufficiently different from clinical ECG data that transfer learning from existing cardiac models would have required more remediation than building a purpose-designed architecture.

We designed a set of custom machine learning and deep learning models for ECG signal processing, engineering features specifically for the characteristics of the wearable sensor including noise filtering, baseline wander correction, and beat segmentation tuned to the device's sampling rate and electrode configuration. The models ware trained on a combination of available public cardiac datasets and KARDI.AI's proprietary wearable data, with careful attention to class imbalance given the relative rarity of atrial fibrillation events in the training population.

Validation was designed from the outset to meet medical device software standards. We maintained a held-out prospective test set, documented all training decisions and hyperparameter choices, and produced performance metrics across clinically relevant subgroups.
 

• Data assessment:
  Characterised the wearable signal noise profile; identified gaps in training data coverage; sourced supplementary public datasets
• Architecture design:
  Custom machine learning and deep learning models for time-series ECG classification, designed for wearable signal characteristics
• Feature engineering:
  Noise filtering, baseline correction, and beat segmentation tuned to the specific device
• Training:
  Multi-class model trained with class-weighted loss to address atrial fibrillation prevalence imbalance
• Validation:
  Prospective held-out test set; subgroup analysis; full documentation for regulatory submission
• Deployment:
  Production-ready API integrated into KARDI.AI's cloud platform infrastructure

The Outcome

Certified for as a Class 2A medical device for clinical use delivered within the product launch timeline

The models achieved clinical-grade accuracy on atrial fibrillation detection from consumer wearable ECG data, passing validation benchmarks required for medical device software certification. KARDI.AI launched on schedule.
 

• Models certified under applicable medical device software standards
• Clinical-grade sensitivity, positive predictability and specificity on wearable ECG signal
• Full regulatory technical documentation delivered alongside the model
• KARDI.AI expanded the engagement to additional cardiac KPI detection use cases following successful deployment

Client Voice

"What we'd been missing was a partner to help us automate the AI solution. We chose VECTOR Labs given their experience in ECG signal evaluation and their ability to deliver KPI detection in quite a short time."
Pavel Digana, KARDI AI Technologies
 

Facing a similar challenge?

If you're building a medical AI product that needs to work with real-world signal data and meet regulatory standards, we'd welcome a technical conversation. We've navigated this process and can help you do the same — without the false starts.

Book a Technical Consultation