On-Device Training

The forward and backward passes of the training loop execute entirely on the device's processor. Loss calculation and gradient computation occur against local data, ensuring raw patient vitals, imaging, or genomic sequences never leave the device. Only the resulting weight updates—not the data itself—are shared in federated settings. This is the foundational mechanism that enforces data locality and regulatory compliance.

A globally initialized model adapts to a specific patient's physiology or a device's unique sensor drift through additional on-device training steps. This combats the non-IID data problem inherent in healthcare, where a generalized model may fail on outlier populations. Techniques like Elastic Weight Consolidation (EWC) prevent catastrophic forgetting of global knowledge during this personalization phase.

An alternative to sharing model weights, this paradigm exchanges only the model's output predictions on a public, unlabeled calibration dataset. A local student model is trained to match the aggregated predictions of peer models. This supports heterogeneous compute environments where devices may have vastly different neural architectures, and it provides an additional layer of privacy by obfuscating the model's internal parameters.

The training process is designed with explicit awareness of the target silicon's limitations. The latency budget, memory footprint, and power envelope of a microcontroller or NPU are incorporated directly into the loss function or neural architecture search. This ensures the personalized model remains deployable on a battery-operated medical wearable without exceeding its thermal or energy constraints.

In a federated loop, locally computed model updates are encrypted and sent to an aggregation server. Techniques like secure multi-party computation or differential privacy are applied to the updates before aggregation. A differential privacy accountant tracks the privacy budget spent during each round of on-device training, providing a mathematical guarantee against membership inference attacks on the training data.

The device operates in a perpetual loop of inference and training, continuously adapting to concept drift in the patient's physiological signals. An Over-the-Air (OTA) update mechanism securely delivers an improved base model from the cloud, which the device then uses as a new starting point for further on-device personalization. A watchdog timer ensures the training process does not compromise the device's fail-safe, real-time monitoring functions.

The process of adapting a globally trained federated model to a specific patient population or device's data distribution through additional on-device training steps. Unlike full retraining, local fine-tuning typically uses a lower learning rate and fewer epochs to prevent overfitting to the small, personalized dataset. This technique is critical for personalized healthcare models that must adapt to individual physiological baselines—such as a cardiac monitor learning a specific patient's normal heart rate variability—without exposing that sensitive data to external servers.

The tendency of a neural network to abruptly and completely overwrite previously learned knowledge when trained on new data. In on-device training, this manifests when a model personalizes to a new patient's data and loses its ability to generalize. Mitigation strategies include:

• Elastic Weight Consolidation (EWC): Constrains important parameters to stay near their old values
• Experience Replay: Maintains a small buffer of representative past samples
• Progressive Neural Networks: Freezes old task weights and adds new capacity for new tasks

This is a fundamental challenge for continuous on-device model personalization in medical wearables.

A regularization technique that mitigates catastrophic forgetting by constraining important model parameters to stay close to their old values while learning new tasks on a device. EWC calculates the Fisher information matrix to identify which weights are critical for previously learned tasks. During subsequent training, a quadratic penalty term is added to the loss function, proportional to how important each weight was. This allows the model to learn new patient-specific patterns while preserving its foundational diagnostic capabilities, making it essential for lifelong learning on medical edge devices.

A federated learning paradigm where local models share their output predictions (logits) on a public, unlabeled dataset instead of sharing model weights or gradients. This approach offers several advantages for on-device training:

• Supports heterogeneous model architectures across different devices
• Provides an additional layer of privacy protection by avoiding direct weight transfer
• Reduces communication overhead compared to transmitting full model updates

In healthcare, this allows a powerful server-side teacher model to transfer knowledge to lightweight on-device student models without exposing patient data.

A mechanism for remotely deploying new firmware or updated machine learning models to a fleet of distributed medical devices via a wireless network connection. OTA systems must ensure:

• Atomic updates: Either fully applied or fully rolled back to prevent bricked devices
• Delta updates: Transmitting only the difference between old and new model weights to conserve bandwidth
• Staged rollouts: Deploying to a small subset of devices first to validate safety

For FDA-regulated medical devices, OTA pipelines must include cryptographic signature verification and maintain a complete audit trail of all model versions deployed to each device.

The principle of processing and storing data physically close to its source on the edge device or local gateway, a fundamental requirement for compliance with strict healthcare data residency regulations. On-device training is the ultimate expression of data locality—training data never leaves the device. This architecture satisfies:

• GDPR data minimization requirements
• HIPAA technical safeguards for protected health information
• National data sovereignty laws requiring in-country processing

Only anonymized model updates, not raw patient data, are transmitted to aggregation servers.