Edge AI for medical devices — on-device ML for ECG, EEG, PPG, and biosignal pipelines

Why edge AI for medical devices is its own discipline

Running an ML model on a smartphone is straightforward. Running one on a battery-powered medical device that has 256 KB of RAM, a 64 MHz Cortex-M, and an FDA submission attached to it is not. Three constraints reshape every design choice:

• Latency. If the inference does not finish before the next ECG window arrives, the model is too big or the runtime is wrong.
• Power. A model that costs 30 mJ per inference and fires once a second turns a multi-day device into a multi-hour device.
• Validation. An FDA submission needs documented training data, version-locked models, regression results, and a way to prove the model in firmware behaves exactly like the model in your notebook.

HydraTech designs around all three. We have shipped on-device biosignal pipelines that meet all three constraints and feed cleanly into your QA/RA partner's submission package.

What edge AI engagements look like

Stack we ship

Layer	What we use
Silicon	Nordic nRF52 / nRF5340, STM32 with CMSIS-NN, ESP32-S3, NXP iMX RT, Linux SoCs (NXP iMX, Rockchip)
Runtimes	TensorFlow Lite for Microcontrollers, CMSIS-NN, Edge Impulse, ONNX Runtime, TFLite, PyTorch with quantization
Training stack	PyTorch, TensorFlow, scikit-learn — for the cases where we own the training pipeline too
Biosignals	ECG, EEG, PPG, SpO₂, IMU activity, audio (voice activity, cough detection)
Verification	Captured dataset replay, on-device unit tests, training/firmware bit-exactness checks, CI on hardware-in-the-loop

Proof — edge AI programs we have shipped

FAQ

Which on-device ML runtimes do you use?

TensorFlow Lite for Microcontrollers, CMSIS-NN, ONNX Runtime, and Edge Impulse for Cortex-M. For Linux SoCs we use TFLite, ONNX Runtime, and PyTorch with quantization. The runtime is chosen for the silicon, not the other way around.

Can edge AI in a medical device get past FDA?

Yes. The FDA's Software as a Medical Device (SaMD) and Predetermined Change Control Plan guidance both have established paths for AI / ML medical devices. We design the inference pipeline, training-data documentation, and verification suite around what your regulatory partner needs.

How do you validate biosignal ML on-device?

Captured biosignal regression suites that run on real hardware in CI. Every firmware build replays a curated dataset through the on-device pipeline and asserts on outputs so a regression in DSP, sensor driver, or model cannot silently degrade clinical accuracy.

Do you train the model too, or just deploy it?

Both. For algorithm-heavy programs we partner with your data science team and own deployment, quantization, and on-device verification. For smaller programs we have shipped end-to-end — feature engineering, training, quantization, on-device inference.