Early prediction and real-time response to life-threatening events is more important than ever. Cutting-edge research in the medical domain has shown that deep learning (DL)-based algorithms can perform advanced diagnoses that are hardly achievable by doctors. Yet, these sophisticated inference techniques are confined to server-scale platforms, and thus unable to process data at its source – the human body. Although recent research in implantable medical devices has made giant steps toward the Internet of Implantable Medical Things (IIoMT), it is still unknown:
(i) Whether DL techniques can be successfully integrated inside a resource-challenged embedded implantable system
(ii) Whether hardware-based DL can provide better energy and latency performance with respect to a CPU-based or cloud-based offloading of the learning task
Technology Overview
Embedded Networked Deep Learning (ENDL) is composed of:
(i) A hardware-based convolutional neural network (CNN) that interfaces with a series of implantable sensors
(ii) A wireless ultrasonic interface that sends the classification results to an external device or receives actuation commands
To study the necessary trade-offs between latency and resource consumption, Northeastern researchers have proposed a mathematical model of the interactions between the ENDL’s components. ENDL is prototyped on a system-on-chip platform and its end-to-end capabilities are demonstrated on an application to predict seizures in epileptic patients, where the platform trains the models using real intracranial electroencephalogram (iEEG) data.
- Boosting scheme improves validity of neural network predictions 
- Improved latency by nine times as compared to full CPU base system 
- Four and a half times less energy consumption than a cloud-based neural network
- Capabilities to communicate predictions with the outside world
- Predict the onset of epileptic seizures well before the occurrence
- Perform actuation to mitigate effects of seizures 
- Notify healthcare personnel on patient condition
- Predict other negative health events from different sensor inputs
- License
- Research collaboration
- Partnering
Patent Information:
-Sensors tech
For Information, Contact:
Mark Saulich
Associate Director of Commercialization
Northeastern University
Daniel Uvaydov
Raffaele Guida
Francesco Restuccia
Tommaso Melodia
Artificial intelligence
Biomedical implants
Deep learning