How machine learning is detecting seizures in people with traumatic brain injuries

By December 7th, 2020

Traumatic brain injuries affect over 69 million people worldwide, with roughly 700,000 of those Australians. 30% of those Australians are likely to develop chronic epilepsy as the result of their injury, but a small and powerful piece of data-driven medtech could change that.  

By applying sensors, machine learning and energy harvesting technology, researchers at CSIRO, CSIRO’s Data61 and Australian medical device company Anatomics could prevent the development of seizure disorders in patients that have undergone decompressive surgery (craniectomy) through the creation of a data-driven and highly personalised treatment plan.

This research is split into three projects:
1. The development of implantable devices equipped with;
2. On-chip machine learning for seizure detection, followed by;
3. The creation of a smart helmet for brain activity monitoring.  

Initially funded by CSIRO’s Probing Bio Systems Future Science Platform, the research involves the design and placement of an array of autonomous electroencephalography (EEG) monitoring and seizure detection sensors to the injured part of the brain exposed during a craniectomy (the removal of part of the skull to prevent pressure to the brain caused by swelling as a result of injury).  

Done during the initial surgery to prevent further invasivenessthe devices are able to monitor the organ’s activity in realtime.   

“Brain activity data post-surgery is especially critical to a patient’s recovery as seizures regularly occur, with the frequency of these events often leading to many patients developing epilepsy,” explains project researcher and Data61 scientist Dr Umut Guvenc.   

“These seizures are often difficult to detect, with past EEG monitoring techniques only able to take place within a hospital using bulky devices and performed for up to 24 hours before being disconnected, providing a snapshot of brain activity during that time.  

This new method can continuously monitor brain activity wirelessly, allowing the patient to be mobile, comfortable and more socially active.”   

The implants are placed on a soft protective silicone shield (Durashield) fitted to the part of the brain where the skull has been removed, with the devices containing an integrated circuit (IC) capable of sensing, signal processing and machine learning. 

Trained using data from Monash University, the on-chip machine learning algorithm can detect even the smallest seizures, which are often so minute they can’t be felt by the patient 

Once detected, the data is communicated wirelessly to a small system that sits within the protective helmet patients wear before being transferred to a mobile and desktop interface to alert the practitioner.  

Machine learning for brain seizure detection

“The information will be used to inform clinicians about the brain activity of the patient and support them in decisions regarding the administering of antiepileptic drugs,” explains project researcher and senior Data61 engineer, Mr Peter Marendy. 

The system architecture of the algorithm is implemented on Field Programmable Gate Array circuits (FPGA) for testing, with the implantable microchips processing and communicating the live data.  

“A simple random forest algorithm was chosen for this application as it’s much less complicated than a CNN (convolutional neural network) in terms of hardware implementation,” says Dr Guvenc 

“The simplicity of this approach results in a considerable reduction in power consumption as well as reduced hardware footprint on the ASIC (Application Specific Integrated Circuit). As a result, a random forest algorithm is significantly more suitable for a hardware implementation with a little compromise on accuracy.” 

During normal brain activity, the implants stay in standby mode to conserve energy while monitoring brain activity for seizures. When a seizure is detected, everything is reactivated and the signal is sampled at higher resolution. 

While this project recently concluded, Anatomics and Data61 have received $1million in funding from the Government’s BioMedTech Horizons Program to develop a smart helmet to monitor brain swelling and other biometrics in stroke and traumatic brain injury patients that have undergone a craniectomy.  

Machine learning for brain seizures

The engineering of the machine learning system used to detect seizures.

Ultimately, the project will see the combination of advanced sensors and microelectronics with encrypted secure data transfer and machine learning to develop a ‘brain machine interface’, enabling neurosurgeons to monitor brain function in realtime. 

“What makes this work novel is that it will provide the data needed for a study on the ideal time when to perform a cranioplasty, research that will ultimately dictate future medical procedures,” says MMarendy.  

A key part of this project will determine the role of brain swelling in relation to the timing of the surgery says Mr Marendayexplaining that the combination of brain swell size data, surgery timing and patient outcome post-cranioplasty could determine a best practise guide for when surgery should be undertaken.  

“The eventual goal is to show there is a relationship between patient outcomes and the timing of skull reconstruction based on the size of the brain, with no research on this topic currently available. It will also be a precursor platform to the helmet for the brain implant as that work matures.

This project is now underway, with an update to come in 2021. 

CSIRO’s Data61 would like to thank the Probing Bio Systems Future Science Platform for their work and funding in the initial project, and CSIRO Business Units Energy and Mineral Resources for their work on developing micro batteries used in the initial and current project.  


Leave a Reply