1. Seizure Detection Algorithm in Neonates Using Limited Channel aEEG
Team Members: Tahj Alli-Balogun, Kyle Thomas
Client: Dr. Zachary Vesoulis

2. Background | Project Scope | Design Specs | Existing Solutions | Schedule
Neonatal Intensive Care Unit (NICU)
Term Births – Birth around 40 weeks
Pre-term – Birth before 37 weeks
Apnea
Respiratory Distress Syndrome
Hypoxic-Ischemic Encephalopathy
Seizures

3. Electroencephalogram (EEG)
Background | Project Scope | Design Specs | Existing Solutions | Schedule
Electroencephalogram (EEG)
Conventional EEG (cEEG)
Electrodes
1-2 hour use
High Resolution
Needs a trained neurologist to interpret
Amplitude-Integrated EEG (aEEG)
Usually 4 electrodes
12-24 hour use
Simple interpretation training
Quick Set-Up

4. Need Statement and Project Scope
Background | Project Scope | Design Specs | Existing Solutions | Schedule
Need Statement and Project Scope
There is a need for an accessible monitoring system to alert clinicians in the neonatal intensive care unit (NICU) about the health status of seizure-prone neonates for potential life-saving intervention.
We propose to deliver a system that will employ a limited channel-EEG to provide a less expensive alternate to current, commercialized software. A device that fulfills this need will include different responses based on the severity of the situation to guide the clinician towards an apt response. In addition, it will have an accessible device interface and be safe for use with infants. We propose to deliver, by the end of April 2019, a system that can both reliably detect seizures in neonates utilizing data from a limited-channel EEG and alert clinicians of an ongoing seizure with distinct and non-chaotic signals appropriate for the NICU. The prototype will include a user manual for detailed instructions on the reproduction and operation of the system.
There is a need for an accessible monitoring system to alert clinicians in the neonatal intensive care unit (NICU) about the health status of seizure-prone neonates for potential life-saving intervention.

5. Design Specifications
Background | Project Scope | Design Specs | Existing Solutions | Schedule
Design Specifications
Specification
Metric
Description
Sensitivity
99%
Seizures detected divided by total number of seizures
Specificity
80%
False seizures detected divided by total number of seizures
Frequency Sampling
> 240 Hz
The frequency rate at which signals from the neonate's brain are sampled
Device Responsivity
< 10 seconds
Delay between initiation of a seizure and device alert
Operating Time
24 hours
This device will operate over multiple neonatal sleep-wake cycles
Device Setup Time
< 10 minutes
Non-expert clinicians should be able to set up system with ease
Production Cost
$200 (USD)
Cost-effective device based on integration with NICU equipment
Device Size
9 cm x 6.5 cm x 2.5 cm
The device will be easily storable and portable for rapid, efficient use

6. Background | Project Scope | Design Specs | Existing Solutions | Schedule

7. cEEG with Physician Interpretation
Background | Project Scope | Design Specs | Existing Solutions | Schedule
cEEG with Physician Interpretation
BioSemi ActiveTwo System
Gold-standard for brain monitoring
Incredible resolution which scales with electrode number
Drawbacks
Requires trained neurologist interpretation
Real-time analysis difficult
Electrodes can damage baby’s skin

8. aEEG with Detection Software
Background | Project Scope | Design Specs | Existing Solutions | Schedule
aEEG with Detection Software
CFM Olympic BrainZ Monitor with RecogniZe software
Electrodes record a limited channel EEG
Display shows raw EEG and aEEG
RecogniZe software
Drawbacks
Expensive (~ $25k-$30k)
Does not integrate well with other signals

9. Other Clinical Signals
Background | Project Scope | Design Specs | Existing Solutions | Schedule
Other Clinical Signals
Angelcare AC517 Baby Breathing Monitor with Video
Detects motion with a sensor pad
Video camera for real-time monitoring
Drawbacks
Does not work well with pre-term babies
Requires constant monitoring

10. Background | Project Scope | Design Specs | Existing Solutions | Schedule
Machine Learning
Classical Machine Learning
Random Forest
Summation of Decision Trees
Mursalin et al.: 97% Specificity Rate
Support Vector Machines
Hyperplane between datasets
Kumar et al.: 95% Specificity Rate

11. Background | Project Scope | Design Specs | Existing Solutions | Schedule
Machine Learning
Classical Machine Learning
Random Forest
Summation of Decision Trees
Mursalin et al.: 97% Specificity Rate
Support Vector Machines
Hyperplane between datasets
Kumar et al.: 95% Specificity Rate

12. Background | Project Scope | Design Specs | Existing Solutions | Schedule
Design Schedule

13. Group Responsibilities
Background | Project Scope | Design Specs | Existing Solutions | Schedule
Group Responsibilities
Responsibilities
Kyle
Tahj
Report Writing X
Point Person for Client Contact
Web Page Upkeep
Literature Search and Clinical Relevance
Lead Software Designer
Signaling and Systems Analyst
Electronics and Machine Design
Cataloguing of Results
Preliminary Presentation
Progress Presentation
V&V Presentation
Final Presentation and Poster

14. References https://i.ytimg.com/vi/TjaXUyQyH34/maxresdefault.jpg
representation-of-the-electrodes-positions-in-the-a-19-Electrodes-b.png
Mursalin, M., Zhang, Y., Chen, Y., Chawla, N. (2017). Automated Epileptic Seizure Detection Using Improved Correlation-Based Feature Selection with Random Forest Classifier. Neurocomputing, 241,
Kumar, Y., Dewal, M. L., Anand, R. S., Epiletic Seizure Detection Using DWT Based Fuzzy Approximate Entropy and Support Vector Machine, Neurocomputing, 133,

15. Thank You!

16. Background | Project Scope | Design Specs | Existing Solutions | Schedule

17. Background | Project Scope | Design Specs | Existing Solutions | Schedule
Machine Learning
Deep Machine Learning
Convolutional Neural Networks (CNNs)
More effective than classical machine learning in image recognition tasks
Require a large dataset