This summer, I had the wonderful opportunity to work in the Penn Center for Neuroengineering and Therapeutics (CNT) to analyze medical signal recordings with machine learning. A brief description of my project is provided below.
Forty-thousand neonates are born annually in the U.S. with congenital heart diseases (CHD), and 1% of these patients experience cardiac arrest. With seizures occurring in many neonates after cardiac surgery, clinicians have monitored CHD patients’ EEG recordings and observed that EEG background patterns change preceding cardiac arrest. Although EEG recordings of CHD patients have been manually reviewed by clinicians, this method has demonstrated the inevitable drawbacks of high cost and subjectivity. To address this problem, I started a project with clinicians at the Children’s Hospital of Philadelphia (CHOP) to apply computational methods for extracting quantitative EEG features and classifying neonatal EEG backgrounds.
The data for my project were obtained from EEG recordings of neonates in Cardiac Intensive Care Units at CHOP. Each patient’s EEG was recorded through 27 scalp electrodes, along with clinician-verified annotations. Upon loading the dataset, I devised a complex pipeline to remove faulty recordings and noise from the data and calculate quantitative EEG features for windowed intervals of the EEG recordings. These features were labeled along with specific annotations, and I trained a popular machine learning algorithm called an Artificial Neural Network (ANN) to predict EEG backgrounds of arbitrary neonate EEG inputs of fixed length. Testing the model over random EEG samples across patients returned an accuracy of 86%, while the results for cross-patient validations were more varied.
Before this program, I had no prior experience with working in an actual lab. One of the things that I most appreciated in the CNT was how the professors and graduate/postdoc students willingly allocated their precious time to share their knowledge and help me throughout my research process. The atmosphere in the lab was very friendly and open-minded to everyone, which allowed me to freely ask questions and improve my research/communication skills overall. On a more technical side, I had the unique opportunity to work with mass recordings of biological waveform data using machine learning. This led me to study other related areas such as digital signal processing and learn useful techniques for handling such large datasets. Overall, I consider my experience at the lab this summer to be very rewarding and meaningful. As a closing remark, I would like to thank Dr. Litt for providing this enriching research opportunity, as well as my PhD advisor John Bernabei for guiding me throughout the project.
This summer, I had the wonderful opportunity to work in the Penn Center for Neuroengineering and Therapeutics (CNT) to analyze medical signal recordings with machine learning. A brief description of my project is provided below.
Forty-thousand neonates are born annually in the U.S. with congenital heart diseases (CHD), and 1% of these patients experience cardiac arrest. With seizures occurring in many neonates after cardiac surgery, clinicians have monitored CHD patients’ EEG recordings and observed that EEG background patterns change preceding cardiac arrest. Although EEG recordings of CHD patients have been manually reviewed by clinicians, this method has demonstrated the inevitable drawbacks of high cost and subjectivity. To address this problem, I started a project with clinicians at the Children’s Hospital of Philadelphia (CHOP) to apply computational methods for extracting quantitative EEG features and classifying neonatal EEG backgrounds.
The data for my project were obtained from EEG recordings of neonates in Cardiac Intensive Care Units at CHOP. Each patient’s EEG was recorded through 27 scalp electrodes, along with clinician-verified annotations. Upon loading the dataset, I devised a complex pipeline to remove faulty recordings and noise from the data and calculate quantitative EEG features for windowed intervals of the EEG recordings. These features were labeled along with specific annotations, and I trained a popular machine learning algorithm called an Artificial Neural Network (ANN) to predict EEG backgrounds of arbitrary neonate EEG inputs of fixed length. Testing the model over random EEG samples across patients returned an accuracy of 86%, while the results for cross-patient validations were more varied.
Before this program, I had no prior experience with working in an actual lab. One of the things that I most appreciated in the CNT was how the professors and graduate/postdoc students willingly allocated their precious time to share their knowledge and help me throughout my research process. The atmosphere in the lab was very friendly and open-minded to everyone, which allowed me to freely ask questions and improve my research/communication skills overall. On a more technical side, I had the unique opportunity to work with mass recordings of biological waveform data using machine learning. This led me to study other related areas such as digital signal processing and learn useful techniques for handling such large datasets. Overall, I consider my experience at the lab this summer to be very rewarding and meaningful. As a closing remark, I would like to thank Dr. Litt for providing this enriching research opportunity, as well as my PhD advisor John Bernabei for guiding me throughout the project.