Machine Learning Applied on EEG Data

Machine Learning Applied on EEG Data to Identify Epileptic Seizure Results Discussion/Conclusion Figure 1: EEG Input Data Application: Based on our results we have concluded that a portable EEG reader with an ML chip could be beneficial to those living with epilepsy. Although our F1 score of .93 may not seem accurate enough for clinical work it actually outperforms real physicians, if action is reserved until a few consecutive samples come back positive, real accuracy will improve(4). Introduction Epilepsy: Epilepsy is a neurological disease which affects 3.4 million people in the United States. Individuals with epilepsy have an imbalance in the activity of inhibitory and excitatory neurons in the brain. This imbalance causes seizures that often lead to a loss of consciousness and involuntary convulsions (1). EEG: Electroencephalography (EEG) is a test that measures the electrical activity of the brain and it is often used to detect an epileptic seizure (1). Recognition: Visual recognition of a seizure from EEG data is not always obvious, thus a machine learning algorithm could more accurately detect the onset of Image credit: (10) a seizure (2). Fig. 1: Blue graphs show EEG for non‐epileptic subjects, Red for epileptic. Seizures can start halfway into sample, complicating analysis. Samples are taken from various electrodes and last for one second. Selected Sample Range=+‐500μV; Data Range = +‐2000μV. Notice the epilepsy does not need to start at the beginning of a sample. Fig. 2: Even though switching from SGD to MLP did not improve precision by that much (.93 to .95), it did improve the more important recall value from .72 to .91. ML: A machine learning (ML) algorithm can learn when a seizure is occurring through training on EEG data that corresponds to seizure and non seizure events. Methods SGD Classifier MLP Classifier Input Layers: EEG waves split into 179 points Seizure or Not? Point 2 Point 3 Final Point Fig. 3: The SGD Classifier algorithm obtained its highest F1 score with 450 iterations and 500 as the hidden layer size. The MLP Classifier algorithm obtained its highest F1 score with 400 iterations and 400 as the hidden layer size. 179 Adapted from: (9) Algorithms: The experiment used a machine learning algorithm trained with EEG data that exhibit either seizure or non‐seizure behavior. Two algorithms were explored: the Stochastic Gradient Descent (SGD) and multi‐layer perceptron (MLP) classifiers from the Sci‐Kit SKLearn Python package (3). EEG Cap: This measuring device is placed on a participant’s head and a conductive gel is placed on each electrode. Waves from the general location of the electrode are recorded. Run Time: When ran on the laboratory computers (3600 MHz) it took 40 seconds to train and a negligible time to test. On a computer with a smaller processor (Raspberry Pi clock speed=200MHz) the run time would be longer, but still small. Confusion Matrix: Because ML algorithms cannot be ‘opened’ and examined, a confusion matrix is a good metric to see what sorts of errors occur. The F1 score is based on this metric, but knowing the specifics allows us to tune our parameters. 2766 36 52 646 Adapted from: (8) Image credit: (6) F1: While a false‐positive (FP) is a hindrance to an epilepsy patients day to day life, a false negative (FN) that ignores an ongoing seizure could prove deadly. Because FN could only be reduced at great expense to FP, the more important value is lower here. Fig. 3 (Parameters Heatmap) Data: Our data is a set of EEG electrode readings (n=11501) each describing a one second sample of EEG record labeled 1 (epileptic), or (2‐5) non‐epileptic, that is represented by 178 integer summation. Start of EEG Optimizing: A balance of layer size and iterations is clearly essential to maximize accuracy. While increasing iterations seems to only improve results, it actually overtrains, making the tool less broadly applicable. With only one user, this may not be as much of a problem due to the narrow application on one person. A commercial product would be partially trained, and would tune itself to its user improving over time. Image credit:(5) Driving: It is legal for those with epilepsy to drive, but such individuals are a at greater risk for collisions due to the unpredictable nature of seizures. Self‐driving cars are not yet safe without driver supervision, but combined with an epilepsy detection device that triggered the car to pull over at seizure onset, the risk of collision could be greatly mitigated. Although this algorithm used one second samples‐‐too long to prevent accidents‐‐because it detected the seizure in samples with seizure duration of less than a second, the data supports the potential of such a product.