telvis07 / kaggle-melbourne-university-seizure-prediction

Analysis for the Human EEG Kaggle competition

Kaggle Seizure Prediction Competition

author: Telvis Calhoun date: February 3, 2017 autosize: true

Kaggle Competition Goal

Detect seizure (preictal) or non-seizure (interical) segments of intracranial electroencephalography (iEEG) data. See Kaggle EEG Competition page for more details.

My Approach:

• Extract basic stats and FFT features for non-overlapping 30-second iEEG windows
• Detect signal drop out and impute missing data with mean for each feature per window
• Predict seizure and non-seizure segments using a stacked model.

Overview of Stacked Model

Model Features

Random Forest

• Min, Max, Mean, Median, Std Err, Absolute Sum per EEG channel per window
• Kurtosis, Skewness per EEG channel per EEG Channel per window
• Entropy of FFT magnitude and phases
• Entropy, Eignenvalues of EEG channels

---

GLM

• Number of windows per file predicted preictal by RF
• Number of windows per file predicted interictal by RF

Model Parameters

RandomForest

• Number of Trees (ntree) - 50
• Number of variables randomly sampled as candidates at each split (mtry) - 74
• Metric - Accuracy
• Number training folds - 5
• Use SMOTE due to preictal/interictal class imbalance

---

GLM

• Weight for Preictal samples: 1.0
• Weight for Interictal samples : 0.25
• Number training folds - 5

GLM Predictions

• Each point represents a 10-minute EEG segment file
• X-axis is the number of 30-second windows predicted as preictal by the RF
• Y-axis is the number of 30-second windows predicted as interictal by the RF
• The closer to an axis, the more certain the predicion

---

Final Thoughts

• This is my first Kaggle competition. I acheived my goal of making a competition submission. See my profile.
• I submitted after the deadline but my submission would have ranked 391 of 2440 submissions. Screenshot
• The Kaggle model stacking tutorial helped me understand cross-fold validation with stacked models.
• Deep's Kernel and Tony Reina's Kernel helped me understand EEG features.
• Special thanks to Hassan Kingravi for suggesting the stacked model.