Page d'accueil

Analysis of mouse sleep with machine learning

This project revolves around the analysis of EEG (electroencephalogram) and EMG (electromyogram) data from 250 mice, collected by UNIL researchers for a biological study on sleep. The data involves three sleep stages: REM (rapid eye movement), NREM (non-rapid eye movement), and Wake. However, the data is unbalanced, with REM representing only about 5% of the total data. Due to the laborious nature of manual sleep data annotation, machine learning techniques are explored to achieve more efficient data interpretation.

The data is transformed and subjected to feature engineering to extract relevant information. Three initial experiments using a random forest classifier achieved excellent results in classifying sleep stages for individual mice, entire breeds, and all mice from different breeds.

To address the issue of unbalanced labels, ensemble methods like the balanced random forest and easy ensemble classifiers were tested, with the balanced random forest showing the most promising results. A combined approach was adopted, where the first model classified wake versus sleep, and the second model classified REM versus NREM. This combined approach yielded performance similar to the balanced random forest method. Feature importance ranking and exploration were utilized to reduce the number of features and simplify the input, enhancing the model's accuracy.

Finally, a comparison was made between our random forest classifier and a Support Vector Machine-based annotation method developed by UNIL researchers, which employed one SVM model for each mouse. We computed the Cohen's Kappa coefficient for this comparison and found that our model performs very similarly to the ensemble of SVM models, with the advantage of being a single model.