The library performs multitask-multiple kernel learning (mt-mkl) for time series analysis. We focus on the analysis of time series acquired on focal epileptic patients. In this case the source of the seizure is limited to a specific area of the brain. For drug resistant patients, the ablation is sometimes a solution. To localize at best the epileptic area, clinicians perform invasive measures, through deep filiform electrodes, which are implanted under skull. Each electrode on average has 10 channels of acquisition, and a correspondent time series. Given the tag assignment (binary label that denotes if the channel has epileptic activity) performed by medical experts, which we use as ground truth, mt-mkl addresses at the same time two problems (i) classification of the activity and the (ii) extraction of relevant features in the frequency domain.
The library consists of
preprocessing step: we filter each time series to remove powerline effects and then we proceed to the computation of wavelet transform - NEW for a total of 83 scales kernel computation: we use different similarity measures. At the moment they are correlation, phase locking value, NEW cross correlation (not the average value) multikernel: this part of the library includes minimization methods scripts: the first is test_wavelet_trasform.py for the computation of the wavelet transform and the kernels learning_pipeline.py learning method through random search cv