Neurostimulation is becoming a more popular treatment approach in individuals with drug-resistant epilepsy. One neurostimulation option is the responsive neurostimulation (RNS), a closed loop interface that monitors the electrical brain activity and applies local electrical current when seizure-like patterns are detected. Changes at the time-frequency domain of the ictal putative signal were described by visual inspection of expert epileptologists. Some of these electrographic seizure pattern modulation (ESPM) showed to be correlated with clinical improvements. Due to the large amount of data and scarce of experts' time, there is a need to develop unsupervised methods to identify ESPM with high precision. For this, we evaluate the capability of one class support vector machine (OCSVM) classifiers ensemble to detect ESPM. We tune the OCSVM hyperparameter and accelerate the training process using a kNN and data density based method. As data input we use different signal representation approaches such as short-time Fourier transform spectrogram, a widely adopted electrophysiology time-frequency representation, and scattering transform, a non-linear signal representation method that uses cascading wavelet modulus decomposition followed by a low pass filter. Finally we compare the classification metrics of the unsupervised approach with few expert labeled signals and generalize the result to signals not labeled by the expert.