Background and aim:
Chronic pain is a major healthcare problem that affects one of every five adults in Europe. Current diagnostic methods are suboptimal, and a variety of methods have been tried to help diagnostics. We aim to classify between different chronic pain models in large animals using a CNN-LSTM based on µECoG recordings from S1.

Methods and materials:
In this study, we employed 16 Danish Landrace pigs to investigate the effects of high frequency stimulation (HFS) in inducing a temporary chronic state and evoking long term potentiation (LTP). Additionally, some of the pigs underwent spared nerve injury (SNI) to induce long-term chronic pain. Electrocorticography (ECoG) recordings were conducted before and after intervention, with low frequency stimulation (LFS) applied. The recorded data was transformed using the continuous wavelet transform. This spectrogram was utilized as input for a CNN-LSTM neural network architecture, aiming to identify unique patterns in the signal that could be correlated with specific pain model groups (LTP, SNI, and control). The design of the model focused on differentiating between these groups.

Results:
An accuracy of 42.8% as achieved for the multiclass model, and 52% for the binary model. Test results showed only correct predictions for the control class for both models. The validation accuracy was 63.9% and 84.0% for the multiclass and binary model respectively. Model interpretability showed differences between the location and patterns of feature attribution from validation to test data.

Conclusions:
The performance of the developed CNN-LSTM model was found to be unsatisfactory for both multiclass and binary classification tasks. To enhance its generalizability, it is crucial to include a larger number of subjects, particularly from the LTP and SNI groups. By incorporating a wider range of inter-subject variance in the training process, it can potentially improve the model's performance and make it more reliable for future use.