Prediction error from CNN: Giving same predictions for all test inputs

Question

I have been training the ICBHI 2017 dataset using CNN. This was done by converting the audio files (lung audio data) into STFT and MFCC. These features were fused (and of course padded to avoid any discrepancy) using [np.dstack] as suggested in this research paper.

The fused data was then fed into a CNN model as shown in the picture below CNN Model 1.

The output is supposed to be a class among 4,6,7 or 8.

The model predicted as follows on the given test data. CNN Model 1 Prediction

The model predicted all to be of one single class. Accuracy in this way is also high (around 89%, note that data for this particular dataset have 90% of audio in particualr class).

I used a much simpler to avoid overfitting CNN Model 2. Even this too showed similar behaviour.

As the paper suggest, I used depthwise separable CNN (MobileNet architecture) and it too showed a similar behaviour.

Kindly provide with good suggestions on whether I have to change the model (and how: reducing layers, increasing layers or changing layers?) or to process the data in any other way.