Obviously, one of two conditions applies:

Remark: it is not that tragical to have a system the gives an output "might be one of the following (short list)" if it is made clear how to proceed and discriminate between the "options" by other means.

And an afterthought: are you already using neuronal networks?

I think that the placement of the the electrodes on the head surface. Specific waveforms can be accessed in specific locations on the scalp. You may need to consult a neurologist specialized in brain activity.

I think the paper in the given link will be useful to you:http://www.measurement.sk/2002/S2/Teplan.pdf

Best wishes

Thank you so much for the answer. I have used neural networks but earlier I have used it to detect only one disease. I am also trying to show a way to discriminate between the options as you said.But how can I improve my feature extraction?

If you are already using NN-based feature extraction, more "learning" with "typical" samples for each disorder might do the 'trick'. As the esteemed Prof. Zekry added, checking and eventually modifying the electrode placing might add to get more discriminative signal patterns.

Depending on which NN topology you are currently using, a change of topology might also add to the discriminative properties.

Wish you success.

Although probably not what you looking for, simple physiological variables might add clarity such as age, EMG, HRV, BMI, reaction time, ...  I think the approach of looking for a number of diseases and trying to access their probabilities and distinguish from healthy controls may prove fruitful. Also the diseases that you mention have co-morbidities.  The  disease descriptions are inherently symptomatic and not based upon brain/body structure (just because we do not have the diagnostic ability to do so).  The results that you are getting may be accurate.  It is just that the diseases overlap.  Ultimately, this is a combine sensor and big data problem.