1. Schizophrenia Classification Using
Artificial Neural Networks and Support Vector Machines
Kamini Jodha

2. The Motivation
Schizophrenia, a mental disorder, characterized by abnormal social behaviour and failure to identify what is real.
Affects 0.3%-0.7% people. Low? No. 24 million people around the world.
Reduces life expectancy by years than the general population.
Almost every patient, once in his/her lifetime attempts suicide.
Very curable if diagnosed early.
But due to its symptom overlap with other mental illnesses (like bipolar disorder) it can only be diagnosed subjectively, by process of elimination.

3. The Data Set Two modalities of MRI scans has been used:
Functional Network Connectivity (describe patterns of the brain function)
Source Based Morphometry (describe patterns of the brain structure)
Total FNC features – 378
Total SBM features – 32
Combine total number of features – 410
Total number of Samples – 86
*(It is a Kaggle competition dataset)

4. Pre-Processing Normalization Normalize value to between 0 – 1
Combine both feature sets
Correlation based feature selection

5. Artificial Neural Network Analysis
MATLAB program
Network Configurations:
1, 2 or 3 hidden layers
Each layer with 3, 5 or 7 perceptrons
70% training
15% for testing
15% for validation
Transfer functions used:
Logistic
Tan Sigmoid
Network retrained 50 times with random sample with replacement.

6. Support Vector Machine Analysis
MATLAB program implementation is done
Parameters Tuned:
Kernels
Radial Basis Function
Linear Kernel
C - support vector cost/penalty
Kernel Scale
- Cross validation
(* Results are yet to be tuned and finalised)

7. Artificial Neural Network Results
Maximum accuracy 97.4%
Sensitivity: 98%
Specificity: 96%
Support Vector Machine Results
Maximum accuracy: 94%
Sensitivity: 96.58%
Specificity: 91.43%
(*Configurations of the parameters are yet to be tuned and finalized)

8. Conclusion
In comparison to the SVM, the neural network achieved higher accuracy by 3%, sensitivity by 2% and specificity by 5% (*Further analysis has to be done)
For the result, as of now, a neural network appears to be a better algorithm for classifying the data.
*(Conclusion is yet to be finalized)

9. Discussion
Other configurations for the ANN and SVM which were not analyzed.
More fine tuning of parameters.
Artificial Neural Network
Learning rates, transfer functions, more/less layers/perceptrons
Support Vector Machines
More Kernels
Different cost function