Neuropsychological impairment in depression is less concise compared to schizophrenia, dementia or other brain disorders. It is varying between patients and over time in the natural course of depression. Furthermore it depends on various co-variables. These characteristics make the detection of depressive patterns in neuropsychological performance very difficult for conventional statistics.
Artificial neural networks are highly parallel nonlinear teachable systems of information processing. They are used for pattern recognition and classification tasks in different fields and can be superior to conventional linear statistics in the analysis of complex data.
The results of 1100 neuropsychological examinations of psychiatric patients with varying diagnoses and healthy controls were used to train different kinds of neural networks. The neuropsychological battery (NEUROBAT) consists of usual test paradigms as optical reaction time, a go-nogo task, recognition and free memory recall, sensorimotor interference and a continuous performance task.
Trained multilayer perceptrons and radial basis function networks allowed a significant recognition of depressive patterns. Patients were classified correctly in up to 71% of cases, whereas up to 64% of depressive disorders were recognized correctly by linear artificial neural networks.
Recognition of depressive neuropsychological patterns seems to be possible by artificial neural networks. But sensitivity and specificity are too low for a possible support of clinical diagnostics. The superiority to linear classification models could not shown clearly. More complex hierarchical neural networks, as they are commonly used in picture recognition, should be tested in future studies in order to improve classification results.