Table 3.
Results of Neural Network Analysis with Cambridge Neurological Test Automated Battery (CANTAB) Test Results as Input Data and Diagnosis, Healthy Control (HC) as Output Variables
| Models | NN#1 | NN#2 | |
|---|---|---|---|
| Input layer | Number of units | 13 | 13 |
| Hidden layers | Number of hidden layers | 2 | 2 |
| Activation function | Hyperbolic tangent | Hyperbolic tangent | |
| Number of units in hidden layer 1 | 3 | 3 | |
| Number of units in hidden layer 2 | 2 | 2 | |
| Output layer | Output variables | aMCI and HC | mCoDy and HC |
| Number of units | 2 | 2 | |
| Activation function | Softmax | Identity | |
| Training | Error term | 32.705 (cross entropy) | 11.392 (sum of squares) |
| % incorrect predictions | 32.1% | 28.6% | |
| ROC curve | Area under the ROC curve | 0.801 | 0.781 |
| Testing | Error term | 13.390 (cross entropy) | 4.242 (sum of squares) |
| % incorrect predictions | 25.0% | 30.4% | |
| Holdout | % incorrect predictions | 33.3% | 17.9% |
| Classification | Sensitivity=56.5%, specificity=81.3% | Sensitivity=80.0% specificity=83.3% |
Abbreviations: aMCI, amnestic mild cognitive impairment; HC, healthy controls; mCoDy, mild cognitive dysfunction (a purer subgroup of aMCI subjects); ROC, receiver operating characteristics.