result

The Stuttgart Neural Network Simulator (SNNS - free-available software simulator) has been used in our studies to build and train networks used in the experiments. We have implemented the "Quickprop" training algorithm with continuous sigmoid activation of neurons, because of the character of data. A classic four divided cross-validation method, multiple random initialisations and network trainings [3] have been used in order to check the correctness and repeatability of the results. The test's results of ten networks were averaged to get partial results.

These partial results, after ultimate averaging for four divisions, provided us with the percentage of correct detected persons (Fig. 2). In our neural network experiments we have used 19 parameters (Table 1, Set I) because some of 26 mentioned features were expressed both in minutes and in percents of total sleep time. There are also parameters, which can be calculated on the basis of the others from that set, for example "stage3 plus stage4". Therefore, we can simplify the analysed set to only 13 parameters (Table 1, Set II).

To evaluate an influence of particular parameters on result of alcoholic addiction detection, we checked how reducing of each parameter from training set (Table 1; Set II) changes the outcome. Fig. 2 shows the results of these experiments. We have also calculated percentage of effective networks' initialisations as a tentative method of estimation the parameter's significance.

The statistical analysis (Fig. 1) shows that some sleep parameters have small differences (100%) standard deviation in both groups. This means that detection of alcoholic addiction based on statistical methods would have too small correctness to be reliable.

The averaged results of neural network analyses are presented in Fig. 2. We have obtained the percentage of correct detections from 69,9% to 76,6% with the standard deviation of 0,83,8% for all experiments. Effective initialisations were from 64% to 97,5%. Based on the results shown in Fig. 2 we can notice, that lack of some parameters in the Set II leads to better results (e.g. "latency to stage 3 and 4"). Simultaneously, a deficiency of the other ones causes noticeable deterioration of the detections' outcomes (e.g. "stage 4 NREM" and "stage 3 NREM"). The significance weight of parameters: "stage 3 NREM" and "stage 4 NREM" was confirmed by the values of the received numbers of the effective initializations and by substantial differences in the averages for the groups. Moreover, the big difference in the averages of "latency to stage 3 and 4" seems to be not correlated with alcohol addiction.

The noticeable differences of the results for the specified sets of parameters (Set I and Set II) indicate that the redundant information decrease reliability of the detections and optimization of the set's composition is required. We have proposed the Set III (Table 1) to check if we can reduce a number of features from the Set II without worsening the correctness of diagnosing. We obtained 76,25% correctness of diagnoses with 51,6% effective initializations for the neuronal network with 5 hidden units.

Simultaneously, we achieved the 76,83% correctness of diagnoses with 96,4% effective initializations for the neuronal network with 7 hidden units. These results show that "weak features" are necessary to optimise the learning process' quality.