**3.1. Descriptive statistics and basic comparisons**

The recruited patient cohort consisted of 54 patients (37 males and 17 females), mean age of 62 years (SD 9.6). The neurological deficit by the Rankin's scale was assess at mRs=1 (37 patients), mRs=2 (16 patients) and mRs=3 (1 patient). Most frequent minor ischaemic strokes occurred in the left carotid system (53.7%), right carotid system (31.4%), both systems (9.3%) as well as in the vertebrobasilar system (5.6%). Most frequently observed are the syndromes of the middle cerebral artery (hemiparesis and involvement of VII and XII cranial nerves), aphasias after a damage of the dominant hemisphere, upper monoparesis and motor aphasia, etc. In particular, during the mean follow-up of 11.1±2.4 months, 8 secondary CV events (14.8%) were observed only in males within a mean period of 5.8±2.7 months. No difference in the age of patients with CV event (61.1±12.6 years) vs. those without (62.1±9.6 years) was found (p>0.05). The one-year risk for CVE was ≈15% (95%CI 7.1÷27.7%). The other main demographic and clinical parameters of the initial cohort of 54 patients were reported in a more detail earlier [Atanassova, Chalakova & Dimitrov, 2008a].

The main results are summarized in Table 1. Although the distributions of CMCT at month 12 in the asymptomatic hemisphere and MEP at month 1 in the symptomatic one were slightly skewed, two tendencies could be clearly observed. While there is clearly a difference in the TMS measures according to the existing symptomatics (i.e., symptomatic or asymptomatic hemisphere), the first one is an increase of CMCT over time with higher values in the symptomatic hemisphere, while the second one is again an increase of MEP over time, but the higher values this time are observed in the asymptomatic hemisphere (p<0.05).

An interesting pattern is, however, that while CMCT first decreased from month 1 to month 3 and then increased (Figure 2), the MEP amplitude, in parallel but opposite, increased in month 3 and then decreased slightly in month 12 (Figure 3). In particular, there was a significant change over time (p<0.001) in CMCT and a multivariate, combined effect of symptomatics and time (grand mean 8.523 ms, 95%CI 8.240-8.805, p=0.01). Notably, there was a statistically significant difference (adjusted for the baseline values at month 1) between the estimated marginal means of CMCT in the symptomatic (10.717 ms, 95%CI 10.191-11.244) and asymptomatic (8.023 ms, 95%CI 8.023-9.077) hemispheres (Figure 2).

Computational Intelligence in Electromyography Analysis – 42 A Perspective on Current Applications and Future Challenges

probabilistic modelling (EasyNN ver.6.0i) were used.

**3.1. Descriptive statistics and basic comparisons** 

**3. Results** 

& Dimitrov, 2008a].

(p<0.05).

ANOVA (general linear models) in the view of the symptomatic and asymptomatic hemispheres (**Figure 2 & Figure 3**). As appropriate, parametric and non-parametric

Parametric regression modelling was used to analyse the data and develop models to predict the outcomes at month 12 (**Table 2**). The significant relationships were later explored and confirmed by probabilistic artificial neural network (ANN) modelling, irrespectively of usual statistical constraints (**Figures 4 & Figure 5**). The stopping rule of learning was assumed when a state of maximum overall correctness of prediction with minimum average learning error was reached [Sarle, 1997]. The p-values less than 0.05 were considered statistically significant. The specialised software packages for statistical (SPSS ver.18) and

The recruited patient cohort consisted of 54 patients (37 males and 17 females), mean age of 62 years (SD 9.6). The neurological deficit by the Rankin's scale was assess at mRs=1 (37 patients), mRs=2 (16 patients) and mRs=3 (1 patient). Most frequent minor ischaemic strokes occurred in the left carotid system (53.7%), right carotid system (31.4%), both systems (9.3%) as well as in the vertebrobasilar system (5.6%). Most frequently observed are the syndromes of the middle cerebral artery (hemiparesis and involvement of VII and XII cranial nerves), aphasias after a damage of the dominant hemisphere, upper monoparesis and motor aphasia, etc. In particular, during the mean follow-up of 11.1±2.4 months, 8 secondary CV events (14.8%) were observed only in males within a mean period of 5.8±2.7 months. No difference in the age of patients with CV event (61.1±12.6 years) vs. those without (62.1±9.6 years) was found (p>0.05). The one-year risk for CVE was ≈15% (95%CI 7.1÷27.7%). The other main demographic and clinical parameters of the initial cohort of 54 patients were reported in a more detail earlier [Atanassova, Chalakova

The main results are summarized in Table 1. Although the distributions of CMCT at month 12 in the asymptomatic hemisphere and MEP at month 1 in the symptomatic one were slightly skewed, two tendencies could be clearly observed. While there is clearly a difference in the TMS measures according to the existing symptomatics (i.e., symptomatic or asymptomatic hemisphere), the first one is an increase of CMCT over time with higher values in the symptomatic hemisphere, while the second one is again an increase of MEP over time, but the higher values this time are observed in the asymptomatic hemisphere

An interesting pattern is, however, that while CMCT first decreased from month 1 to month 3 and then increased (Figure 2), the MEP amplitude, in parallel but opposite, increased in month 3 and then decreased slightly in month 12 (Figure 3). In particular, there was a significant change over time (p<0.001) in CMCT and a multivariate, combined

correlations between CMCT and MEP at various times were also performed.

There was a significant increase over time (p<0.001) in MEP amplitude, however, the multivariate, combined effect of symptomatics and time was not significant (grand mean 8.310 mV, 95%CI 7.922-8.697, p=0.309). Certainly, there was a statistically significant difference between the estimated marginal means of MEP amplitude in the symptomatic (6.888 mV, 95%CI 6.340-7.437) and asymptomatic (9.731 mV, 95%CI 9.182-10.279) hemispheres, but this was observed since month 1 and continued as such till month 12 (Figure 3).

**Figure 2.** General linear modelling (repeated ANOVA) of CMCT changes from month 1 till month 12

Modelling of Transcranial Magnetic Stimulation in

coefficient β\* p-value

Symptomatics -0.354 <0.001

month 1 0.642 <0.001

Symptomatics 0.183 0.051

One-Year Follow-Up Study of Patients with Minor Ischaemic Stroke 45

Since these changes appeared to be parallel, we tested also the correlations between the measurements of TMS parameters at different months. In particular, there was a weak inverse, but significant correlations between CMCT and MEP (Spearman's Rho=-0.45-0.46, p<0.05). Notably, the highest positive correlations were observed between CMCT at month 1 and the following months (0.60-0.81, p<0.05) as well as between MEP at month 1 and the following months (0.78-0.87, p<0.05). The latter relationships provided the opportunity to

The parametric regression modelling indicated that the CMCT outcome at month 12 can be predicted by the initial values at month 1 and whether or not these have been observed in the symptomatic or asymptomatic hemisphere (Fmodel=33.323, p<0.001, Table 2). The same is valid for the MEP amplitude outcome at month 12 (Fmodel=55.0.09, p<0.001, Table 2), although the role of the symptomatic as a predictor is with a marginal statistical significance

month 12 CMCT at month 1 0.448 <0.001

Notes: \*The predictor "symptomatics" is a categorical variable referring to the particular hemisphere, with two categories: symptomatic and asymptomatic. The constants, unstandardized coefficients β and their standards errors are available from the authors upon request. Abbreviations: TMS, transcranial magnetic stimulation; CMCT, central

**Table 2.** Parametric regression modelling to predict the TMS outcomes in 40 patients at month 12

when the average error decreased below the target value of 0.049.

The above relationships were further investigated by a probabilistic modelling, employing artificial neural network (ANN) methodology, which has not the usual constrains of a parametric regression analyses (Figure 4 & Figure 5). The ANN for modelling and predicting CMCT at month 12 contained 9 nodes with 2 hidden layers, with two potential predictors: CMCT at month 1 and symptomatics (symptomatic or asymptomatic hemisphere) (Fig. 4). The structure for predicting the resulting outcome node was obtained

Amplitude of MEP at

model and predict the outcome at month 12 in the two TMS parameters.

(p=0.051).

Outcome parameters in 40 MIS

Central motor conduction time at

Amplitude of motor evoked potential at month 12

motor conduction time; MEP, motor evoked potential.

**3.2. Statistical and probabilistic modelling of the outcome at 12 months** 

patients Independent variables Standardized

**Figure 3.** General linear modelling (repeated ANOVA) of MEP amplitude changes from month 1 till month 12

Since these changes appeared to be parallel, we tested also the correlations between the measurements of TMS parameters at different months. In particular, there was a weak inverse, but significant correlations between CMCT and MEP (Spearman's Rho=-0.45-0.46, p<0.05). Notably, the highest positive correlations were observed between CMCT at month 1 and the following months (0.60-0.81, p<0.05) as well as between MEP at month 1 and the following months (0.78-0.87, p<0.05). The latter relationships provided the opportunity to model and predict the outcome at month 12 in the two TMS parameters.

### **3.2. Statistical and probabilistic modelling of the outcome at 12 months**

Computational Intelligence in Electromyography Analysis – 44 A Perspective on Current Applications and Future Challenges

**Figure 3.** General linear modelling (repeated ANOVA) of MEP amplitude changes from month 1 till

month 12

The parametric regression modelling indicated that the CMCT outcome at month 12 can be predicted by the initial values at month 1 and whether or not these have been observed in the symptomatic or asymptomatic hemisphere (Fmodel=33.323, p<0.001, Table 2). The same is valid for the MEP amplitude outcome at month 12 (Fmodel=55.0.09, p<0.001, Table 2), although the role of the symptomatic as a predictor is with a marginal statistical significance (p=0.051).


Notes: \*The predictor "symptomatics" is a categorical variable referring to the particular hemisphere, with two categories: symptomatic and asymptomatic. The constants, unstandardized coefficients β and their standards errors are available from the authors upon request. Abbreviations: TMS, transcranial magnetic stimulation; CMCT, central motor conduction time; MEP, motor evoked potential.

**Table 2.** Parametric regression modelling to predict the TMS outcomes in 40 patients at month 12

The above relationships were further investigated by a probabilistic modelling, employing artificial neural network (ANN) methodology, which has not the usual constrains of a parametric regression analyses (Figure 4 & Figure 5). The ANN for modelling and predicting CMCT at month 12 contained 9 nodes with 2 hidden layers, with two potential predictors: CMCT at month 1 and symptomatics (symptomatic or asymptomatic hemisphere) (Fig. 4). The structure for predicting the resulting outcome node was obtained when the average error decreased below the target value of 0.049.

Modelling of Transcranial Magnetic Stimulation in

One-Year Follow-Up Study of Patients with Minor Ischaemic Stroke 47

Legend: Yellow circles (No.0–1) on the left indicate 2 input variables. The magenta circle (No.16) on the right is the output variable (outcome). The nodes of two hidden layers are grouped vertically and coloured in cyan: hidden layer 1 (nodes No.2–9); hidden layer 2 (nodes No.10–15). ANN nodes description: Each node contains small bar charts indicating the basic functional parameters – net input (cyan bar), activation (magenta bar), bias (orange bar) and error (yellow bar). The hidden nodes are connected by lines, showing the type and strength of weights: the red and green

**Figure 5.** Artificial neural network trained with 40 patients to predict the MEP amplitude outcome at

lines indicate negative and positive weights, respectively. The thicker the line is, the heavier the weight.

month 12

Legend: Yellow circles (No.0–1) on the left indicate 2 input variables. The magenta circle (No.9) on the right is the output variable (outcome). The nodes of two hidden layers are grouped vertically and coloured in cyan: hidden layer 1 (nodes No.2–5); hidden layer 2 (nodes No.6–8). ANN nodes description: Each node contains small bar charts indicating the basic functional parameters – net input (cyan bar), activation (magenta bar), bias (orange bar) and error (yellow bar). The hidden nodes are connected by lines, showing the type and strength of weights: the red and green lines indicate negative and positive weights, respectively. The thicker the line is, the heavier the weight.

**Figure 4.** Artificial neural network trained with 40 patients to predict the CMCT outcome at month 12

The modelling confirmed the finding from the parametric regression analysis (β=0.448, Table 2) for a slightly higher relative (predictive) importance of the CMCT at month 1 (0.301) than symptomatics (β=-0.354, relative importance = 0.290, not shown).

The ANN for modelling and predicting MEP at month 12 contained 16 nodes with 2 hidden layers, with two potential predictors: MEP at month 1 and symptomatics (symptomatic or asymptomatic hemisphere) (Fig. 5). The structure for predicting the resulting outcome node was obtained when the average error decrease below the target value of 0.049. The modelling confirmed the finding from the parametric regression analysis (β=0.642, Table 2) for a quite higher relative (predictive) importance of the MEP at month 1 (relative importance = 230.19) than symptomatics (β=0.183, relative importance = 62.06, not shown).

Computational Intelligence in Electromyography Analysis – 46 A Perspective on Current Applications and Future Challenges

Legend: Yellow circles (No.0–1) on the left indicate 2 input variables. The magenta circle (No.9) on the right is the output variable (outcome). The nodes of two hidden layers are grouped vertically and coloured in cyan: hidden layer 1 (nodes No.2–5); hidden layer 2 (nodes No.6–8). ANN nodes description: Each node contains small bar charts indicating the basic functional parameters – net input (cyan bar), activation (magenta bar), bias (orange bar) and error (yellow bar). The hidden nodes are connected by lines, showing the type and strength of weights: the red and green lines

**Figure 4.** Artificial neural network trained with 40 patients to predict the CMCT outcome at month 12

The modelling confirmed the finding from the parametric regression analysis (β=0.448, Table 2) for a slightly higher relative (predictive) importance of the CMCT at month 1 (0.301)

The ANN for modelling and predicting MEP at month 12 contained 16 nodes with 2 hidden layers, with two potential predictors: MEP at month 1 and symptomatics (symptomatic or asymptomatic hemisphere) (Fig. 5). The structure for predicting the resulting outcome node was obtained when the average error decrease below the target value of 0.049. The modelling confirmed the finding from the parametric regression analysis (β=0.642, Table 2) for a quite higher relative (predictive) importance of the MEP at month 1 (relative importance = 230.19) than symptomatics (β=0.183, relative importance

indicate negative and positive weights, respectively. The thicker the line is, the heavier the weight.

than symptomatics (β=-0.354, relative importance = 0.290, not shown).

= 62.06, not shown).

Legend: Yellow circles (No.0–1) on the left indicate 2 input variables. The magenta circle (No.16) on the right is the output variable (outcome). The nodes of two hidden layers are grouped vertically and coloured in cyan: hidden layer 1 (nodes No.2–9); hidden layer 2 (nodes No.10–15). ANN nodes description: Each node contains small bar charts indicating the basic functional parameters – net input (cyan bar), activation (magenta bar), bias (orange bar) and error (yellow bar). The hidden nodes are connected by lines, showing the type and strength of weights: the red and green lines indicate negative and positive weights, respectively. The thicker the line is, the heavier the weight.

**Figure 5.** Artificial neural network trained with 40 patients to predict the MEP amplitude outcome at month 12

Computational Intelligence in Electromyography Analysis – 48 A Perspective on Current Applications and Future Challenges
