The medical records of the patient as well as other characteristics are required for the clinical diagnosis of silicosis. In order to arrive at an accurate diagnosis, it is necessary to consider not only the patient's medical history but also a chest X-ray that has been performed in accordance with the ILO standard. In this particular investigation, it was determined based on the patient history and many other diagnostic characteristics utilizing ANFIS.
The computer program was executed in a MATLAB environment (version R2013a, The MathWorks Inc., United States of America), and the fuzzy logic toolbox was utilized in its execution. ANFIS topologies including four input membership functions were learned throughout this process. For the ANFIS model, "working time," "burden of smoke," "profusion score," and "symptom" were all acceptable input membership functions.
The data set is divided in to two distinct sections. The first section of the data, which consists of 375 pairs of observations, is utilized in the process of model training. The model is tested using the second collection of data, which consists of 185 different pairs of observations.
The hybrid learning algorithm for ANFIS has been used in this study. The average error value of training models for four inputs is given in Table 2.
Table 2
Average Error of Testing and Training Model
Used Membership Function (For Each of the Inputs) | RMSE (training) | RMSE (testing) |
trimf | 0,0482 | 0,0540 |
trapmf | 0,6667 | 0,0999 |
gbellmf | 0,05091 | 0,0780 |
gaussmf | 0,048393 | 0,070493 |
gauss2mf | 0,05304 | 0,36367 |
When the test results are compared, it is seen that the model with the smallest error gives the best estimate (trimf RMSE = 0,0640). The model is estimated by trimf.
Model validation was presented some statistical methods such as mean squared error (MSE), root mean squared, root mean squared error (RMSE), Mean Absolute Error (MAE), the coefficient of multiple determinations (R2) might be used to present model validation. It is expected RMSE, MSE and MAE have lowest value and R2 has highest value for the best model validation.
Table 3
Statistics for Model Validation
Statistics for model validation | ANFIS MODEL |
R2 | 0,939180927 |
MSE | 0,002915715 |
RMSE | 0,053997365 |
MAE | 0,000938536 |
ANFIS models have good statistics for model validation. R2 is almost 0,94 and this value well enough for accuracy. It has been demonstrated that the ANFIS model developed by Elif et al to detect erythematous-squamous disease correctly accuracy is 0,95 (Übeyli & Güler, 2005) In another study conducted by Kianaz Rezaaei et al to detect peptic ulcer, the disease was estimated with an accuracy of 0.98 (Rezaei, Hosseini, & Mazinani, 2014) It has been also demonstrated that the ANFIS model developed for the diagnosis of heart valve disease by Avcı and Turkoglu. can detect the disease with 0,93 accuracy (Avci & Turkoglu, 2009). In this specific study, a model accuracy rate of 0.92 was found when comparing the projected risk score to the actual risk score. It has been determined that the level of accuracy achieved by the findings of this investigation is satisfactory.
In this model, MSE is 0,0029, RMSE is 0,0539, MAE is 0,0009 and have low value as seen. It is desirable to be low value, because these values represent "error value". The low level of these statistics in the model indicates that the model performance is high.
Table 3
Correlation Between Real Risk and Predicted Risk | r (Real & Predicted) | p |
0,981 | < 0,001 |
It is possible to demonstrate the effectiveness of the model created with trained data loaded in the ANFIS data field, by comparing it with real outputs. The computer program was performed on SPSS (IBM, version 25). There was determined statistically significant strong and positive relationship between real outputs and ANFIS predicted outputs at the level of α = 0,05 (r = 0,981;p < 0,001).
It was determined that the correlation between the predicted values and the actual values obtained with the new Fuzzy model, which was established by adding the Erythrocyte Sedimentation rate of the patients to the existing input parameters, is 0.632. This was determined by determining the value of the correlation between the predicted values and the actual values obtained with the old model. This demonstrates that the model generated using ANFIS makes more accurate predictions (r = 0.632; p0.001) than other models.
Table 4
Comparing real outputs and ANFIS predicted outputs.
| Mean | n | Std-dev. | Std. Error Mean | t (Real & Predicted) | P value |
Real Risk Score | 0,624 | 185 | 0,178 | 0,01313 | 0,279 | 0,780 |
Predicted Risk Score | 0,618 | 185 | 0,189 | 0,01394 |
P-value, t test statistics, mean, std deviation and std. Error mean values were given in Table 4. Accordingly, no statistically significant difference was found between ANFIS prediction outputs and actual outputs in terms of averages at the level of α = 0,05 (t = 0,279; p > 0,05). As a result, it can be said that ANFIS predicted model is successful in predicting the real risk scores.
Figure 1 shows the comparison of ANFIS predicted outputs and actual outputs. Actual risk outputs and forecast risk outputs were collected on almost the same line. It was also graphically revealed of predictive power of the model.