3.1 Subject demographics
Ninety PD patients underwent PET with two inflow volumes: 1500 ml (20 patients, group 1) and 2000 ml (70 patients, group 2). There were no differences in the distributions of age, gender, underlying diseases, or durations of PD treatment between the two groups (see Table 1).
===Table 1 inserted here===
3.2 Significant variables used in this study
We identified a set of variables that provided the best discrimination between the two groups (see Table 2). The discriminator variables providing the best discrimination are height, weight, BMI, WC, effluent volume, P0, and P4 in IPP. Only those five were significantly and practically extracted. That is, Student's t-test suggests that the two groups are statistically different at a significant level of 0.05 about the five variables mentioned above (see Table 2).
==Table 2 inserted here===
3.3 Formulas predicting inflow volume
3.3.1 EFA used for selecting domain variables using the original data
Two domains composed of (1) P0, P4, and “effluent volume”; and (2) BMI and waist circumference were observed by using EFA.
The three methods (i.e., alternatives from A to C in Table 3) used for discriminating the study groups are equivalently equal, with a high correct prediction rate (94.4%). The AUCs approach 0.98. The only difference results are from the standard errors(SE). See Figure 2. It is noting that the smaller SE(=0.01) is from Logistic regression (LG). The predicted variable using LG is Z (= 4.32974 + 3.85477 * F1 + 3.83008 * F2 based on the sample factor scores) for calculating optimal dialysate DV, see Table 3.
==Table 3 and Figure 2 inserted here===
We plotted the rotation axis with the two domain scores of 90 patients in Figure 3. It is evident from using the rotation method at 35 degrees of the angle that only five subjects were misclassified. All other equations (i.e., discrimination functions and logistic regression) have a high correlation (>0.97) with one of the rotation equations in Figure 3. As a result, the predicted accuracies of those classifications are similar (see Table 3). Interested readers are invited to san the QR-code in Figure 3 to see the EFA process for extracting factors and featured variables from the study data.
== Figure3 inserted here===
Figure 3 The plot of study data and a new axis of Z (new axis) for alternative A.
However, the personal factor scores are usually unknown when answering the app on a smartphone for classifying appropriate DV. The personal factor scores are thus required to be transformed from the original responses on the App.
3.3.2 Prediction accuracy using variables including effluent volume
Two Equations for obtaining the personal factor score are shown at Step D in Table 4 using multiple regression analysis to estimate model parameters. Through this, we can get an accuracy rate of 92.22% shown at Step E in Table 4.
However, the effluent volume might be a collinear variable to the predicted DV. Furthermore, patients are to know the appropriate DV without the previous record of eluent volume at the first treatment on PD. The accuracy rate using featured variables but eluent volume to predict types of the PD DV for patients is required.
==Table 4 inserted here===
3.3.3 Validation when Excluding the Variable of Effluent Volume
Due to the close relationship between the inflow and effluent volume as well as the new CAPD patients who have not the experience of applying the type of DV, step D in Table 5 was performed by using LG according to the sample factor scores. The accuracy rate is 89.47% and the ROC=0.976, see the bottom in Figure 2.
Similar to the previous section, the personal factor scores should be known by using the model parameters of multiply regression analysis at Step G in Table 5. That is, two factors were predicted by two respective equations with factor scores at step G. Parameters were derived from factor scores in EFA (i.e., P0 and P4 in Factor 1 and BMI and WC in Factor 2 for individual PD patients). Through this, the original patients using the validation data can be regrouped again by Steps H.
The prediction formula is thus determined at Step H (i.e., Z = 2.80498 + 2.41593 * F1 + 2.74587 * F2 with mode H in Table 3) developed excluding the variable of effluent volume. The correct prediction rate is 89.01% with AUC = 0.958, slightly lower than that, including effluent volume at the correct prediction rate of 92.22%, but much higher than the previous study at 80.68%[4].
==Table 4 inserted here===
3.4 An App for Determining DVs for PD patients
An app with a QR code as seen at the top in Figure 4 is demonstrated in Additional File 1. Interested readers are invited to scan the QR code in Figure 4 to see the group classification response when entering the required data (i.e., P0, P4, BMI, and WC [cm]) at the left-top corner at Step 1 in Figure 4.
The results are immediately shown on the screen at Steps 2 and 3 in Figure 4 according to the cutting point at 3.79, see Step H in Table 5 . If the effluent volume has been entered to predict DV jointly at Step 1in Figure 4, the next calculation for DV and the visual display are presented at Steps 3 in Figure 4. The confidence is denoted by the probability of the classification on the axis Y suggested for physicians who can adjust DV more specifically, not limited to either 1500 or 2000 ml, as was the case in daily practice.. For instance, the confidence(=probability) for classifying the DC of 1500 ml (on the curve 1500 ml) is 0.75 in Figure 4. higher than 0.5(or odds=1.0) at the intersection of these two curves, implying the tendency of the classification is toward 1500 ml instead of toward 2000 ml.
== Figure 4 inserted here===