| Background and Objection:Urinary stone is a common disorder of urology, and the prevalence of urolithiasis is higher in some provinces of southern China. Following the improvement of nutritional condition, the incidence of upper urinary tract stone became significantly higher in the recent decade. Since the advent in the early80s of the last century, extracorporeal shock wave lithotripsy (ESWL)cured numerous patients with upper urinary tract stones, who had to undergo surgery before. ESWL has been proved to be safe, effective, less painful, rapid to recovery and inexpensivedue to its minimally invasive character, and low morbidity. However, for larger upper urinary tract stones, ESWL may need to repeat several times or combine others methods of treatment. But repetition of ESWL may lead to irreversible damage to kidney. Furthermore, the process of stone fragments removing is often accompanied by some complications such as renal colic, fever, hematuria, steinstrasse, which may also result in conversion to surgery. These morbidities may compromise the clinical outcome of ESWL-treated calculi, thus not all upper urinary tract calculi are suitable for ESWL.The patient would pay for the consequences of unsuccessful ESWL with a cost of money, time, and an experience of distressful medical treatment.If the outcome of ESWL can be predicted, on one hand, the clinicians could develop relatively correct strategy to avoid the patient’s unnecessary cost, to reduce the cost of medical services; On the other hand, the prediction could help patients be well informed and help doctors share the decision-making with patients.In this study, logistic regression (LR) analysis and artificial neural network (ANN) were used to establish the predictive model of ESWL for renal and ureteral calculi, respectively.Under logistic regression models, the important impact factors influencing outcome of ESWL for upper urinary tract stones were determined. By ANN, the importance of predictor variables was evaluated. A comparative study afterwards was done between LR model and ANN to compare the accuracy of prediction and pros&cons between these two models of prediction. Eventually the cut-off values of optimal probabilities were identified to facilitate the clinical application.Methods and materialsPre-ESWL clinical data of325cases of kidney stones treated with ESWL and those of1065cases of ureteral calculi were included retrospectively. The information of gender, urinary irritation symptom, hematuria, renal colic, stone location, stone side, age, height, weight, history of stone, stone length, stone width was collected.All cases of kidney stones and ureteral stones were grouped separately with a randomized method. For LR analysis and modeling, about70%of cases were assigned to training-sample group, and approximate30percent of cases were assigned to holdout-sample group. For ANN, about56%of the cases were assigned to training-sample group, and about14%of the cases assigned to test-sample group, about30%of cases were assigned to the holdout-sample group.χ2test was performed for the significance test of relationship between the ESWL outcome and every categorical variable (gender, urinary irritation symptom, hematuria, renal colic, stone location, stone side).The significant test of continuous predictive variables (age, body mass index, history of stone, the product of stone length&width) were done by univariate logistic regression to initially identify the corresponding influential factors of ESWL outcome.Then all the predictive variables which have P<0.25, together with other important predictive variables were included as candidate predictive variables for multivariate analysis. Thereafter logistic stepwise regression analysis (forward:LR) was performed. The training sample and the holdout sample were respectively substituted to the LR model, and the respective predictive probabilities for ESWL outcome were obtained. Then the area under curve (AUC), sensitivity, specificity and were calculated to evaluate the accuracy of prediction after drawing receiver operating characteristic curve (ROC).Three-layer forward neural network model was constructed. Hidden layer was set to1layer; the parameters of the input layer comprise all10predictive variables. Predictive factors were set to be gender, urinary irritation symptom, hematuria, renal colic, stone location, and stone side. Predictive covariates were age, BMI, history of stone, stone size. Value of gender was coded by male=1, female=2. And the value of urinary irritation symptom, hematuria, renal colic were set to yes=1, no=2. In the cases of kidney stone, location of stone in the kidney was set to the upper calyces=1, the middle calyces=2, the lower calyces=3, the pelvis=4, combined location=5. In the cases of ureter stone, it was set to upper ureter=1, middle ureter=2, lower ureter=3.The output layer is ESWL-treated calculi outcome. Successful and failed outcome is set to1and0, respectively. All cases were assigned randomly basing on partitioning variable aforementioned."Automatic architecture selection" was chosen, implying that the minimum number of units in hidden layer is1, and the maximum number is50. The type of training was chosen as "batch"(for renal calculi sample) or "mini-batch"(for ureteral calculi sample),"optimization algorithm selection"(for renal calculi sample) and "scaled conjugate gradient"(for ureteral calculi sample). Initial Lambda value is0.0000005, and the initial Sigma value of is0.00005, the interval center is0, the interval offset is±0.5.The predictive pseudo probabilities of ESWL outcome were obtained.0.50was set as the cut-off value of predictive pseudo probabilities, and the area under curve (AUC), sensitivity, and specificity were calculated to evaluate the accuracy of prediction after drawing receiver operating characteristic curve (ROC). The efficacy together with advantages&disadvantages of LR predictive model and ANN were then evaluated and compared.The statistical package of SPSS20.0published by IBM(?) were used statistically, by which χ2test, univariate logistic regression, multivariate logistic analysis and modeling were performed. Hosmer-Lemeshow goodness of fit test for LR model and the χ2test are used for significance test of LR model. P value is accurate to three decimal. P<0.05is for significant difference. The statistical package of Medcalc(?) was used for drawing of ROC curve and the comparison of AUC and for calculation of Youden index, then the cut-offs value of optimal predictive probabilities were identified.Results:A. Prediction of ESWL outcome for kidney stones casesAmong a total of325cases of kidney stones,250cases (76.9%) were free of stones at3months.Post-ESWL auxiliary treatment was required in75cases who failed to respond to the ESWL treatment. The longest follow-up attained three months.The success rate of ESWL was significantly affected by stones side, stone location, urinary irritation symptom, and hematuria.Additionally, after further LR multivariate analysis to the training-sample, history of stone, hematuria and the product of stone length&width were found to be3independent impact factors of the outcome of ESWL treatment of kidney stones (P<0.05) with AOR (Adjusted Odds Ratio)(95%confidence interval) being0.977(0.964-0.989),12.388(3.443-44.565),0.192(0.113-0.323), respectively.χ2value of LR prediction model is108.938(P<0.001), which means the predictive variables have significant explanatory power for the dependent variable (ESWL outcome). The result of Hosmer-Lemeshow goodness of fit test indicated that the prediction model could represent the data well (χ2=6.927, df=8, P=0.545).To evaluate the accuracy of prediction of LR model, the prediction-observation cross classification table was employed. The cut-off value of probabilitywas set at0.5. The results showed that the training sample’s sensitivity, specificity, overall accuracy rate were94.4%63.3%86.0%, respectively. The sensitivity, specificity, overall accuracy rate of holdout samples were100%,20.0%,88.4%, respectively.The ANN automatically removed the typical modeling process "redundant" unit, and established an input layer with of19units, a hidden layer of6units and an output layer of2units. The activation function of input was set as "hyperbolic tangent", The activation function of output was set as "softmax". The order of importance of predictive variables showed that stone size, history of stone, hematuria, stone location were listed on the top five. At0.5of cutoff value of probability, the prediction of ANN showed that the sensitivity, specificity, overall accuracy rate of the training sample were98.4%,72.5%and90.8%,respectively; Those of the testing sample were97.4%,55.6%and89.6%. Those of holdout sample were93.3%,46.7%and86.5%.The AUC of the LR model is0.625(95%confidence interval:0.525-0.718), whereas ANN model is0.856(95%confidence interval:0.774-0.917). Comparing with AUC=0.5, the former’s P value is0.122, and the latter’s P value0.001. The medical statistics Packages Medcalc was employed for testing the difference of two AUCs, the z value is3.988, P=0.0001, which means there exist statistically significant differences between the AUCs of two models. That is to say, the prediction of ANN model has higher ability than that of the LR model.The optimal cutoff value of probability of ANN was identified to0.595, with consideration of balancing the sensitivity and specificity. At this cutoff value, the sensitivity and specificity of ANN model were thought to be92%and60%, respectively.Prediction of ESWL outcome for ureter stones casesAmong a total of1,065cases of ureter stones,874cases (82.1%) were free of stones at3months,191cases failed of ESWL. All failed cased were converted to other auxiliarytreatments. The longest follow-up duration reached3months. Gender, stone side, stone location, urinary irritation symptom, renal colic were found to have significant impacts on the outcome of ESWL. Univariate logistic regressions indicated age, BMI, history of stone, stone size were among the significant contributors to the success rate of ESWL.Further using LR multivariate analysis to training samples, renal colic, stone location (upper part of ureter and middle part of ureter), stone size were found to be4independent impact factors of outcome of ESWL treatment for ureter stones, with AOR (95%confidence interval) were1.508(0.999-2.277),0.651(0.391-1.086), 0.374(0.191-0.731),0.246(0.151-0.396), respectively.χ2test was applied to verify the significance of LR prediction model, χ2value was54.460(P<０.001), which means these predictive variables have significant explanatory power for the dependent variable (ESWL outcome). Hosmer-Lemeshow goodness of fit indicated that there was no statistically significant difference (χ2=8.406,df=8,P=0.395).To evaluate the accuracy of prediction of LR model, the prediction-observation cross classification table was employed. The cut-off value is set at0.5. The results showed that the training sample’s sensitivity, specificity, overall accuracy rate were98.1%,4.2%,82.9%, respectively, meanwhile the sensitivity, specificity, overall accuracy rate of the holdout samples were99.3%,11.3%,84.7%, respectively.The ANN automatically removed the typical modeling process "redundant" unit, and established an input layer with of17units, a hidden layer of5units and an output layer of2units."Hyperbolic tangent" was selected as the activation function of input. The activation function of output was "softmax". The order of importance of predictive variables showed that stone size, stone location, history of stone time, age and BMI were listed on the top five. Atthe cutoff value of0.5, the prediction of ANN showed that the sensitivity, specificity, overall accuracy rate of the training sample were98.8%,12.7%,84.0%.respectively; Those of the testing sample were99.2%,16.7%,89.3%. Those of the holdout sample were97.8%,9.4%,83.2%.The AUC of the logistic regression model was0.729(95% confidence interval:0.676-0.777), whereas ANN model was0.751(95% confidence interval:0.700-0.797). Comparing with AUC=0.5, theirs P values were less than0.001. With the medical statistics Packages Medcalc(?) for non-parametric tests, the z value is0.750, P=0.4534, which means there is no statistically significant differences between the two AUC. That is to say, the prediction of ANN model may have the same predictive efficacy as that of the LR model.The cutoff value of probabilities of ANN was identified to0.7694, in consideration of balancing the sensitivity and specificity. The optimal sensitivity and specificity were thought to be81%and60%, respectively.Conclusion:This study found that history of stone.the product of stone length&width, hematuria are important impact factors for outcome of ESWL for kidney stones, whereas renal colic, stone location in ureter, and stone size have independent impacts on outcome of ESWL for ureter stones. Logistic stepwise regression prediction model and ANN were also constructed. In comparison with LR model, ANN with self-learning, parallel computing functions has a predictive ability. Although the meaning of predictive variables of ANN is hard to explain, it could still be used to predict ESWL efficacy as a powerful tool. ANN should contribute to the screening of patients more suitable for ESWL treatment. It may help fully share decision-making with patients, and reduce patients’ care costs and economic burden. |
PDF Full Text Request