OBJECTIVE: To compare the results of a neural network versus a logistic regression model for predicting early (0-3 months) pancreas transplant graft survival or loss. RESEARCH DESIGN AND METHODS: This study was a cross-sectional, secondary analysis of demographic and clinical data from 117 simultaneous pancreas-kidney (SPK), 35 pancreas-after-kidney (PAK), and 8 pancreas-transplant-alone (PTA) patients (n = 160). The majority of patients were men (57%) and were white (90.1%), with a mean age of 39 +/- 8.09 years. Of the patients, 23 (14.4%) experienced early graft loss, which included any loss owing to technical or immunological causes, and death with a functional graft. Data were analyzed with a logistic regression model for multivariate analysis and a backpropagation neural network (BPNN) model. RESULTS: A total of 12 predictor variables were chosen from literature and transplant surgeon recommendations. A logistic model with all predictor variables included correctly classified 93.53% of cases. Model sensitivity was 35.71%; specificity was 100% (pseudo-R2 0.24). Of the predictors, history of alcohol abuse (odds ratio [OR] 32.39; 95% CI 1.67-626.89), having a PAK or PTA (OR 13.6; 95% CI 2.20-84.01), and use of a nonlocal organ procurement center (OPO) (OR 4.51; 95% CI 0.78-25.96) were most closely associated with early graft loss. The BPNN model with the same 12 predictor variables correctly predicted 92.50% of cases (R2 0.71). Model sensitivity was 68%; specificity was 96%. Of the predictors, the three variables most closely associated with graft outcome in this model were recipient/donor weight difference >50 lb, having a PAK or PTA, and use of a nonlocal OPO. CONCLUSIONS: First, the BPNN model correctly predicted 92.5% of graft outcomes versus the logistic model (93.53%). Second, the BPNN model rendered more accurate predictions (>0.70 = loss; <0.30 = survival) versus the logistic model (>0.50 = loss; <0.50 = survival). Third, the BPNN model was more sensitive (68%) than the logistic model (35.71%) to graft failures and demonstrated an almost threefold increase in explained variance (R2 = 0.71 vs. 0.24). These results suggest that the BPNN model is a more powerful tool for predicting early pancreas graft loss than traditional multivariate statistical models.
OBJECTIVE: To compare the results of a neural network versus a logistic regression model for predicting early (0-3 months) pancreas transplant graft survival or loss. RESEARCH DESIGN AND METHODS: This study was a cross-sectional, secondary analysis of demographic and clinical data from 117 simultaneous pancreas-kidney (SPK), 35 pancreas-after-kidney (PAK), and 8 pancreas-transplant-alone (PTA) patients (n = 160). The majority of patients were men (57%) and were white (90.1%), with a mean age of 39 +/- 8.09 years. Of the patients, 23 (14.4%) experienced early graft loss, which included any loss owing to technical or immunological causes, and death with a functional graft. Data were analyzed with a logistic regression model for multivariate analysis and a backpropagation neural network (BPNN) model. RESULTS: A total of 12 predictor variables were chosen from literature and transplant surgeon recommendations. A logistic model with all predictor variables included correctly classified 93.53% of cases. Model sensitivity was 35.71%; specificity was 100% (pseudo-R2 0.24). Of the predictors, history of alcohol abuse (odds ratio [OR] 32.39; 95% CI 1.67-626.89), having a PAK or PTA (OR 13.6; 95% CI 2.20-84.01), and use of a nonlocal organ procurement center (OPO) (OR 4.51; 95% CI 0.78-25.96) were most closely associated with early graft loss. The BPNN model with the same 12 predictor variables correctly predicted 92.50% of cases (R2 0.71). Model sensitivity was 68%; specificity was 96%. Of the predictors, the three variables most closely associated with graft outcome in this model were recipient/donor weight difference >50 lb, having a PAK or PTA, and use of a nonlocal OPO. CONCLUSIONS: First, the BPNN model correctly predicted 92.5% of graft outcomes versus the logistic model (93.53%). Second, the BPNN model rendered more accurate predictions (>0.70 = loss; <0.30 = survival) versus the logistic model (>0.50 = loss; <0.50 = survival). Third, the BPNN model was more sensitive (68%) than the logistic model (35.71%) to graft failures and demonstrated an almost threefold increase in explained variance (R2 = 0.71 vs. 0.24). These results suggest that the BPNN model is a more powerful tool for predicting early pancreas graft loss than traditional multivariate statistical models.