Comparison study between neural network model and multiple linear regression in predicting pulmonary artery obstruction index in pulmonary embolism

Objective: To compare the predictive capability of multiple linear regression (MLR) and neural network model (NNM) for pulmonary artery obstruction index (PAOI) in pulmonary embolism. Methods: One hundred and forty-seven APE patients (79 male, 68 female) were collected from March 2015 to July 2016 in our hospital and randomly divided into training group and testing group with the ratio of 3 : 1. Four indexes, including total volume (V), total length (L), total degree of embolism (D) and total number of clots (N) were calculated by computer assisted detection. Qanadli index (Q) as CT PAOI was calculated manually. With SPSS 14.2 modeler, the predictive value of Qanadli index ((Q)) was calculated by MLR and NNM respectively, with Qanadli index as dependent variable and V, L, D, N as independent variables. SPSS 22.0 Spearman test was used to analyze the correlation between (Q) and Q. Mean absolute error (MAE), mean relative error (MRE), root mean square error (RMSE) were used to quantify the accuracies of two methods. Results: MLR equation was (Q) = 10.98 + 1.37 × V + 0.06 × L, model fitting was 0.764. NNM included one hidden layer and two neurons with accuracy of 0.868. In training group, the correlation between (Q) and Q in NNM (r = 0.932, P < 0.01) was higher than MLR (r = 0.879, P < 0.01); in testing group, the correlation between (Q) and Q in NNM (r = 0.875, P < 0.01) was higher than MLR (r = 0.868, P < 0.01). In training group, MAE, MRE and RMSE of NNM (5.144, 0.274, 6.957) were significantly lower (t = 3.402, P = 0.002) than MLR (6.784, 0.282, 8.700); in testing group, MAE, MRE and RMSE of NNM (6.643, 0.312, 9.195) were significantly lower (t = 3.383, P = 0.002) than MLR (8.505, 0.334, 10.361). Conclusion: NNM is a better model in predicting CT pulmonary artery obstruction index of APE patients. (authors)