Bodil Andersson, Roland Andersson, Mattias Ohlsson, Johan Nilsson
Prediction of Severe Acute Pancreatitis at Admission to Hospital Using Artificial Neural Networks
LU TP 11-49
Abstract:
Background/Aims: Artificial neural networks (ANNs) are non-linear pattern recognition techniques, which can be used as a tool in medical decision-making. The aim of this study was to construct and validate an ANN model for early prediction of the severity of acute pancreatitis (AP).
Methods: Patients treated for AP from 2002 to 2005 (n = 139) and from 2007 to 2009 (n = 69) were analyzed to develop and validate the ANN model. Severe AP was defined according to the Atlanta criteria.
Results: ANNs selected 6 of 23 potential risk variables as relevant for severity prediction, including duration of pain until arrival at the emergency department, creatinine, hemoglobin, alanine aminotransferase, heart rate, and white blood cell count. The discriminatory power for prediction of progression to a severe course, determined from the area under the receiver-operating characteristic curve, was 0.92 for the ANN model, 0.84 for the logistic regression model (p = 0.030), and 0.63 for the APACHE II score (p < 0.001). The numbers of correctly classified patients for a sensitivity of 50 and 75% were significantly higher for the ANN model than for logistic regression (p = 0.002) and APACHE II (p < 0.001).
Conclusion: The ANN model identified 6 risk variables available at the time of admission, including duration of pain, a finding not being presented as a risk factor before. The severity classification developed proved to be superior to APACHE II.
Background/Aims: Artificial neural networks (ANNs) are non-linear pattern recognition techniques, which can be used as a tool in medical decision-making. The aim of this study was to construct and validate an ANN model for early prediction of the severity of acute pancreatitis (AP).
Methods: Patients treated for AP from 2002 to 2005 (n = 139) and from 2007 to 2009 (n = 69) were analyzed to develop and validate the ANN model. Severe AP was defined according to the Atlanta criteria.
Results: ANNs selected 6 of 23 potential risk variables as relevant for severity prediction, including duration of pain until arrival at the emergency department, creatinine, hemoglobin, alanine aminotransferase, heart rate, and white blood cell count. The discriminatory power for prediction of progression to a severe course, determined from the area under the receiver-operating characteristic curve, was 0.92 for the ANN model, 0.84 for the logistic regression model (p = 0.030), and 0.63 for the APACHE II score (p < 0.001). The numbers of correctly classified patients for a sensitivity of 50 and 75% were significantly higher for the ANN model than for logistic regression (p = 0.002) and APACHE II (p < 0.001).
Conclusion: The ANN model identified 6 risk variables available at the time of admission, including duration of pain, a finding not being presented as a risk factor before. The severity classification developed proved to be superior to APACHE II.
LU TP 11-49