BACKGROUND: The ability to rapidly and accurately triage, evacuate, and utilize appropriate interventions can be problematic in the early decision-making process of trauma care. With current methods of prehospital data collection and analysis, decisions are often based upon single data points. This information may be insufficient for reliable decision-making. To date, no studies have attempted to utilize data at multiple time points for purposes of enhancing prediction, nor have studies attempted to synthesize prediction models with data reflecting both large-vessel venous and arterial injuries. Therefore, we performed a retrospective study to examine the potential utility of dynamic neural networks in predicting mortality using highly discretized uncontrolled hemorrhagic shock data. METHODS: One hundred forty-three swine with either grade V liver injuries or 2.8-mm aortotomies had hemodynamic data collected every minute throughout injury and resuscitation. The independent variables used as inputs to the polynomial neural networks (PNNs) included systolic blood pressure and mean arterial pressure (MAP). These inputs were used to predict mortality in individual swine 1 hour after injury using data up to 20 minutes after injury. Survival models were compared based on discrimination power (DP), i.e., where specificity equals sensitivity, and area under the receiver operating characteristic (ROC) curve (c-statistic). The Hosmer-Lemeshow (H-L) statistic was used to measure model calibration. RESULTS: The best PNN model predicted mortality at 60 minutes utilizing data from injury to 20 minutes after injury. This model produced a ROC area of 0.919, a DP of 0.857, and a H-L value of 16.47. A DP of 0.857 means that 85.7% of the survivors are correctly predicted to survive, and 85.7% of the nonsurvivors are predicted to die. MAP of survivors and nonsurvivors were graphed for comparative purposes. As this graph illustrates, the use of MAP alone cannot discriminate survivors from nonsurvivors. CONCLUSION: This study demonstrates that PNN models can effectively harness the dynamic nature of uncontrolled hemorrhagic shock data, despite utilizing data from large-vessel arterial and venous injuries. Utilizing the dynamic nature of hemorrhagic shock data in PNNs may ultimately allow the development of novel decision assist devices.
BACKGROUND: The ability to rapidly and accurately triage, evacuate, and utilize appropriate interventions can be problematic in the early decision-making process of trauma care. With current methods of prehospital data collection and analysis, decisions are often based upon single data points. This information may be insufficient for reliable decision-making. To date, no studies have attempted to utilize data at multiple time points for purposes of enhancing prediction, nor have studies attempted to synthesize prediction models with data reflecting both large-vessel venous and arterial injuries. Therefore, we performed a retrospective study to examine the potential utility of dynamic neural networks in predicting mortality using highly discretized uncontrolled hemorrhagic shock data. METHODS: One hundred forty-three swine with either grade V liver injuries or 2.8-mm aortotomies had hemodynamic data collected every minute throughout injury and resuscitation. The independent variables used as inputs to the polynomial neural networks (PNNs) included systolic blood pressure and mean arterial pressure (MAP). These inputs were used to predict mortality in individual swine 1 hour after injury using data up to 20 minutes after injury. Survival models were compared based on discrimination power (DP), i.e., where specificity equals sensitivity, and area under the receiver operating characteristic (ROC) curve (c-statistic). The Hosmer-Lemeshow (H-L) statistic was used to measure model calibration. RESULTS: The best PNN model predicted mortality at 60 minutes utilizing data from injury to 20 minutes after injury. This model produced a ROC area of 0.919, a DP of 0.857, and a H-L value of 16.47. A DP of 0.857 means that 85.7% of the survivors are correctly predicted to survive, and 85.7% of the nonsurvivors are predicted to die. MAP of survivors and nonsurvivors were graphed for comparative purposes. As this graph illustrates, the use of MAP alone cannot discriminate survivors from nonsurvivors. CONCLUSION: This study demonstrates that PNN models can effectively harness the dynamic nature of uncontrolled hemorrhagic shock data, despite utilizing data from large-vessel arterial and venous injuries. Utilizing the dynamic nature of hemorrhagic shock data in PNNs may ultimately allow the development of novel decision assist devices.
Authors: Peter Hu; Gregory Defouw; Colin Mackenzie; Christopher Handley; Steven Seebode; Phil Davies; Douglas Floccare; Yan Xiao Journal: AMIA Annu Symp Proc Date: 2006
Authors: Anthony Wertz; Andre L Holder; Mathieu Guillame-Bert; Gilles Clermont; Artur Dubrawski; Michael R Pinsky Journal: Crit Care Explor Date: 2019-10-30