Predictive model for estimating the death risk of in-hospitalized injured patients from a large scale disaster： a pattern recognition study based on patient data sets after the 2008 Wenchuan Earthquake, China

Abstract

Objective Massive earthquake is one of disasters resulting in huge numbers of heavy and serious casualties. Identifying risk factors that lead to organ failure and death is crucial for improving trauma service performance. Pattern recognition technique （PRT） is a new tool for mining important information and in turn can generate new knowledge from huge amount of data. Here we use PRT to identify patterns in the cause of deaths of trauma patients from a massive earthquake in the Wenchuan, China. Methods Weconducted a retrospectively data mining study. The data used is from a total of 2, 316 casualties ambulated to the Sichuan Academy of Medical Sciences （SAMS） Trauma Service from May 12 to 20, 2008 after a massive earthquake. Before analysis, data preprocessing and cleansing were conducted. We categorized patient data by survival/non-survival and MODS/non-MODS. According to the result of distribution test, quantitative data was described by mean ＋ standard deviation （SD） or median （ quartile）, Student t testing or Wilcox testing was employed. Qualitative data was described by ratio, X2 testing of fisher testing was employed. After mortality and multiple organ dysfunctional syndromes （MODS） related variables are acquired, partial least squares discriminant analysis （PLS-DA） algorithm was used to establish mortality and MODS correlation model. We adopted two principle components to establish PLS projection plotting, and used variable important projection （VIP） to screen variables that correlated with clinical outcome. Receiver operating characters （ROC） curve was used for sensitivity and specificity analysis. Results The records of 1919 patients were selected by data cleansing, and 31 demographical, physiological-biological parameters and intervention factors were acquired as exposure variables. There were 36 in-hospital death cases, and 17 MODS cases. MODS related mortality was 47. 1% （8/17）. In PLS-DA, the first two principal components in the scatter plot could distinguish survival, MODS and deceased patients. For predicting mortality and MODS, the AUC of ROC was 0. 882 and 0. 979, respectively. VIP indicator （ variable importance for the projection） of PLS-DA identified 8 physiological variables （pH, BE, PaCO2, PaO2, HCO3-1, SBHCO3, Cr, volume of fluid resuscitation at the first day in-hospital） comprised a pattern related to in-hospital death event and MODS. Condusions This study shows that a significant pattern （comprised by a set of physiological-biological and fluid resuscitation intervention when patient reach to trauma service） emerges which can predict survival probability for hospitalized casualties injured during a massive earthquake. Application of this model may provide a tool to help disaster health providers identify most-at-risk patients, especially after a massive disaster when limited medical resources have to cope with huge numbers of victims. Key words: Earthquake injury; Trauma; Mortality; Big data; Multiple organ dysfunctional syndrome; Pattern recognition technique; Partial least squares discriminant analysis; Data mining; Case- control