Prediction of Mortality in a Patient With Acute Poisoning

Objective: This study aimed to develop a scoring system for predicting the in-hospital mortality of acute poisoning patients at the emergency department (ED). Methods: This was a retrospective analysis of the Injury Surveillance Cohort generated by the Korea Center for Disease Control and Prevention (KCDC) from 2011–2018. We developed the new-Poisoning Mortality Scoring system (new-PMS) to generate a prediction model using the derivation group (2011– 2017 KCDC cohort). Points were computed for each category of each variable. The sum of these points was the new-PMS. The validation group (2018 KCDC cohort) was subjected to external temporal validation. The performance of new-PMS in predicting mortality was evaluated using receiver operating characteristic (ROC) curves for both groups. For simple interpretation in clinical settings, risk groups were categorized as very low, low, intermediate, and high according to the new-PMS; we suggested the mortality curve according to new-PMS. Results: Of 57326 poisoning cases, 42568 were selected. Of these, 34352 (80.7%) and 8216 (19.3%) were enrolled in the derivation and validation groups, respectively. New-PMS was the sum of points for each category of 10 predictors. The range of new-PMS was -20 to 3420 points. The area under the ROC curve of new-PMS was 0.942 (95% CI: 0.934–0.949) and 0.946 (95% CI: 0.930–0.963) for the derivation and validation groups, respectively. The mean predicted mortality and the observed mortalities of the high-risk group (new-PMS ≥ 1048) were 9.7% (95% CI: 9.3 – 10.0) and 10.0% for the derivation group and 8.4% (95% CI: 7.7 – 9.1) and 7.4% for the validation groups, respectively. Conclusions: New-PMS showed good performance in predicting in-hospital mortality for both groups. As mortality sharply increased with the high risk-group of the new-PMS, early hemodynamic stabilization of acute poisoning patients at the ED may improve their clinical outcomes. New-PMS contributes to clinical decision-making

of poisoning patients depends on not only their physiological condition but also the unique characteristics of the poisoning. The type of substance, route of exposure, and intent of poisoning affect the outcomes of acute poisoning patients. In addition, the toxic substance and its lethality are often unknown. A prediction model of mortality for acute poisoning patients has to consider both poisoningrelated characteristics and the patient's physiological condition; moreover, it must be applicable to poisoning patients of all ages. The objective of this study was to develop a scoring system for predicting mortality in acute poisoning patients at the emergency department (ED). This work will assist in treatment allocation and therapeutic decision-making for acute poisoning patients at the early stages of poisoning.

Study design and selection of study patients
This study was a retrospective analysis of a prospective cohort from 23 EDs, namely the Injury Surveillance Cohort, which was generated by the Korea Center for Disease Control and Prevention (KCDC) from 2011 to 2018. This registry comprised prospectively collected data on the epidemiology and outcome variables of injury patients who presented at an ED [8]. The registry included poisoning cases as a type of injury. We selected poisoning patients from this cohort. This selected registry included the baseline characteristics of poisoning patients: age; sex; time-related factors, such as ED presenting time and poison exposure time; poisoning-related variables, such as the intent of poisoning, route of exposure, type of substance (7 categories and 44 types of substance); and initial vital signs at ED presentation, such as systolic blood pressure (SBP), heart rate (HR), respiration rate (RR), body temperature (BT), and AVPU scale of mental status. The registry also contained outcome-related variables, such as mortality at the ED or after hospital admission.
Patients who were transferred from the initial ED to another hospital; those who had incomplete data of poisoning-related variables, initial physiological condition-related variables, or outcome-related variables; and those who died on arrival at the ED were excluded from this study ( Figure 1).
The selected study population was divided into two groups, namely the derivation group for the prediction of in-hospital mortality and the validation group for the external validation of the developed prediction model ( Figure 1).
The Institutional Review Board of the Korea University hospital approved this study (IRB No. 2020AN0195).

Data analysis
The primary outcome was in-hospital mortality. We compared the characteristics of the poisoning patients between the derivation and validation groups (Table 1). Age, sex, time from exposure to ED presentation, classes of substance, intent of poisoning, route of exposure, vital signs of the patient at ED presentation, and in-hospital mortality were analyzed (Table 1). For analysis, the variables related to poisoning characteristics were categorized as follows: intent of poisoning: 1) unintentional, 2) intentional, and 3) unknown; route of exposure: 1) dermal, ocular, or injection; 2) oral; and 3) inhalation; and toxic substance included 44 kinds of substances that were classi ed into eight categories from A to H. For categorization of substances, we considered the classi cation in the types of substances. And then we categorized the substances in the same classi cation according to the mortality index (MI) of each substance: A) pharmaceutical agents with MI of less than 0.5%, B) pharmaceutics with MI 0.5 -5%, C) arti cial toxic substances with MI of less than 1.0%, D) arti cial toxic substances or pesticides with MI of 1.0 -10.0%, respectively, E) arti cial toxic substances or pesticides with MI of 11.0 -20.0%, respectively, F) paraquat with MI of 52.5%, G) gases with MI of less than 1.0%, H) natural toxic substances with MI of less than 1.0% (Table 2)(An additional le 1 shows this in more detail [see Additional le 1]). The patient's physiological variables included age, SBP, HR, RR, BT, and mental status (AVPU scale), in accordance with the predictors in SAPS-II [9]. However, because SAPS-II does not include RR score, we categorized RR according to the normal range (12-24 breaths/min) [10,11].

Development of the new poisoning mortality scoring system
We developed the new-Poisoning Mortality Scoring system (new-PMS) to generate a prediction model for the derivation group (2011-2017 data of the KCDC cohort) ( Figure 1). First, we compared poisoning-and physiological condition-related variables between the patients who survived and were discharged (survivor subgroup) and those who died at the hospital (in-hospital death subgroup) among the derivation group (Table 3). We selected variables that had statistical and clinical signi cance in acute poisoning as predictors for developing the new-PMS [12]. Points for each category of each predictor were computed using multivariable logistic regression, in which the regression coe cient for each category of each predictor was converted into points by dividing the smallest regression coe cient in the model (Table 4) [13]. The sum of these points of each category in each predictor was the new-PMS.

Performance evaluation of the new-PMS
We analyzed the performance of the new-PMS in predicting mortality using receiver operating characteristic (ROC) curves in both the derivation and validation groups. The validation group (2018 data of the KCDC cohort) was subjected to external temporal validation.
For simple interpretation in a clinical setting, we created the following risk groups: very low, low, intermediate, and high risk according to the quartile range of the new-PMS. Real mortalities were investigated in the derivation and validation groups, respectively [12]. Moreover, we generated a mortality curve according to the new-PMS in the derivation group.

Statistical analysis
Continuous variables were reported as the median with interquartile ranges (IQR). Differences in the medians were compared using the Mann-Whiney U-test. Categorical variables were compared using the chi-square test. All statistical analyses were performed using SPSS version 20.0 (IBMSPSS Inc., Chicago IL, USA). Two-tailed p values < 0.05 were considered statistically signi cant.

Selection of the study population and outcomes
Of 57326 poisoning cases, 14758 (25.7%) were excluded: 3399 were transferred out of the ED, 239 had unknown outcomes, 10953 had incomplete data on poison-related variables, and 167 died on arrival at the ED (Figure 1). Of the 42568 included patients, 34352 (80.7%) and 8216 (19.3%) were enrolled in the derivation and validation groups, respectively ( Figure 1). Among the study population, the median time from exposure to ED presentation was 2 h (interquartile range: 1.0-2.0 h). The incidence of in-hospital mortality was 909 (2.6%) and 135 (1.6%) for the derivation and validation groups, respectively (p < 0.001) ( Table 1). The characteristics of the derivation and validation groups are presented in Table 1.

Development of the new-PMS in the derivation group
We compared characteristics between the survivor and in-hospital death subgroups in the derivation group ( Table 2). The demographics, poisoning-related variables, and initial vital signs in each subgroup are shown in Table 2. Patients in the in-hospital death subgroup showed higher likelihood of older age, male sex, intentional poisoning, oral ingestion, and pesticide poisoning; they also initially presented with low SBP, high HR, high RR, and altered mental status compared with those in the survivor subgroup. Time from exposure to ED presentation was not signi cantly different between the survivor and in-hospital death subgroups (p = 0.057).
We selected 10 predictors from these variables considering clinical reasoning and statistical signi cance. The 10 predictors (age, sex, type of substance, intent of poisoning, route of poisoning, SBP, HR, RR, BT, and AVPU scale) and each category of each predictor are presented in Table 4. Multivariable logistic regression was used to calculate the points for each category of each predictor. First, we estimated the regression coe cient (B) for each category of each predictor in the multivariable logistic regression model. Next, the base constant B was selected as the smallest regression coe cient (B) in the model. The base constant B was 0.062 in the multivariable model (Table 4). We converted the regression coe cient (B) for each category of each predictor into points using the formula B/0.005 (Table 3) [13] [13]. The points for each predictor were 0 to 480 for the age categories, 0 and 88 for the sex categories, 0 to 215 for the intent categories, 0 to 202 for the route categories, 0 to 1174 for the substance categories, -19 to 382 for the SBP categories, 0 to 195 for the HR categories, -1 to 140 for the RR categories, 0 and 133 for the BT categories, and 0 to 411 for the AVPU categories ( Table 4). The new-PMS was the sum of the points of each predictor. The minimum to maximum range of the new-PSS was -20 to 3420 points.

Performance evaluation of the new-PMS
The performance of the new-PMS in predicting in-hospital mortality in acute poisoning was signi cantly high, with an area under the curve (AUC) of 0.942 (95% con dence interval [CI]: 0.934-0.949) in the derivation group (p<0.001) (Figure 2(A)). External temporal validation analysis of the new-PMS also showed a signi cantly high AUC of 0.946 (95% CI: 0.930-0.963) (p < 0.001) in the validation group (Figure 2(B)).
The median value of the new-PMS was 786 (inter quartile range: 593-1047) points in the derivation group. In risk-grouping for simple interpretation in clinical settings, patients were classi ed, according to the quartile range of the new-PMS, into four categories: ≤593 points, very low risk; 594-786 points, low risk; 787-1047 points, intermediate risk; and ≥1048, high risk ( Table 5). The mean predicted mortality and the observed mortalities of the high-risk group (new-PMS ≥1048) were 9.7% (95% CI: 9.3 -10.0) and 10.0% for the derivation group and 8.4% (95% CI: 7.7 -9.1) and 7.4% for the validation groups, respectively ( Table 5).
The mortality curve of the new-PMS showed an S-shape in the derivation group ( Figure 3). In the new-PMS of 1048 points or above, the probability of in-hospital mortality increased very sharply (Figure 3).

Discussion
In this study, we developed the new-PMS to predict the probability of mortality in acute poisoning patients. The new-PMS is a simpli ed scoring system that is easy to use in clinical practice. The new-PMS comprises 10 predictors, including patient demographics, poisoning-related factors, and patient's initial vital signs.
Speci c prediction outcome models in toxicology have limited value when applied to a wide range of poisoning patients. The PSS has been used in toxicology as a disease-speci c scoring system [1]. It was developed in the 1990s in Europe to describe a patient's most severe symptomatology [1,4]. However, the PSS has several subjective criteria, is time-consuming, and is likely to be of little use in some types of poisoning, limiting its clinical utility [4]. The development of a new poisoning severity scoring system is required for clinical use. The new-PMS developed in this study has several bene ts, namely the use of objective predictors, the rapid assessment of mortality risk, and early applicability in clinical settings.
Several severity of illness models that have been used in the intensive care unit can be applied to acute poisoning patients. Silakhori [14] reported that the APACHE-II, APACHE-IV, SAPS-II, and SOFA have acceptable discriminatory power for poisoning patients, and APACHE-II can be used for mortality prediction in the early days of admission. However, this previous study included only 150 patients. When predicting the outcome of acute poisoning patients in clinical settings, we have to consider not only the physiological condition of the patient but also the unique characteristics of the poisoning. In the present study, 10 predictors were selected for the development of the new-PMS, considering toxicologic factors and the patient's physiological status. The predictors related to the patient's physiological status were categorized into age and vital signs, in accordance with SAPS-II [15], which has been commonly used to predict outcomes in poisoning patients. The new-PMS re ected the two major characteristics of acute poisoning patients, namely the characteristics of the poisoning and of the early physiological condition after the poisoning.
Given the unique characteristics of individual xenobiotics, many researchers have attempted to apply physiological scoring systems in patients with speci c xenobiotic poisoning [14][15][16][17][18][19]. However, outcome prediction models for speci c toxic substances have a limitation of generalization. In the current study, the new-PMS showed excellent performance in predicting mortality, with an AUC of over 0.9 in all acute poisoning patients, regardless of the cause of poisoning, type of substance, age, and sex. The present study was an attempt to develop a new scoring system for outcome prediction in poisoning patients as an alternative to the PSS.
We used the multivariable logistic regression method to assign points for each category of each predictor. This method is commonly used for the development of prognosis prediction models [12,13]. This approach has been used in numerous studies to create a risk scoring system [20,21]. The reference category of each predictor was determined considering the lowest mortality or normal physiological variable value. For example, the mortality of the 40-year-old age group was 0.04%, which is the lowest among all age groups ( Table 2).
The performance of the new-PMS was excellent according to the general guideline of the AUC in both the derivation and validation groups [22]. In simulation studies, the external validation of a prediction model requires a minimum of 100 events of the primary outcome, as a small external validation study is unreliable and inaccurate [23-25]. Our validation group had 135 mortality cases in a total of 8216 poisoning cases.
For easy use in clinical settings, we constructed the risk groups and a mortality curve of the new-PMS. The observed mortalities of the derivation and validation groups also increased according to the riskgrouping. We expect that the new-PMS may be useful for objective discrimination of the very-low-risk group at the poisoning call center and for allocating acute poisoning patients to poisoning treatment centers at the pre-hospital setting. Furthermore, the risk of mortality sharply increased in acute poisoning patients with the new-PMS of approximately 1048 points or above. These results suggested that early hemodynamic stabilization for high-risk poisoning patients at the ED may improve their clinical outcomes. The new-PMS will contribute to clinical decision-making and the therapeutic guidance of patients with acute poisoning.

Limitations
First, in this study, we excluded cases that had missing values of poisoning-, outcome-, and vital signsrelated variables. The traditional 'complete cases' analysis may lead to selection bias of subjects and statistically ine cient results [12]. In addition, we excluded DOA patients from this study because we considered that these patients required no speci c treatments for acute poisoning. Second, the amount of exposure in the cases of oral ingestion and the duration of exposure in the cases of inhalation or surface absorption are important for predicting the outcomes of acute poisoning patients. Unfortunately, our cohort did not have data on the amount of exposure and the envenomation, such as animal bites. However, the new-PMS developed in this study included the unique characteristics of poisoning, such as the intent and route of poisoning. Third, we categorized the toxic substances into 8 categories comprising 44 speci c substances, as all the 44 substances could not be inputted to the multivariable logistic regression. The clinical severity of a poisoning can range from asymptomatic to life-threatening, depending on speci cs related to the toxin. For example, speci c toxic substances, such as paraquat, are known to have high mortality by themselves without the consideration of other predictors [26]. In this study, we considered paraquat as a separated category. Further study with using machine learning method may improve the prediction of mortality with consideration of speci c toxic substances. Lastly, the real observed mortalities in this study were as low as 2.6% for the derivation group and 1.6% for the validation group. There is a risk of overestimating/over tting the predictive performance of the model if the number of predictors is much larger than the number of outcome events [12].

Conclusion
Because outcome prediction systems for poisoning patients are rarely studied, we developed a single scoring system to accurately predict outcomes in poisoning patients with a wide range of demographics.
The new-PMS showed good performance in predicting the in-hospital mortality of acute poisoning patients, both in the derivation and validation groups. Because the risk of mortality sharply increased with the high risk-group of the new-PMS, early hemodynamic stabilization for acute poisoning patients at ED presentation may improve their clinical outcomes. The new-PMS will assist in clinical decision-making for patients with acute poisoning, thus improving their outcomes.

Declarations
Ethics approval and consent to participate This study was approved by the Institutional Review Board of the Korea University hospital (#IRB No. 2020AN0195). This study was a retrospective study based on the de-identi ed administrative database, so the informed consents were waived.

Consent for publication
Not applicable

Availability of data and materials
The datasets during and/or analysed during the current study available from the corresponding author on reasonable request.

Competing interests
The authors declare that they have no competing interests. Neither the entire paper nor any part of its contents have been published or accepted by another journal. The paper has not been submitted in its entirety, or in part, to any other journal.

Funding
This work was funded by the Korea University (K1925141). Kap Su Han received these funding. This funding source had no role in the study design, data collection, data analysis, data interpretation, or writing of the report Authors' contributions KSH, SJK, and SWL conceived and designed the study and wrote the manuscript. KSH managed and analyzed the data, including data quality control. LJS provided advice on statistics. All authors contributed substantially to the writing of the manuscript.   Table 2. Category of exposed of substances according to the class of the substance and the mortality index in the derivation group Category Name of substance  Base constant B was selected as the smallest regression coe cient in the model, which was 0.005.
The new-PMS was the sum of the point of each variable. The possible range of new-PMS was -20 to 3420 points.   Table 4. PMS, Poisoning Mortality Score b) Predicted mortality = 1/(1 + e -z ), z = -9.708 +0.005*newPMS Figure 1 Selection of study patients. Of 42568 study patients, 34352 (80.7%) and 8216 (19.3%) were enrolled in the derivation and validation groups, respectively. PMS, Poisoning Mortality Score.