Introduction
Acute pancreatitis (AP) is one of most common gastrointestinal diseases that requiring acute hospital admission [
1], with about 2,814,972 incident cases occurred each year in the globe [
2]. Most patients present with mild AP, which is self-limiting and often resolves within one week. Approximately 1/5 of patients develop moderate or severe AP, with necrosis of the pancreatic tissue and extrapancreatic organ dysfunction, and a substantial mortality rate of 20–40% [
3‐
6].
According to the revised Atlanta classification, the severity of AP can be defined as mild, moderately severe, or severe [
7]. Given the unpredictable course of AP, a plethora of clinical and biochemical scoring systems have been developed to predict the prognosis of AP, such as Acute Physiology and Chronic Health Evaluation II and Ranson’s score [
8,
9]. These scoring systems are based on a combination of vital signs, certain laboratory data, and radiographic findings, which are frequently cumbersome to calculate and thus limited their application in clinical practice [
9,
10].
Several individual serum tests have been studied as predictors of severe AP, including C-reactive protein (CRP), blood urea nitrogen (BUN), procalcitonin, and others [
10,
11]. Among them, high level of CRP was found to be the most promising marker that correlated with pancreatic necrosis and severe AP course [
12]. However, CRP levels may be influenced by liver disease [
13], and its levels peak at 72–96 h after symptom onset, which may limit its prognostic accuracy as an early indicator [
14,
15].
Hypocalcemia is one of the most common electrolyte abnormalities among AP [
16,
17]. It’s proposed that hypocalcemia might result from fat necrosis and the subsequent “calcium soaps” formation [
18]. Current research has revealed that the in-depth reason for the reduction in serum calcium is the suppression of cytomembrane Ca2 + ATP enzyme at the outbreak of AP, which results in an influx of ectocytic calcium ions and intracellular calcium overload [
19]. Since intracellular calcium overload is the central mechanism of acinar cell injury in AP [
19], theoretically low serum calcium level can reflect both the mechanism of pancreatic injury, and the consequence of peripancreatic fat tissue necrosis. In hospitalized patients, serum calcium is measured routinely and repeatedly, and each patient’s calcium trajectory provides a repository of longitudinal quantitative data. There is a paucity of report on using this longitudinal data for phenotyping and prognosticating patients with AP.
The aim of this study was to classify AP into novel sub-phenotypes based on their calcium trajectories. We hypothesized that each of these trajectory groups would have different physiological characteristics and clinical outcomes. Looking at calcium trends in addition to a point-specific calcium level might better predict prognosis.
Methods
Source of data
We conducted this retrospective study based on the sizeable critical care database of Medical Information Mart for Intensive Care (MIMIC)-IV (version 1.0) [
20]. The MIMIC-IV database is an integrated and comprehensive clinical dataset containing all the patients admitted to the ICUs of Beth Israel Deaconess Medical Center in Boston, MA, from 2008 to 2019. There were 524,520 distinct hospital admissions for 257,366 adult patients (aged 16 years or above) during the study period. Besides, we also searched for eligible patients from the old version of MIMIC-III between 2001 and 2007. One author (M.J) who obtained access to this database was responsible for data extraction. Since the study was an analysis of a third-party anonymized publicly available database, institutional review board (IRB) approval from our institution was exempted.
Participants
Patients with AP (ICD-9: 5770) were identified using the revised Atlanta definition [
7], which is based on the fulfilment of two of three features: (1) abdominal pain, (2) serum amylase (or lipase) of at least three times the upper limit of normal, or (3) findings consistent with AP on contrast-enhanced CT, MRI, or transabdominal ultrasound. In MIMIC database, total serum calcium (TSC) was mostly measured, thus in this study TSC was used for analysis. Only TSC measurements from Hour 0 (defined as time of admission) to Hour 24 were included. The TSC data from Hour 0 to Hour 24 were split into 1-hour blocks of time. If a patient had multiple TSC measures in that 1-hour block, the earliest TSC measurement was used. Patients who had at least two TSC measurements during 0–24 h were included.
Outcome definition
In-hospital mortality was the primary outcome. Other clinical outcomes such as ICU admission, length of hospitalization, antibiotic use, intubation rate, and renal replace treatment (RRT) were secondary outcomes.
Management of missing data
In this study, the continuous variables CRP and lactate had more than 50% of values missing, so they were excluded from the analysis. Most of the other continuous variables (e.g., albumin, bicarbonate, BUN) had less than 5% of values missing, the missing values were handled with single imputation in the analysis [
21].
Sensitivity analysis
To test the robustness of our findings, we also constructed trajectory models by excluding patients who died or were transferred before Hour 24, using 4-hour time block of TSC measurement, based on the TSC measurements in patients with calcium infusion during the first 24 h after admission, as well as using available serum ionized calcium. The consistency of the discriminability of these models was investigated.
Statistical analysis
Categorical variables are presented as number and proportions, continuous variables are summarized as the mean ± standard deviation (SD) or the median (interquark range [IQR]) according to the data distribution. Demographics, comorbidities, laboratory results, and processes of care (e.g., intubation, RRT) were compared between survivors and non-survivors. In addition, TSC measurements (initial, mean, maximum, minimum, and TSC variability, defined as the SD of TSC measurements during the first 24 h) were also compared between survivors and non-survivors. For these comparisons, categorical and continuous variables were tested for statistical significance using Chi-square tests, Fisher’s exact tests, or Wilcoxon rank-sum tests, as appropriate.
We used group-based trajectory modeling to identify sub-phenotypes based on TSC trajectories from the first 24 h of TSC data. Group-based trajectory modeling is a specialized mixture modeling method that is used to identify groups of individuals following similar trajectories for a particular variable [
22]. The Bayesian information criterion and average posterior probability were used for model selection, which ensure that a well-fitting model is chosen. We explored two-, three-, and four-group trajectories, with both linear and quadratic models investigated. Group-based trajectory modeling was conducted using the ‘traj’ package in Stata software (StataCorp).
Demographics, comorbidities, physiological characteristics, and clinical outcomes were compared between the identified trajectory groups with analysis of Chi-square tests and variance rank tests. Multivariable logistic regression analysis was performed to evaluate the independent association between TSC trajectories and in-hospital mortality (controlling for age, sex, race and comorbidities), using “stable normal-calcium” as the reference group. Two-tailed P values < 0.05 were considered statistically significant for all tests, and analyses were performed using Stata version 16.0 software.
Discussion
We reported a novel approach to identify sub-phenotypes in patients with AP on the basis of TSC trajectories. Through group-based trajectory modeling, we identified four groups of patients that follow distinct TSC trajectories. We found significant physiological differences between the trajectory groups, as well as considerable variation in in-hospital mortality, with the “Very low TSC, slow resolvers” having a mortality rate six times higher than the “Stable normal-calcium”. We found that patients with fluctuating high level TSC also showed a higher mortality. These findings have important significance for understanding the heterogeneity of AP and may inform future research to determine treatment strategy across different sub-phenotypes.
Recent evidence proved that the identification of biologically meaningful subgroups of a disease could lead to more personalized management strategies. For example, Bhavani and colleagues identified sub-phenotypes with different inflammatory responses in sepsis using temperature trajectories [
23], with the sub-phenotypes displaying different clinical outcomes. Similarly, Bhatraju and colleagues [
24] identified sub-phenotypes of acute kidney injury (AKI) based on the creatinine trajectory, with the finding that non-resolving patients had 68% higher mortality. AP is a late entry to the field of sub-phenotyping. Neyton and colleagues [
25] used unsupervised clustering of transcriptomic, proteomic, and metabolomic data to describe four sub-phenotypes of AP (have not yet been subject to peer review;
www.researchgate.net/publication/330852956). Although the results are promising, the findings are based on small sample size (54 cases) and it costly and complicated to use in practice. The simple, routine, and widespread laboratory test of serum calcium have great potential for AP phenotyping, because it is readily available, inexpensive, and have standardized reference ranges. The universal availability of longitudinal TSC trajectories could facilitate more resource-intensive diagnostics for the classification of clinically relevant AP sub-phenotypes.
It is well documented that point-specific hypocalcemia is significantly associated with AP severity, and hypocalcemia is one of the components of Ranson’s criteria to assess the severity of AP [
16,
26,
27]. Peng and colleagues have recently reported that serum calcium on admission is independently associated with persistent organ failure in AP and may serve as a potential prognostic factor [
28]. Notably, we discovered that transient hypocalcemia (TSC < 8.4 mg/dL) was more common in non-survivors compared with survivors (70.3% vs. 30.6%) in the 24 h from presentation. We also reported that TSC variability was higher in non-survivors than in survivors. Our study underscores the value of evaluating the trajectory of the TSC rather than a static test; and these data support the prognostic implication of dynamic TSC measurements and their benefit over using a single TSC measurement.
Importantly, we discovered four TSC trajectory groups: “Very low TSC, slow resolvers”; “Moderately low TSC”; “Stable normal-calcium”; and “Fluctuating high TSC”. Studies have elucidated the mechanisms of extracellular calcium entry and calcium-mediated acinar cell injury and death in AP [
19], so we hypothesize that the four groups might represent the different degree and duration of pancreatic parenchymal injury. Given the low mortality rate of the “Stable normal-calcium”, we assume that they may represent a transient calcium influx and quickly recovered sub-phenotype. For “Very low TSC, slow resolvers” and “Moderately low TSC” phenotypes, they may represent a relatively longer duration of calcium toxicity induced organelle dysfunction sub-phenotypes. For the three groups, serum calcium should be interpreted as the mediator of gallstones, alcohol, drugs or any other causes that trigger the pathological cellular pathways and organelle dysfunction in AP. For “Fluctuating high TSC” group, we hypothesize that hypercalcemia secondary to calcium infusion, malignancy (e.g., some from osteoclastic damage from metastatic cancer to bone), hyperparathyroidism or other disease might be the direct cause of AP.
Identifying the specific phenotypes could lead to a better understanding of the pathophysiology of AP and may inform new treatment targets. It’s well known that the mitochondrial dysfunction secondary to cellular calcium overload leads to ATP depletion, and ultimately results in acinar cell necrosis [
29,
30]. Based on the central importance of calcium concentration toxicity, ORAI1 channel inhibitors that prevent extracellular calcium enter into the acinar cells have been developed [
31,
32]. ORAI1 inhibitors have been demonstrated to prevent necrosis in mouse models of AP and human acinar cells, alleviating both the local and systemic extent of organ injury [
31]. Furthermore, TRO40303 was found to prevent ATP depletion caused by cytosolic calcium overload, which also has therapeutic potential for AP [
33]. The benefit of these novel strategies that targeting calcium overload in AP should be validated in future multicenter trial. This has shed light on the new treatment approach of AP, especially in severe AP that still has a high mortality rate up to now.
Our study has some limitations that should be acknowledged. First, The MIMIC database consists of data collected from electronic health records from 2001 to 2019; this has led to significant missing data for inflammatory markers, such as ESR and CRP. Thus, we couldn’t confirm the differential inflammatory basis for these hypothesized sub-phenotypes. In addition, the management of AP may have changed over this period of time, which might bring confounding effects on our analysis. Third, we did not control for the etiology of AP due to its unavailable in the database, and further study are needed to confirm whether these sub-phenotypes are consistent in AP caused by different etiology. Fourth, we could not remove the effect of calcium infusion on TSC trajectories. However, modelling based on AP patients who received calcium infusion within the first 24 h didn’t change the TSC trajectories in a significant way compared with patients without calcium infusion (Figs.
S1 and
S2). Furthermore, the TSC concentration might be influenced by the level of serum albumin. However, sensitivity analysis that modelling based on the serum ionized calcium (Fig.
S3) showed similar trajectories, demonstrating that the phenotyping was stable.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.