a b s t r a c t

Objective: Acute pancreatitis is a common inflammatory disorder that may develop into severe AP (SAP), resulting in Life-threatening complications and even death. The purpose of this study was to explore two different machine learning models of multilayer perception-artificial neural network (MPL-ANN) and partial least squares-discrimination (PLS-DA) to diagnose and predict AP patients’ severity.

Methods: The MPL-ANN and PLS-DA models were established using candidate markers from 15 blood routine pa- rameters and five serum biochemical indexes of 133 mild acute pancreatitis (MAP) patients, 167 SAP (including 88 moderately SAP) patients, and 69 healthy controls (HCs). The independent parameters and combined model’s diagnostic efficiency in AP severity differentiation were analyzed using the Area Under the Receiver Operating Characteristic Curve .

Results: The Neutrophil to lymphocyte ratio is the most useful marker in 20 parameters for screening AP patients [AUC = 0.990, 95% confidence interval (CI): 0.984-0.997, sensitivity 94.3%, specificity 98.6%]. The MPL-ANN model based on six optimal parameters exhibited better diagno- stic and predict performance (AUC = 0.984, 95% CI: 0.960-1.00, sensitivity 92.7%, specificity 93.3%, accuracy 93.0%) than the PLS-DA model based on five optimal parameters (AUC = 0.912, 95% CI: 0.853-0.971, sensitivity 87.8%, specificity 84.4%, accuracy 84.8%) in discriminating MAP patients from SAP patients.

Conclusion: The results demonstrated that the MPL-ANN model based on routine blood and serum bio- chemical indexes provides a reliable and straightforward daily clinical practice tool to predict AP patients’ severity.

(C) 2021 Published by Elsevier Inc.

Abbreviations: AMY, amylase; ANN, artificial neural networks; APACHE, Acute Physiology and Chronic Health Evaluation; ARDS, acute respiratory distress syndrome; AUC, area under the ROC; CI, confidence intervals; CRP, C-reaction protein; GS, Glasgow severity; HCs, healthy controls; HGB, hemoglobin concentration; HTC, hematocrit; IQR, inter-quartile range; LPS, lipase; LYM, lymphocyte count; LYM-R, percentage of lympho- cytes; MAP, mild acute pancreatitis; MODS, Multiple organ dysfunction syndromes; MONO, Monocyte count; MONO-R, percentage of monocyte; MPL, multilayer perceptron; NEU, Neutrophil count; NEU-R, percentage of neutrophils; NLR, the neutrophil to lympho- cyte ratio; PAMY, ancreatic amylase; PCT, procalcitonin; PDW, platelet distribution width; PLS-DA, partial least squares-discrimination; PLT, platelet count; RBC, Red blood cell count; RDW-CV, RBC distribution width-coefficient of variation; RDW-SD, RBC distribution width-standard deviation; ROC, receiver operating characteristic curve; SAP, Severe acute pancreatitis; SD, standard deviation; SIRS, systemic inflammatory response syn- drome; VIP, variable importance; WBC, white blood cell.

* Corresponding author at: Department of Laboratory Medical, Affiliated Hospital of Southwest Medical University, 25 Taiping Street, Luzhou, Sichuan 646000, China.

E-mail address: [email protected] (J. Liu).

¹ These authors contributed equally to this work.

² These authors are co-corresponding authors.

1. Introduction

Acute pancreatitis (AP) is an inflammatory condition of the pancreas that frequently leads to Systemic Inflammatory Response Syndrome , multiple organ dysfunction syndromes (MODS), and even death without early intervention [1]. The global incidence of AP ranges from 5 to 30 cases per 100,000 individuals per year and is still rising in recent years [2]. Most AP cases are mild, but approximately 15-20% of AP pa- tients can develop severe AP (SAP) because of serious organ failure and pancreatic necrosis, resulting in mortality rates as high as 30% [3,4]. Therefore, the diagnosis and prediction of AP patients’ severity in the early stage can help guide therapeutic intervention to prevent the devel- opment of severe conditions and, ultimately, save patients’ lives [5].

Many studies focus on the early diagnosis and treatment of SAP based on clinical manifestations, laboratory tests, and imaging examination.

https://doi.org/10.1016/j.ajem.2021.01.044 0735-6757/(C) 2021 Published by Elsevier Inc.

However, it is difficult to predict the disease’s course because of AP’s complexity and varieties [6]. Although several traditional scoring systems can predict SAP [7,8], they are complicated and time-consuming [9]. On the contrary, a few routine blood and biochemical indexes have gained significant interest in recent years due to their wide availability, low cost, and ability to reveal patients’ conditions. The abnormal blood rou- tine examination results, such as the increased Neutrophil to lymphocyte ratio was associated with the SAP and precisely predicted AP patients’ severity [10,11]. A recent study also shows that an elevated serum C-reaction protein (CRP) level is a strong predictor of poor progno- sis in AP patients [12]. Besides, serum amylase (AMY) and lipase (LPS) are useful Blood parameters in diagnosing AP patients because of rela- tively high specificity and sensitivity and longer duration time above normal levels [13].

Although these indicators were demonstrated well with AP patients’ severity, a single marker may not comprehensively reflect patients’ multifactorial nature because of limited prediction power [14]. There- fore, some sophisticated statistical methods by combining multiple bio- markers are critically needed to diagnose and predict AP patients’ severity. In recent years, machine learning models such as the multi- layer perceptron artificial neural networks (MPL-ANN) and partial least squares-discrimination (PLS-DA) methods were widely used for the classification of diseases and prediction of clinical outcomes [15-17]. ANN is a non-linear machine learning method proposed as supplements to standard statistical models to explore underlying non-linear interac- tions of interconnected predictors and predict multifaceted biological events [18,19]. On the contrary, the linear PLS-DA models were used to identify potential biomarkers and have the potential to predict dis- eases’ severity rather than a yes/no binary outcome [20]. Therefore, the study aimed to explore both the MPL-ANN and PLS-DA models’ abil- ity to diagnose and predict AP patients’ severity based on blood routine and serum biochemical indicators.

1. Methods
   1. Patients and data collection

We retrospectively analyzed the medical records of 300 patients with AP admitted to the emergency department of the Affiliated Hospital of Southwest Medical University from 20 January 2017 to 21 November 2019. The diagnosis and disease severity classification of AP were accord- ing to the Revised Atlanta Classification 2012 [4]. There were 133 MAP pa- tients (MAP group) and 167 SAP (including 88 moderately SAP) patients (SAP group). Patients with the following conditions were excluded from the study: active malignancy, history of drug abuse, severe cardiovascular, respiratory, hepatic, renal, or mental diseases. Besides, 69 healthy subjects who underwent physical examination were enrolled as healthy controls (HCs) from 1 October 2019 to 15 November 2019. The screen criteria of HCs were as follows: (a) no history of pancreatic; (b) no history of bile duct diseases; (c) no noticeable abnormalities confirmed by physical ex- amination; and (d) not pregnant or lactating women. Demographic infor- mation of all participants, including gender, age, previous history (e.g., hypertension and diabetes), etiology (e.g., biliary, hyperlipemia), and other complications, was collected from electronic medical records. Only the first episode for each patient was included in our analysis. The study was conducted according to the principles of the Declaration of Hel- sinki. The ethics review board of the Affiliated Hospital of Southwest Med- ical University approved this study. Informed consent for an individual patient was not obtained since all data were retrieved retrospectively from the laboratory test information system without additional blood samples or Laboratory analysis. Laboratory data mainly included two parts: serum index of AMY, LPS, pancreatic amylase (PAMY), procalcitonin , CRP, and a series of routine blood parameters of white blood cell count, NLR, absolute neutrophil count (NEU), ab- solute lymphocyte count (LYM), absolute monocyte count (MONO), the percentage of neutrophils (NEU-R), the percentage of lymphocytes

(LYM-R), the percentage of monocyte (MONO-R), red blood cell count (RBC), the concentration of hemoglobin (HGB), hematocrit (HTC), RBC distribution width-standard deviation (RDW-SD), RBC distribution width-coefficient of variation (RDW-CV), absolute platelet count , and platelet distribution width . Laboratory data of the patients were recorded on admission and during the first 72 h of hospitalization.

• 1. Statistical analysis of the clinical data

Normally or near normally distributed variables are reported as mean with standard deviation (SD) and compared using Student’s t– test or variance analysis. Non-normally distributed data are expressed as median with inter-quartile range (IQR) and compared using the Kruskal-Wallis H test or Mann Whitney U test, as appropriate. Categor- ical variables were presented as frequency and percentage and com- pared using the chi-square test. We use the area under the receiver operating characteristic (ROC) curve (AUC) to determine each parameter’s optimal cut-off values with sensitivity and specificity in di- agnosing and predicting AP patients’ severity. The significance level was set at a P-value less than 0.05. Statistical analyses and the MPL-ANN model’s establishment were performed using the SPSS 25.0 statistical software (IBM, Hangzhou, China). The PLS-DA model’s establishment was performed using the SIMCA-13 software package (Umetrics, Kinnelon, NJ, United States).

• 1. MPL-ANN model design

The ANN model used in the current study was an MPL network conducted as a fully connected feed-forward network with a back- propagation training algorithm. We initially screened significantly in- creased indexes between the MAP group and the SAP group, and then the candidate markers with moderate diagnostic performance (AUC > 0.70) were further selected to build the MPL-ANN model. All the data were automated standardized (values minus mean divided by standard deviation) in the MPL-ANN model. After finishing the can- didate markers’ standardization of data, the AUC, total error, and the correct percent predicted of the MPL-ANN model were used to identify the optimal combination panel. Only the panel with both the relatively higher AUC, correct percent predicted, and relatively lower across- entropy error is the optimal combination panel. Finally, we used 70% of randomly selected samples as the training set (MAP group: n = 92, SAP group: n = 122) and the remaining 30% as the testing set (MAP group: n = 41, SAP group: n = 45) to assess the MPL-ANN model’s di- agnosis and prediction ability.

• 1. PLS-DA model design

In the study, the entire preprocessed dataset between MAP and SAP was arranged in the independent variable matrix to construct the PLS-DA model. The dependent variable matrix was categorical and contained artificial values of 0 or 1, corresponding to the AP severity (0 for MAP and 1 for SAP). Similar to MPL-ANN, seven standardized indicators (significant increase between the MAP group and the SAP group, and AUC > 0.70) were initially screened into the PLS-DA model, and the optimal combination panel was further established using candidate indexes with the variable importance in the projection (VIP) values exceeding 1.0 [21]. We used the R²X (measures the accu- mulative variance explained by the selected number of latent variables), R²Y (measures the goodness of fit), and the Q² (the k-fold cross- validated, Predictive ability of the model) to evaluate the PLS-DA model’s quality [22]. Data on the training and testing sets were consis- tent with those in the MPL-ANN model. We used the training set to de- velop the model in fitting and cross-validation, and the testing set to evaluate the model’s actual predictive ability.

1. Results
   1. Demographic, clinical characteristics, and severity parameters

We enrolled a total of 300 AP patients in the study with a mean age of (47.6 +- 13.4) years. 63.0% of AP patients were males, 21.3% with a history of hypertension, 22.3% with a history of diabetes mellitus, 36.3% with Biliary tract diseases, and 35.0% with the etiology of hyperli- pemia. Among all AP patients, there were 133 MAP patients and 167 SAP patients. There were no apparent differences in age, gender, history of hypertension and diabetes, or etiology between the MAP and SAP groups (P > 0.05). Detailed information on the baseline characteristics is presented in Table 1.

As shown in Table 2, the NLR, WBC, NEU, MONO, NEU-R, RDW-CV, and PDW levels were different among the MAP group, SAP group, and HCs comparing 15 blood routine indicators. These indexes were the

Table 1

Demographic data on clinical characteristics of the AP patients.

highest in SAP patients, followed by MAP patients, and the lowest in HCs patients (P < 0.05). Conversely, the LYM, LYM-R, and MONO-R levels were significantly lower in the SAP group than those in the MAP and HCs group (P < 0.05). However, there was no statistically significant difference among the three groups in PLT, HTC, RBC, and HGB levels (P > 0.05). The peripheral blood CRP level in the SAP group was signif- icantly higher than that in the MAP and HCs groups (P < 0.05). In terms of biochemical indicators, the serum concentrations of PCT, AMY, PAMY, and LPS in the SAP group were also significantly higher than those in the MAP and HCs groups (P < 0.05).

• 1. ROC analysis between the AP group and HCs

The ROC curve analysis was performed on the AP patients’ group (n = 300) and HCs (n = 69) to evaluate routine blood and serum bio- chemical indicators’ diagnostic value. The optimal cut-off levels and evaluative indexes of all markers with AUC >= 0.7 (Table 3). The ROC curve analysis showed that NLR, LYM-R, NEU-R, NEU, WBC, CRP, and LPS could serve as a screening index due to their high diagnostic accu- racy (AUC > 0.90). The AUC for LYM, PAMY, MONO, PCT, RDW-SD,

and AMY was between 0.70 and 0.80, suggesting a moderate diagnostic

Characteristics Total

(n = 300)

Age, mean (SD)	47.6 (13.4)	47.3 (13.8)	47.9 (13.2)	0.655
Male, n (%)	189 (63.0%)	91 (68.4%)	98 (58.7%)	0.083
History Hypertension, n (%)	64 (21.3%)	25(18.8%)	39(23.4%)	0.339
Diabetes, n (%)	67 (22.3%)	32(24.1%)	35(21.0%)	0.522

MAP

(n = 133)

SAP

(n = 167)

P-value

accuracy for AP. Among these parameters, the NLR had the highest AUC value of 0.990 (95% confidence intervals [CI], 0.984 to 0.997) with a cut- off value of 0.929, a sensitivity of 95.3%, and a specificity of 98.6%, so it was a reliable predictive indicator for distinguishing AP patients from HCs.

Etiology, n (%) Biliary	109 (36.3%)	50 (37.6%)	59 (35.3%)	0.685	3.3. ROC analysis between the MAP group and SAP group
Hyperlipemia Others	104 (35.0%) 87 (29.0%)	47 (35.3%) 36 (27.1%)	57 (34.1%) 51 (30.6%)	0.827 0.510	The ROC analysis showed the highest AUC for NEU-R (AUC = 0.778), followed by LYM-R, NLR, NEU, and WBC for differentiation between the
Complications, n (%) intestinal obstruction	47 (15.6%)	0 (0%)	47 (28.1%)	<0.001	MAP and SAP groups. In terms of biochemical indicators, the AUC values
pulmonary infection	42 (14.0%)	0 (0%)	42 (25.1%)	<0.001	of AMY, PAMY, and CRP for distinguishing SAP from MAP were 0.737,
MODS	19 (6.3%)	0 (0%)	19 (11.4%)	<0.001	0.720, and 0.706, respectively (Table 4). The AUC values of single indica-

ARDS 7 (2.3%) 0 (0%) 7 (4.2%) 0.017

electrolyte disturbances	15 (5.0%)	0 (0%)	15 (9.0%)	<0.001
Respiratory failure	13 (2.6%)	0 (0%)	13 (7.8%)	0.001

SIRS 6 (2.0%) 0 (0%) 6 (3.6%) 0.027

ARDS, acute respiratory distress syndrome; MAP, mild acute pancreatitis; MODS, multiple organ dysfunction syndromes; SAP, severe acute pancreatitis; SD, standard deviation; SIRS, systemic inflammatory response syndrome.

tors were generally insufficient for distinguishing SAP from the MAP.

• 1. The establishment of optimal MPL-ANN model and prediction

The ROC analysis showed that the individual variables (NLR, NEU-R, LYM-R, NEU, AMY, WBC, PAMY, CRP) provided limited ability to

Table 2

Comparing laboratory parameters among MAP group, SAP group, and HCs, [median (IQR)].

Variables	HCs (n = 69)	MAP (n = 133)	SAP (n = 167)	P-value
WBC, x109/L	5.90 (4.91-6.74)	11.3 (8.77-14.6)	14.9 (12.1-19.0)	<0.001
NEU, x109/L	3.28 (2.66-4.04)	9.30 (7.05-12.3)	13.2 (10.4-16.7)	<0.001
LYM, x109/L	1.93 (1.69-2.31)	1.17 (0.84-1.61)	0.88 (0.67-1.21)	<0.001
NLR	1.58 (1.20-2.20)	8.24 (4.53-12.6)	14.4 (9.94-21.2)	<0.001
MOMO, x109/L	0.33 (0.26-0.40)	0.54 (0.37-0.81)	0.66 (0.46-0.94)	<0.001
NEU-R, %	55.7 (49.8-63.3)	83.4 (75.3-87.3)	89.0 (86.0-91.4)	<0.001
LYM-R, %	35.1 (27.7-40.4)	9.90 (6.85-16.7)	6.10 (4.12-8.80)	<0.001
MONO-R, %	5.50 (4.75-6.45)	5.10 (3.72-6.85)	4.50 (3.20-5.60)	<0.001
RDW-SD, fl	41.9 (40.7-43.4)	44.4 (42.2-47.4)	45.8 (42.6-48.1)	<0.001
RDW-CV, %	12.7 (12.5-13.3)	13.1 (12.6-13.5)	13.4 (12.9-14.1)	<0.001
PDW, %	16.3 (16.1-16.6)	16.5 (16.0-16.9)	16.7 (16.3-17.0)	<0.001
CRP, mg/L	4.11 (3.17-4.82)	36.0(14.4-75.8)	78.0 (35.2-155.5)	<0.001
PCT, ng/mL	0.02 (0.02-0.49)	0.14 (0.05-0.41)	0.45 (0.12-1.65)	<0.001
AMY, U/L	129.5 (120.4-142.9)	191.0 (93.2-394.3)	634.4 (217.6-1200.3)	<0.001
PAMY, U/L	54.1 (43.9-59.1)	157.6 (71.8-340.8)	511.2 (182.6-1055.8)	<0.001
LPS, U/L	55.6 (37.3-77.5)	279.1 (123.9-648.5)	871.7 (254.5-1688.7)	<0.001
PLT, x109/L	202.0 (176.0-237.5)	196.0 (153.5-266.0)	191.0 (146.0-235.0)	0.090
HCT	0.44 (0.41-0.47)	0.42 (0.39-0.46)	0.43 (0.37-0.48)	0.203
RBC, x1012/L	4.67 (4.40-4.98)	4.60 (4.14-5.13)	4.71 (4.08-5.24)	0.621
HGB, g/L	141.0 (133.5-152.0)	144.0 (128.0-164.5)	145.0 (124.0-167.0)	0.748

AMY, amylase; CRP, C-reaction protein; HCs, healthy controls; HGB, hemoglobin concentration; HTC, hematocrit; LPS, lipase; LYM, lymphocyte count; LYM-R, percentage of lymphocytes; MAP, mild acute pancreatitis; MONO, monocyte count; MONO-R, percentage of monocyte; NEU, neutrophil count; NEU-R, percentage of neutrophils; NLR, the neutrophil to lymphocyte ratio; PAMY, pancreatic amylase; PCT, procalcitonin; PDW, platelet distribution width; PLT, platelet count; RBC, red blood cell count; RDW-CV, RBC distribution width-coefficient of var- iation; RDW-SD, RBC distribution width-standard deviation; SAP, severe acute pancreatitis; WBC, white blood cell.

Table 3

ROC analysis of laboratory parameters between the AP group and HCs (AUC > 0.70).

Parameters	AUC (95% CI)	Cut-off	Sensitivity	Specificity	P-value
NLR	0.990 (0.984-0.997)	0.929	0.953	0.986	<0.001
LYM-R	0.989 (0.983-0.997)	0.918	0.947	0.971	<0.001
NEU-R	0.989 (0.980-0.997)	0.924	0.943	0.971	<0.001
NEU	0.984 (0.974-0.995)	0.933	0.933	1.000	<0.001
WBC	0.965 (0.948-0.982)	0.875	0.933	0.942	<0.001
CRP	0.919 (0.890-0.948)	0.787	0.917	0.870	<0.001
LPS	0.911 (0.880-0.942)	0.706	0.807	0.899	<0.001
LYM	0.871 (0.834-0.907)	0.671	0.787	0.884	<0.001
PAMY	0.837 (0.795-0.878)	0.671	0.830	0.841	<0.001
MOMO	0.835 (0.794-0.875)	0.607	0.650	0.957	<0.001
PCT	0.747 (0.664-0.830)	0.583	0.887	0.696	<0.001
RDW-SD	0.744 (0.685-0.802)	0.437	0.640	0.797	<0.001
AMY	0.742 (0.694-0.790)	0.599	0.700	0.899	<0.001

AMY, amylase; AUC, area under the receiver operating characteristic curve; CRP, C-reac- tion protein; HCs, healthy controls; LPS, lipase; LYM, lymphocyte count; LYM-R, percent- age of lymphocytes; MONO, monocyte count; NEU, neutrophil count; NEU-R, percentage of neutrophils; NLR, the neutrophil to lymphocyte ratio; PAMY, pancreatic amylase; PCT, procalcitonin; RDW-SD, RBC distribution width-standard deviation; ROC, receiver operat- ing characteristic curve; WBC, white blood cell.

Table 4

ROC analysis of laboratory parameters between the MAP group and SAP group (AUC > 0.70).

Parameters	AUC	95% CI	P-value
NEU-R	0.778	0.725-0.831	<0.001
LYM-R	0.759	0.704-0.815	<0.001
NLR	0.758	0.702-0.813	<0.001
NEU	0.758	0.704-0.812	<0.001
WBC	0.731	0.674-0.787	<0.001
AMY	0.737	0.681-0.793	<0.001
P-AMY	0.720	0.662-0.778	<0.001
CRP	0.706	0.649-0.764	<0.001

AMY, amylase; AUC, area under the receiver operating characteristic curve; CRP, C-reac- tion protein; LYM-R, percentage of lymphocytes; MAP, mild acute pancreatitis; NEU, neu- trophil count; NEU-R, percentage of neutrophils; NLR, neutrophil to lymphocyte; PAMY, pancreatic amylase; ROC, receiver operating characteristic curve; SAP, severe acute pan- creatitis; WBC, White blood cell counts.

discriminate MAP from SAP (AUC < 0.80). Thus, we included these vari- ables into the MPL-ANN modeling process to find an optimal combina- tion of features for predicting AP severity. By comprehensively comparing the number and type of input variables and prediction perfor- mance (relatively higher AUC and correct percent predicted, lower cross- entropy error), the optimal combination of features needed six variables, including NEU-R, NEU, LYM-R, CRP, AMY, and PAMY (Table 5). The MPL- ANN yielded an accuracy of 93.0% on the training set and 89.7% on the testing set. The optimal cut-off value was 0.530 with an AUC of 0.984

(95% CI, 0.960 to 1.00), a sensitivity of 92.7%, and a specificity of 93.3%. The predicted pseudo-probability of both MAP and SAP groups showed an excellent predictive value in the MPL-ANN model (Fig. 1).

• 1. The establishment of optimal PLS-DA model and prediction

By comprehensively comparing the R²X, R²Y, Q² (Table 6), and the VIP value >1.0 (Fig. 2) in the PLS-DA, the optimal combination panel needed five variables, including the NEU-R, WBC, CRP, AMY, and PAMY. The PLS-DA model yielded an accuracy of 76.6% in the training set and 88.9% in the testing set. The optimal cut-off value was 0.460, with an AUC value of 0.912 (95% CI, 0.897 to 0.986), a sensitivity of 87.8%, and a specificity of 84.4%. The PLS-DA model’s score plots showed a good separation of both the MAP and the SAP groups (Fig. 3).

• 1. Comparison of the MPL-ANN model with the PLS-DA model

As shown in Table 7, the MPL-ANN model was superior to the PLS- DA model showing superior sensitivity (92.7% vs 87.8%), specificity (93.3% vs 84.4%), accuracy (93.0% vs 84.9%), Youden’s index (0.860 vs 0.722), and AUC (0.984 vs 0.912, P = 0.022). ROC curves of the MPL-

ANN and PLS-DA models based on routine blood and serum biochemical indexes are shown in Fig. 4.

1. Discussion

In the present study, we estimated the abilities of routine blood and serum biochemical indicators to diagnose and predict AP’s severity. We demonstrated that the NLR, as previous studies reported [10,23], is a handy index for screening AP patients. The NLR integrates two opposing and complementary components of the immune pathway and repre- sents the balance between inflammatory activating factor neutrophils and inflammatory regulatory factor lymphocytes [24]. Neutrophils can propagate inflammation and tissue destruction in AP via activation of a cascade of inflammatory cytokines, proteolytic enzymes, and oxygen free radicals [25], and corresponds with SIRS development and progres- sion to MODS [10]. On the contrary, the apoptosis of peripheral blood lymphocytes increased in AP due to Fas/FasL overexpression, causing a significantly increased NLR in blood routine analysis [26]. As the NLR can be easily calculated, the NLR value may serve as a useful marker to screen AP patients in daily clinical practice.

Our study further explored the predictive value of 15 blood routine parameters and five serum biochemical indexes for AP severity. Eight indicators were significantly different between the MAP and SAP groups, with the AUC ranging from 0.706 to 0.778, showing a moderate discriminatory ability. These blood routing indexes (NEU-R, LYM-R, NLR, NEU, WBC) and biochemical indicators (AMY, PAMY, and CRP) are inexpensive, conveniently available in clinical settings. CRP is an

Table 5

The establishment of the optimal MPL-ANN model.

Combination	AUC	cross-entropy error	correct percent predicted (%)
NEU-R + AMY + PAMY+CRP	0.919	106.9	82.7
NEU-R + LYM-R + AMY + PAMY+CRP	0.916	103.2	83.3
NEU-R + NLR + AMY + PAMY+CRP	0.926	101.4	85.7
NEU-R + WBC + AMY + PAMY+CRP	0.927	100.6	83.3
NEU-R + NEU + AMY + PAMY+CRP	0.934	95.0	85.0
NEU-R + NEU + NLR + AMY + PAMY+CRP	0.941	90.0	87.7
NEU-R + NEU + WBC + AMY + PAMY+CRP	0.930	100.0	87.0
NEU-R + NEU + LYM-R + AMY + PAMY+CRP	0.951	84.1	87.4
NEU-R + NEU + LYM-R + AMY + PAMY	0.904	113.5	80.0
NEU-R + NEU + LYM-R + NLR + AMY + PAMY+CRP	0.947	86.2	86.7
NEU-R + NEU + LYM-R + WBC + AMY + PAMY+CRP	0.949	85.6	88.0
NEU-R + NEU + LYM-R + WBC + NLR + AMY + PAMY+CRP	0.943	89.0	88.7

AMY, amylase; ANN, artificial neural networks; AUC, area under the receiver operating characteristic curve; CRP, C-reaction protein; LYM-R, percentage of lymphocytes; MPL, multilayer perceptron; NEU, neutrophil count; NEU-R, percentage of neutrophils; NLR, neutrophil to lymphocyte; PAMY, pancreatic amylase; WBC, white blood cell counts;

Fig. 1. Predicted pseudo-probability of the MPL-ANN model for predicting AP patients’ severity.

Table 6

The establishment of optimal PLS-DA model.

Combination	R2X	R2Y	Q2	Correct (%)
NEU-R + AMY + PAMY+CRP	0.80	0.41	0.40	81.3
NEU-R + LYM-R + AMY + PAMY+CRP	0.69	0.44	0.40	82.0
NEU-R + WBC + AMY + PAMY+CRP	0.68	0.46	0.44	85.3
NEU-R + NEU + AMY + PAMY+CRP	0.68	0.46	0.44	85.0
NEU-R + NEU + WBC + AMY + PAMY+CRP	0.60	0.46	0.44	83.2
NEU-R + NEU + LYM-R + AMY + PAMY+CRP	0.60	0.46	0.44	83.3
NEU-R + NEU + LYM-R + AMY + PAMY	0.64	0.34	0.32	75.0
NEU-R + NEU + LYM-R + WBC + AMY + PAMY	0.60	0.46	0.44	83.7

+CRP

AMY, amylase; CRP, C-reaction protein; LYM-R, percentage of lymphocytes; NEU, neutro- phil count; NEU-R, percentage of neutrophils; PAMY, pancreatic amylase; PLS-DA, partial least squares-discrimination; WBC, white blood cell counts.

acute-phase reactant released by the liver in response to increased in- terleukin (IL) levels (e.g., IL-1 and IL-6) and TNF-? released by activated macrophages. Khanna reported a cut-off level of 150 mg/L CRP within

the first 48 h of symptom onset is sensitive (80-86%) and specific (61-84%) for predicting SAP [27]. Meanwhile, the CRP level measured on the 3rd day of admission could also predict AP patients’ systemic complications [28]. Besides, Basak et al. found that the CRP levels can in- crease the Ranson score’s efficacy for predicting the severity of AP [29]. Although there is a 24-48 h latency period before CRP levels increase, limiting its utility as an early predictor of severity, CRP remains a useful predictor for the rise of levels [27,28].

AP’s pathogenesis is complicated, but inflammation and tissue ne- crosis occur with the release of Pancreatic enzymes, including AMY and LPS [11]. A literature review by Ismail et al. indicated that serum LPS offered a higher sensitivity than serum AMY in diagnosing AP [11]. Like Ismail et al.’s conclusions, our ROC analyses showed that LPS yielded moderate diagnostic accuracy than AMY in AP diagnosis. Moridani et al. concluded that the measurement of PAMY had improved the sensitivity and specificity in the diagnosis of AP [30], which was con- sistent with our results.

Recent studies showed that a series of indexes such as the WBC, NLR, LPS, AMY, and PCT, could help determine AP’s severity following diag- nosis [11,31,32]. Nevertheless, the present study found that a single index, either routine blood parameters or serum LPS, AMY, CRP, and PCT, fails to ultimately reveal AP’s severity because of lower AUC values (AUC < 0.80). Therefore, we used the MPL-ANN and PLS-DA models based on different combination panels to increase accuracy. Compared with the PLS-DA model, the MPL-ANN model exhibited a higher diagno- sis and prediction ability to distinguish SAP patients from the MAP pa- tients. Since ANNs work in a non-linear fashion, the models may better describe the interactions between health risk factors [15]. The ad- vantages of ANNs are that few prior assumptions or knowledge about data distributions are required; they can model complex non-linear re- lationships, and they are robust and tolerant of missing data and input errors. Several authors also have used the ANN techniques to develop predictive models for managing patients with AP with varying degrees of success. Mofidi et al. used ten clinical and laboratory variables to de- velop the ANN model from a retrospective review of 664 patients with AP [33]. The ANN model was significantly more accurate than both the Acute Physiology and Chronic Health Evaluation II and Glas- gow severity (GS) score at Predicting progression to a severe course in this patient population. The MPL-ANN model in our study also exhibited higher Predictive performance than a similarly predictive model devel- oped by six variables (duration of pain until arrival at the emergency de- partment, creatinine, hemoglobin, alanine aminotransferase, heart rate, and WBC count) [34].

PLS-DA model has been widely used in high-dimensional data anal- ysis, especially in metabolomics, to maximize group separation. Com- pared with the widely used ANN method and the logistic regression method [35], there are few studies concerning the PLS-DA model’s

Fig. 2. Variables with the variable importance (VIP) > 1.0 in the PLS-DA model.

Fig. 3. The PLS-DA prediction model (blue stars represent the MAP patients; red circles represent the SAP patients).

Table 7

Comparison of the MPL-ANN model and the PLS-DA model for predicting AP patients’ severity.

Variable	MPL-ANN model	PLS-DA model	t P-value
Combination Sensitivity, %	NEU-R + NEU + LYM-R + AMY + PAMY+CRP 92.7	NEU-R + WBC AMY + PAMY +CRP 87.8
Specificity, %	93.3	84.4
Accuracy, %	93.0	84.9
Youden’s	0.860	0.722
index AUC (95% CI)	0.984 (0.960-1.00)	0.912	2.299 0.022
		(0.853-0.971)

AMY, amylase; ANN, artificial neural networks; AUC, area under the receiver operating characteristic curve; CRP, C-reaction protein; LYM-R, percentage of lymphocytes; MPL, multilayer perceptron; NEU, neutrophil count; NEU-R, percentage of neutrophils; PAMY, pancreatic amylase; PLS-DA, partial least squares-discrimination; WBC, white blood cell.

diagnosis in AP. In practice, the relationship between the response and covariates may be very complicated; it may be even worse than useless to fit a non-linear model to a linear relationship sometimes. PLS-DA is another linear classification method that utilizes the knowledge of the group to identify discrimination among groups. In recent years, the PLS-DA model has been used in clinical management and disease prog- nostication, predicting patients’ survival with head and neck Squamous cell carcinoma, and predicting warfarin outcomes atrial fibrillation [17,36,37]. In the study, PLS-DA constructs and optimizes a linear classi- fication model based on the covariance between inflammation-based markers and the final diagnosis (MAP or SAP), with optimal discrimina- tion between patient groups. Based on the linear relationship principle in the PLS-DA model, the NEU-R and LYM-R both involved in the MPL- ANN model can not simultaneously appear because of their opposite composition relation in WBC count. Our analysis revealed that both the MAP and SAP groups were well discriminated from each other, whereas the diagnostic performance of the PLS-DA is inferior to the MPL-ANN model. This finding indicated that using a single model for summarizing data is often sub-optimal for training machine learning models, and significantly more accurate predictions can be achieved with the analysis of different principles-based approaches.

Our study has some limitations. There are a few missing values in the retrospective study, and these data were imputed by random numbers generated using the estimated model. We also fail to compare both the MPL-ANN and PLS-DA prediction models with the Grading stan- dards, such as Ranson and APACHE-II score, and other data information because of limited blood routing indexes and Biochemical parameters.

Fig. 4. ROC curves of the MPL-ANN and PLS-DA models based on routine blood and serum biochemical indexes.

Therefore, a multicenter prospective study with a larger patient cohort is needed to minimize possible biases and confounding factors to evalu- ate MPL-ANN and PLS-DA models’ significance.

1. Conclusion

In conclusion, the advantages of routine blood indexes and serum biochemical parameters are easy to obtain and dynamically monitor at the bedside. The MPL-ANN model, based on routine blood and serum biochemical indexes, provides a reliable and straightforward daily clinical practice tool to predict AP patients’ severity and help clinicians in early clinical interference strategies to prevent AP’s progress to SAP.

Declaration of Competing Interest

The authors have declared no conflicts of interest.

Acknowledgments

The work was supported by grants from the Science and Technology Department of Sichuan Province (Grant no. 2017TJPT0003) and the Health and Family Planning Commission of Sichuan Province (Grant no. 18PJ104).

References

1. Lankisch PG, Apte M, Banks PA. Acute pancreatitis. Lancet. 2015;386(9988):85-96. https://doi.org/10.1016/S0140-6736(14)60649-8.
2. Tee YS, Fang HY, Kuo IM, Lin YS, Huang SF, Yu MC. Serial evaluation of the SOFA score is reliable for predicting mortality in acute severe pancreatitis. Medicine (Bal- timore). 2018;97(7):e9654. https://doi.org/10.1097/MD.0000000000009654.
3. Habtezion A, Gukovskaya AS, Pandol SJ. Acute pancreatitis: a multifaceted set of or- ganelle and cellular interactions. Gastroenterology. 2019;156(7):1941-50. https:// doi.org/10.1053/j.gastro.2018.11.082.
4. Banks PA, Bollen TL, Dervenis C, Gooszen HG, Johnson CD, Sarr MG, et al. Acute pan- creatitis classification working group. Classification of acute pancreatitis-2012: revi- sion of the Atlanta classification and definitions by international consensus. Gut. 2013;62(1):102-11. https://doi.org/10.1136/gutjnl-2012-302779.
5. Lee WS, Huang JF, Chuang WL. Outcome assessment in acute pancreatitis patients. Kaohsiung J Med Sci. 2013;29(9):469-77. https://doi.org/10.1016/j.kjms.2012.10.007.
6. Zerem E. Treatment of severe acute pancreatitis and its complications. World J Gastroenterol. 2014;20(38):13879-92. https://doi.org/10.3748/wjg.v20.i38.13879.
7. Kim BG, Noh MH, Ryu CH, Nam HS, Woo SM, Ryu SH, et al. A comparison of the BISAP score and serum procalcitonin for predicting the severity of acute pancreatitis. Korean J Intern Med. 2013;28(3):322-9. https://doi.org/10.3904/kjim.2013.28.3.322.
8. Wan J, Shu W, He W, Zhu Y, Zhu Y, Zeng H, et al. Serum creatinine level and APACHE-II score within 24 h of admission are effective for predicting persistent organ failure in acute pancreatitis. Gastroenterol Res Pract. 2019;2019:8201096. https://doi.org/10.1155/2019/8201096.
9. Cho SK, Kim JW, Huh JH, Lee KJ. Atherogenic index of plasma is a potential bio- marker for severe acute pancreatitis: a prospective observational study. J Clin Med. 2020;9(9):2982. https://doi.org/10.3390/jcm9092982.
10. Han C, Zeng J, Lin R, Liu J, Qian W, Ding Z, et al. The utility of neutrophil to lympho- cyte ratio and fluid sequestration as an early predictor of severe acute pancreatitis. Sci Rep. 2017;7(1):10704. https://doi.org/10.1038/s41598-017-10516-6.
11. Huang L, Chen C, Yang L, Wan R, Hu G. neutrophil-to-lymphocyte ratio can specifi- cally predict the severity of hypertriglyceridemia-induced acute pancreatitis com- pared with white blood cell. J Clin Lab Anal. 2019;33(4):e22839. https://doi.org/ 10.1002/jcla.22839.
12. Almeida N, Fernandes A, Casela A. Predictors of severity and in-hospital mortality for acute pancreatitis: is there any role for C-reactive protein determination in the first 24h? GE Port J Gastroenterol. 2015;22(5):187-9. https://doi.org/10.1016/j.jpge. 2015.05.004.
13. Ismail OZ, Bhayana V. Lipase or amylase for the Diagnosis of acute pancreatitis? Clin Biochem. 2017;50(18):1275-80. https://doi.org/10.1016/j.clinbiochem.2017.07.003.
14. Park HS, In SG, Yoon HJ, Lee WJ, Woo SH, Kim D. Predictive values of neutrophil- lymphocyte ratio as an early indicator for severe acute pancreatitis in the emergency department patients. J Lab Physicians. 2019;11(3):259-64. https://doi.org/10.4103/ JLP.JLP_82_19.
15. Zakrzewski AC, Wisniewski MG, Williams HL, Berry JM. Artificial neural networks reveal individual differences in metacognitive monitoring of memory. PLoS One. 2019;14(7):e0220526. https://doi.org/10.1371/journal.pone.0220526.
16. Borzouei S, Soltanian AR. Application of an artificial neural network model for diag- nosing type 2 diabetes mellitus and determining the relative importance of risk fac- tors. Epidemiol Health. 2018;40:e2018007. https://doi.org/10.4178/epih.e2018007.
17. Zheng S, Zhong J, Chen Y, Ma Z, He H, Qiu W, et al. Metabolic profiling of plasma in ges- tational diabetes mellitus using liquid chromatography and Q-TOF mass spectrometry. Clin Lab. 2017;63(7):1045-55. https://doi.org/10.7754/Clin.Lab.2017.161110.
18. Shaabanpour Aghamaleki F, Mollashahi B, Nosrati M, Moradi A, Sheikhpour M, Movafagh A. Application of an artificial neural network in the diagnosis of chronic lymphocytic leukemia. Cureus. 2019;11(2):e4004. https://doi.org/10.7759/cureus. 4004.
19. Song WY, Zhang X, Zhang Q, Zhang PJ, Zhang R. clinical value evaluation of serum markers for early diagnosis of Colorectal cancer. World J Gastrointest Oncol. 2020; 12(2):219-27. https://doi.org/10.4251/wjgo.v12.i2.219.
20. Marzetti E, Landi F, Marini F, Cesari M, Buford TW, Manini TM, et al. Patterns of cir- culating inflammatory biomarkers in older persons with varying levels of physical performance: a partial least squares-discriminant analysis approach. Front Med (Lausanne). 2014;1:27. https://doi.org/10.3389/fmed.2014.00027.
21. Tong J, Zhang H, Zhang Y, Xiong B, Jiang L. Microbiome and Metabolome analyses of milk from dairy cows with subclinical Streptococcus agalactiae mastitis-potential bio- markers. Front Microbiol. 2019;10:2547. https://doi.org/10.3389/fmicb.2019.02547.
22. Kalogiouri NP, Aalizadeh R, Dasenaki ME, Thomaidis NS. Authentication of Greek PDO kalamata table olives: a novel non-target high resolution mass spec- trometric approach. Molecules. 2020;25(12):2919. https://doi.org/10.3390/ molecules25122919.
23. Qi X, Yang F, Huang H, Du Y, Chen Y, Wang M, et al. A reduced lymphocyte ratio as an early marker for predicting acute pancreatitis. Sci Rep. 2017;7:44087. https://doi. org/10.1038/srep44087.
24. Jeon TJ, Park JY. Clinical significance of the neutrophil-lymphocyte ratio as an early predictive marker for adverse outcomes in patients with acute pancreatitis. World J Gastroenterol. 2017;23(21):3883-9. https://doi.org/10.3748/wjg.v23.i21.3883.
25. Yimam M, Lee YC, Kim TW, Moore B, Jia Q. Analgesic and anti-inflammatory effect of UP3005, a botanical composition containing two standardized extracts of Uncaria gambir and Morus alba. Pharm Res. 2015;7(Suppl. 1):S39-46. https://doi.org/10. 4103/0974-8490.157995.
26. Pinhu L, Qin Y, Xiong B, You Y, Li J, Sooranna SR. Overexpression of Fas and FasL is associated with infectious complications and severity of experimental severe acute pancreatitis by promoting apoptosis of lymphocytes. Inflammation. 2014;37(4): 1202-12. https://doi.org/10.1007/s10753-014-9847-8.
27. Khanna AK, Meher S, Prakash S, Tiwary SK, Singh U, Srivastava A, et al. Comparison of Ranson, Glasgow, MOSS, SIRS, BISAP, APACHE-II, CTSI scores, IL-6, CRP, and procalcitonin in predicting severity, organ failure, pancreatic necrosis, and mortality in acute pancreatitis. HPB Surg. 2013;2013:1-10. https://doi.org/10.1155/2013/ 367581.
28. Fisic E, Poropat G, Bilic-Zulle L, Licul V, Stimac D. The role of IL-6, 8, and 10, sTNFr, CRP, and pancreatic elastase in the prediction of systemic complications in patients with acute pancreatitis. Gastroenterol Res Pract. 2013;2013:282645. https://doi.org/ 10.1155/2013/282645.
29. Basak F, Hasbahceci M, Sisik A, Acar A, Alimoglu O. Can C-reactive protein levels in- crease the accuracy of the Ranson score in predicting the severity and prognosis of acute pancreatitis? A prospective cohort study. Turk J Gastroenterol. 2017;28(3): 207-13. https://doi.org/10.5152/tjg.2017.16686.
30. Moridani MY, Bromberg IL. Lipase and pancreatic amylase versus total amylase as biomarkers of pancreatitis: an analytical investigation. Clin Biochem. 2003;36(1): 31-3. https://doi.org/10.1016/s0009-9120(02)00419-8.
31. Staubli SM, Oertli D, Nebiker CA. laboratory markers predicting severity of acute pancreatitis. Crit Rev Clin Lab Sci. 2015;52(6):273-83. https://doi.org/10.3109/ 10408363.2015.1051659.
32. Pezzilli R, Billi P, Miglioli M, Gullo L. Serum amylase and lipase concentrations and lipase/amylase ratio in assessment of etiology and severity of acute pancreatitis. Dig Dis Sci. 1993;38(7):1265-9. https://doi.org/10.1007/BF01296077.
33. Mofidi R, Duff MD, Madhavan KK, Garden OJ, Parks RW. Identification of severe acute pancreatitis using an artificial neural network. Surgery. 2007;141(1):59-66. https://doi.org/10.1016/j.surg.2006.07.022.
34. Andersson B, Andersson R, Ohlsson M, Nilsson J. Prediction of severe acute pancrea- titis at admission to hospital using artificial neural networks. Pancreatology. 2011;11 (3):328-35. https://doi.org/10.1159/000327903.
35. Bartosch-Harlid A, Andersson B, Aho U, Nilsson J, Andersson R. Artificial neural net- works in pancreatic disease. Br J Surg. 2008;95(7):817-26. https://doi.org/10.1002/ bjs.6239.
36. Cao W, Liu JN, Liu Z, Wang X, Han ZG, Ji T, et al. A three-lncRNA signature derived from the Atlas of ncRNA in cancer (TANRIC) database predicts the survival of pa- tients with head and neck squamous cell carcinoma. Oral Oncol. 2017;65:94-101. https://doi.org/10.1016/j.oraloncology.2016.12.017.
37. Bawadikji AA, Teh CH, Kader MABSA, Wahab MJBA, Sulaiman SAS, Ibrahim B. Plasma metabolites as predictors of warfarin outcome in atrial fibrillation. Am J Cardiovasc Drugs. 2020;20(2):169-77. https://doi.org/10.1007/s40256-019-00364-2.