Clinical Application of Early Warning Scoring Based on BiLSTM-Attention in Emergency Obstetric Preexamination and Triage.

Maternity is a special category of population and the criteria for emergency prescreening cannot be directly applied to adults. Therefore, a set of criteria for grading maternal conditions should be established. In this paper, we have combined the semantic analysis technique of BiLSTM-Attention neural network and fuzzy defect risk assessment method, to develop a hybrid approach, to preprocess the text of emergency obstetric prescreening information. Furthermore, we have used word2vec to characterize the word embedding vector and highlight the features related to the degree of defects of emergency obstetric prescreening information through the attention mechanism and obtain the semantic feature vector of the warning information. BiLSTM-Attention neural network has the dual advantages of extracting bidirectional semantic information and giving weight to important judgment information which has effectively improved the semantic understanding accuracy. Experimental tests and application analysis show that the judgment model which is based on proposed method has accurately classified and graded the defects of emergency obstetric prescreening alerts. Additionally, the accuracy and microaverage value are used as evaluation indexes.

1. Introduction

The emergency prescreening triage criterion is a classification standard which is based on the degree of urgency and criticality of the patient's condition. It is a tool to assist triage staff in clinical triage [1]. At present, there is a lack of specific and operational emergency obstetric prescreening triage criteria in China. Since the implementation of the comprehensive two-child policy in 2016, the number of elderly and high-risk pregnant women has increased significantly. Thus, emergency medicine, as an important department for emergency care of critically ill mothers, is facing great challenges. Despite the rapid changes in maternal conditions, there are often signs before sudden changes in conditions, which are important for reducing maternal deaths and improving adverse outcomes if they are recognized early and treated in a timely manner [2-4]. Safe and effective emergency prescreening triage criteria can accurately identify patients with acute and critical conditions, ensure patient safety, and improve the efficiency of emergency care [5, 6]. The emergency medical professional committee of the Chinese nursing association, in collaboration with the quality control center of emergency medicine of Zhejiang Province, has conducted evidence-based research, nationwide surveys, retrospective big data analysis, and retrospective data analysis. Based on evidence-based research, nationwide survey, retrospective big data analysis, Delphi expert consultation, and nationwide application, we have developed simple, scientific, and quantifiable emergency prescreening and triage criteria of the emergency obstetric.

With the trend of digitalization and intelligent transformation of hospital industry, the awakening of massive monitoring and alert information resources has become an important initiative for the construction of multifunctional and highly resilient power grid in the dispatching department in the form of energy Internet [7]. The traditional data collection and monitoring system relies on the subjective experience of emergency obstetric prescreening personnel to determine the defect level of emergency obstetric prescreening information, which has problems of low efficiency and misclassification. In order to help the monitoring personnel to quickly and accurately grasp the upper window alarm information, realize the defect identification and risk analysis of emergency obstetric pretesting alarm information, and provide auxiliary decisions for the regulation of defective fault disposal, data mining of alarm information by using deep learning and NLP (natural semantic processing technology) improve the efficiency and accuracy of defect impact level grading and promote the manual BiLSTM-based defect risk analysis. Attention's intelligent technology is for defect risk warning of power grid alert information in a big data platform for emergency obstetric prescreening [8, 9].

In a study on the text of emergency obstetric pretest warning information, Jing Cao et al. [4] established a semantic framework-based defect text mining model to solve the problem of unstructured expressions that cannot be extracted accurately and verified the feasibility and effectiveness of the proposed mining technique by transformer defect text arithmetic. A method of identifying defective text information of power equipment which is based on dependent syntactic analysis is proposed to improve the accuracy of semantic analysis [10]. A BiLSTM-Attention-based text classification of emergency obstetric pretest faults is proposed to make intelligent judgments on fault defect appearances and improve the accuracy of defect classification recognition.

The classification, particularly for the abovementioned literature for equipment failure defects, is only focused on post hoc analysis and data lean control, while the real-time diagnosis of the upper window alarm information has the ability to provide auxiliary decision making and risk warning for obstetric emergency prescreening defect troubleshooting by healthcare providers [11].

In order to further improve ability of deep learning to identify defects and risk analysis in alert information text, we are proposing a natural semantic analysis-based alert information text defect risk warning method in this paper. In proposed model, we have combined the semantic analysis technique based on BiLSTM-Attention neural network [12] with the fuzzy defect risk assessment method. Preprocessing the alert information text using BiLSTM-Attention neural network has the dual advantages of extracting bidirectional semantic information and giving weight to important judgment information which effectively improves the semantic. Additionally, accuracy of comprehension is effectively improved if the proposed technique is applied. Experimental tests and application analysis show that the proposed judgment model has the capacity to accurately classify and grade the defects of emergency obstetric prescreening information, with accuracy and microaverage as evaluation indexes. Moreover, it has better classification effect than typical artificial intelligence algorithms and achieves defect risk warning of warning information.

The remaining parts of this manuscript are organized as given below. In subsequent section, emergency obstetric prescreening and triage classification index is described in detail which is followed by alarm information defect classification diagnosis procedure which is proposed in this paper. In Section 4, emergency obstetric prescreening risk assessment program is described in detail which is followed by a comprehensive discussion on the material and methods utilized in this paper. Experimental observations are presented in Section 6 which is followed by the concluding remarks and future directives.

2. Emergency Obstetric Prescreening and Triage Classification Index

The indicators, which are the same as the adult criteria, are described with reference to the adult criteria and interpretation [13] and those indicators which are specific to obstetrics are interpreted as follows. Sudden loss of consciousness, altered degree of consciousness, and disorders of consciousness are defined as varying degrees of reduction or loss of the patient's ability to awaken and perceive stimuli in themselves and their surroundings [14]. Sudden onset of loss of consciousness in patients is the most serious disorder of consciousness and is classified as grade 1. Unlike the criteria for adult emergency pretest triage, considering impact of disease on the fetus, sudden changes in the degree of consciousness, such as drowsiness, syncope, disorientation, etc., are also indicators of critical signs and require urgent treatment.

2.1. Severe Abdominal Pain in Pregnant Women

Using a descending ladder mentality, list potentially life-threatening conditions for both mother and baby with severe abdominal pain as the main complaint such as preeclampsia, uterine rupture, placental abruption, ectopic pregnancy rupture, and pregnancy combined with surgical emergencies which are pancreatitis, requiring immediate measures to stabilize hemodynamics or relieve fetal distress in utero.

2.2. Umbilical Cord Prolapses outside the Cervical OS

The clinical manifestation of cord prolapse is when the umbilical cord comes out of the cervical opening and descends into the vagina or is even exposed to the vulva after the rupture of the fetal membranes. If the umbilical cord prolapses, the umbilical cord is pressed between the fetal previa and the pelvis which causes fetal hypoxia and even complete loss of fetal heartbeat. Fetal death occurs if umbilical cord blood circulation is blocked for more than 7-8 minutes [15] and emergency treatment is required.

2.3. Signs of Impending Labor

The second stage of labor is the period of fetal delivery, when uterine contractions are frequent and can last up to 1 min specifically with an interval of only 1 to 2 min. When the fetal head descends and presses on the pelvic floor tissues, the mother has a reflexive feeling of defecation and involuntary downward forceful breath-holding movements, perineum is dilated and thinned, and anal sphincter is relaxed. The fetal head is exposed at the vaginal opening during contractions and then retracts back into the vagina during the interval of contractions which is called fetal head revelation [13]. Such clinical manifestations in pregnant women indicate imminent delivery and require urgent treatment.

2.4. Pregnancy Convulsions

A convulsion is a strong or rhythmic contraction of a part of the patient's limb, a side of the limb, or the whole body muscles, which may be accompanied by impaired consciousness [16]. The presence of convulsions in pregnant women should be considered as eclampsia. Pregnant women with eclampsia are prone to lip and tongue bites, falls, fractures, and other traumatic injuries during convulsions and condition progresses rapidly which is the main cause of maternal and infant mortality and requires emergency treatment [13].

3. Alarm Information Defect Classification Diagnosis

Alarm information defect classification diagnosis is divided into two subcategories: (i) semantic analysis algorithm flow and (ii) classification diagnosis training process. Both of these categories are described below.

3.1. Semantic Analysis Algorithm Flow

Based on the BiLSTM-Attention neural network method, it is able to explore the embedded semantic features in a large amount of alert information and performance index is superior in the tasks of defect text record recognition and defect degree classification [17]. Therefore, the classification process of emergency obstetric prescreening alert information is considered as the semantic recognition and classification of unstructured text data and the improved algorithm and evaluation indexes given in [17] are combined to propose a BiLSTM-Attention-based classification algorithm for grid alert information as shown in Figure 1.

Figure 1

Algorithm flow of alert message text classification.

3.2. Classification Diagnosis Training Process

Classification diagnosis training process is divided into four submodules which are described in detail below.

3.2.1. Text Preprocessing

SCADA system supervisory alarm messages are comprehensive statements describing faults, abnormalities, out-of-limits, changes in position, and notifications of substation equipment [18]. In this paper, we focus on abnormal alarm messages (including false fault signals), which have a significant impact on the grid operation and are diagnosed as defects (hereinafter collectively referred to as “alarm messages”). Therefore, we have to preprocess the unstructured textual alarm information (anomaly) on the SCADA system window and perform three steps of textual lemmatization, cleaning to identify the noun entities of power equipment, and standardization to obtain a set of defective keywords that match the alarm information, such as busbar switch, transformer, reclosing, protection device, merging unit, etc. The comparison results of preprocessing are shown in Table 1.

Table 1

Warning message text preprocessing.

Before pretreatment	Processing results
Low oil pressure of 110 kV bus tie switch of Daixi substation	“Daixi substation” “110 kV” “bus tie switch” “low oil pressure” “reclosing” “locking”
Low SF air pressure alarm of 110 kV kunlongyang switch in kunyang substation	“Kunyang substation” “110 kV” “kunlongyang” “switch” “SF” “low air pressure” “alarm”

3.2.2. Word Embedding Layer

The proposed word2vec model is a typical NLP model, which is an open source word vector computation tool introduced by Google in 2013 [12], including a preprocessing module and a shallow and two-layer neural network. The neural network layer outputs a distributed representation of the entity features of the power equipment as shown in Figure 2.

Figure 2

Word2vec model.

Combined with the word2vec word embedding method given in [12], formal representation of the alert message text is defined aswhere x denotes the ith word in the text S. For each word x, for example, the warning message “low oil pressure reclosing blocking at 11 kV female switch at Band Creek” in Figure 1, there is a word vector matrix: Wword ∈ R. In the training process, the word vector matrix is used to transform each word into a word vector representation before it is input into the BiLSTM layer, and finally the word vector representation of the noun of the power equipment entity for the given warning message is obtained, as shown in the following equation:

Thus, for a given alarm message text S will be transformed into a real matrix emb={e1, e2,…, e} with information about the degree of defects and fed into the next layer of the model.

3.2.3. BiLSTM Neural Network Layer

In the field of deep learning and natural semantic processing techniques, the BiLSTM method extracts bidirectional semantic information for long text information which make full use of the back-to-back reverse feature information [19]. Structure of the proposed BiLSTM model is shown in Figure 3.

Figure 3

Proposed BiLSTM model.

The proposed model uses a bidirectional LSTM network gating mechanism, mainly consisting of memory cell state c, input gate i, output gate o, etc.:where i is the input gate; c is the stored cell state; o is the output gate; x is the input at time t; h is the output value at time t − 1; σ is the Sigmoid activation function; W is the input weight matrix.

The vector representation matrix emb={e1, e2,…, e} of each word x of an alarm message text S is used as the hidden state output of the BiLSTM network gating and then stitched together by the position rule as the input of each cell at each moment. For a given alarm message text, “Low oil pressure reclosing blocking at the 110 kV female switch in the Belt Creek substation,” the following steps are performed:

Step 1 .

Preprocess alert information by word2vec word embedding and then enter the bidirectional LSTM network to extract semantic features.

Step 2 .

LSTM in positive direction first extracts positive features of the alarm information; for example, positive hidden state is obtained from “with stream change ⟶ 110 kV ⟶ busbar switch ⟶ low oil pressure ⟶ reclosing ⟶ latching,” that is, .

Step 3 .

Obtain the reverse feature extraction of the alarm information by LSTM in opposite direction; for example, reverse hidden state is obtained from “with stream change ← 110 kV ← busbar switch ← low oil pressure ← reclosing ← locking,” that is .

Step 4 .

After extracting the forward and reverse features, respectively, forward LSTM output of and reverse LSTM output of are spliced at each position to obtain a complete sequence of hidden states:

By splicing the forward and reverse hidden state sequences, the BiLSTM is able to extract the bidirectional semantic information, so that the model is capable of mining the important information in the whole sentence of the alert message text in positive and negative order in relation to the importance of the defect, ensuring that the defective features are not lost. Finally, set of hidden states output from the BiLSTM network layer is characterized as H : [h1, h2,…, h].

3.2.4. Attention Mechanism Layer

The attention mechanism is derived from the human visual attention to the focus area, and it can improve the accuracy of semantic understanding by simulating the human brain to focus on a specific area to obtain more effective value information.

In the attention mechanism model, the attention probability distribution of the defect importance feature vector in the alert message text output by the BiLSTM model at time n to the final state a:where N is the number of alert message text sequences; h′ is the complete sequence of hidden states of the alert message at time n; h′ is the jth vector of hidden states in the alert message at time n; a is the distribution of the attention of the hidden state h′ on h′ at time n. The final features of the attention-based alert message text at time n are calculated as follows:where the representational power of word embedding becomes stronger with increasing values of a and h, and the more information about the defective importance discriminating features of the alert information is obtained from the attention mechanism.

Finally, the Softmax calculation in the output layer results in a labeling dimension of 1 × 4 for the defect type of the alert message, and its probability distribution iswhere T is the number of defect type labels, T = 4, including general defects, important defects, urgent defects, and others. V is the weight of the alert information base generated during the training process.

For each electrical equipment entity noun input to the BiLSTM layer, there is a specific label in the training set, which is the category of electrical equipment entity; the probability of each output label is calculated by equation (7), and label with highest probability is the discriminator of defect degree of the alarm message.

In the processing process, a vector representation of the dimensions is obtained through normalization layer and probability distribution is obtained with the following labels, general defect (0.1), important defect (0.2), emergency defect (0.5), and others (0.2), text of the alarm message “110 kV female switch oil at the Belt Creek substation. Therefore, probability of the alarm message “Low voltage reclosing blocking” is “0.5,” and classification category is “emergency defect.”

4. Emergency Obstetric Prescreening Risk Assessment Program

The process of emergency obstetric prescreening defect risk warning system which is based on semantic analysis, preferably those proposed in this paper, is shown in Figure 4.

Figure 4

Flow of the defect risk warning system based on semantic analysis.

5. Proposed Model (Materials and Methods)

5.1. General Information

In China, hospitals have started to implement MEWS method of treatment in maternal prescreening triage in 2016 and 160 cases of pregnant women, preferably seen from January 2016–2018 months, were used as the observation group, while 160 cases of pregnant women treated in the conventional way in prescreening triage, from December 2013 to December 2015, were used as the control group. The enrolled subjects had complete clinical data, hospital stay of not less than one day, age 20–40 years, and signed informed consent willing to cooperate with the study, while excluding those who were not willing to cooperate with the study, referred to a higher level hospital, and had a hospital stay of less than one day. Control group: age 29.4 ± 3.2 years; pregnancy 2.9 ± 1.4 times; gestational weeks 33.3 ± 5.6 weeks; 102 cases of prim gravida and 58 cases of menstruating mothers. In the observation group: age 29.8 ± 3.4 years; gestation 2.5 ± 1.7 times; gestational week 33.9 ± 5.4 weeks; 105 prim gravid women and 55 menstruating women. There was no significant difference between the two maternal groups in the comparison of the previous information (P > 0.05), comparable.

5.2. Proposed Methodology

In control group, pregnant women were prescreened and triaged in the usual way and prescreened and triaged by medical and nursing staff in accordance with the “Pilot Guidelines for Grading the Condition of Emergency Patients (Draft for Comments)” (Family Planning Commission, 2011), which is divided into four levels according to the degree of condition, including endangered patients (level 1), critically ill patients (level 2), acute patients (level 3), and nonacute patients (level 4), with level 1 and level 2 pregnant women admitted to the level 1 and level 2 mothers are admitted to the ICU and resuscitated, level 3 mothers are sent to the closely observed treatment area, and level 4 mothers are sent to the treatment area to wait.

In the observation group, the mothers were prescreened and triaged by the MEWS method, and prescreening and triage medical personnel quickly assessed the vital signs and consciousness of mothers according to the MEWS method and graded their conditions according to the scoring criteria (total score is the sum of five items).

Grade 4: 0

Grade 3: 1–2

Grade 2: 3–5

Grade 1: >5 or single item 5 or <5 but with serious indicators.

According to the grading, different measures were taken to deal with the patient, in which the score of 5 was the critical value for identifying critical illness, grade 4 was sent to the general obstetric ward, grade 2-3 was sent to the obstetric care unit, grade 1 was sent to the ICU for targeted treatment, and if the mortality rate increased significantly if the score was >9, the patient was promptly sent to the resuscitation area for treatment of specific scoring table of MEWS as shown in Table 2.

Table 2

Specific scores of the modified early warning scoring system.

Column	3 points	2 points	1 points	0 points	1 points	2 points	3 points
Consciousness	—	—	—	Clear	Sound reacts	Respond to pain	Nothing
Heart rate (times/min)	—	≤ 40	41–50	51–100	101–110	111–129	≥ 130
Breathing (times/min)	—	9	-	9–14		21–29	≥ 30
Systolic blood pressure (mmHg)	70	71–80	81–100	101–199	—	≥ 200	—
Body temperature (°C)	—	35	—	35–38.4	—	≥ 38.5	—

5.3. Observed Indicators

The duration of maternal treatment, length of stay, occurrence of obstetric adverse events, and number of transfers to the ICU during hospitalization were recorded for statistical analysis in both groups.

5.4. Statistical Processing

The data were processed using SPSS 20.0. The number of obstetric adverse events and the number of cases transferred to ICU during hospitalization were expressed as percentages %, chi-square (x2) test, and time to save and length of hospitalization were expressed as means (), t-test, with P < 0.05 as a statistically significant difference.

To record and compare the duration of maternal rescue and hospitalization and transfer to ICU during hospitalization between the two groups, the observation group was significantly shorter than control group in terms of rescue time and hospitalization time while rate of transfer to ICU during hospitalization was significantly lower than that of control group, and the differences between the two groups were statistically significant (P < 0.05); see Table 3.

Table 3

Recording and comparing the duration of maternal rescue and hospitalization and transfer to ICU during hospitalization in both groups.

Group	Treatment time (min)	Length of stay (d)	Transfer rate to ICU during hospitalization (%)
Observation group (160)	29.49 ± 2.35	11.02 ± 5.84	2(1.25)
Control group (160)	38.27 ± 3.45	19.23 ± 6.95	23 (14.38)
t/s2	26.6052	11.4591	19.1349
P	P ≤ 0.001	P ≤ 0.001	P ≤ 0.001

To record and compare maternal adverse obstetric events in the two groups, total incidence of maternal obstetric adverse events in the observation group was significantly lower than that in the control group, and the difference between the two groups was statistically significant (P < 0.05), recording and comparing the maternal obstetric adverse events in the two groups [n(%)], observation group (160), bleeding 2 (1.25), shock 0 (0), emergency 2 (1.25), total 4 (2.50), control group (160), bleeding 12 (7.50), shock 3 (1.88), emergency 17 (10.62), total 32 (20.00), x2 24.5383,P0.0000.

6. Experimental Results and Discussion

The feature space of network learning on data was visualized in Figure 5 where each point stands for an image in the dataset. The color of point reflects the image class. It is evident that the samples in some classes were clustered while those in some other classes were separated. For example, Class-1 (dark-blue) and Class-9 (yellow) were close to each other, yet far away from other classes, because they belong to the same cluster.

Figure 5

The feature space of network learning process.

Further, Figure 6(a) provides the confusion matrix of the root nodes of level 1. Figure 6(c) shows clustering of the confusion classes, obtained by applying the spectrum copolymerization algorithm. This clustering is interpreted as a data hierarchy generated automatically from the data. Note that the optimization makes the samples in the same cluster closer and those in different clusters further away. In particular, the MLR enhanced the robustness of model classification. The separation in the feature space minimizes the connectivity loss function without affecting the classification performance, and in fact, this is shown to slightly improve performance.

Figure 6

Proposed model result on the collected dataset. (a) Confusion matrix of level 1/root node. (b) Influence of number of clusters on error. (c) Symmetric coclustering at K = 3.

Figure 6(b) shows the effect of varying the cluster size K on the error rate (due to misclassification). It was learned that the error minimized at K = 3.

7. Conclusion and Future Work

Maternity is a special category of population and the criteria for emergency prescreening and triage cannot be directly applied to adult criteria. Therefore, there is a need to establish a set of criteria for classifying maternal conditions according to their characteristics. In this paper, we have combined the semantic analysis technology based on BiLSTM-Attention neural network with the fuzzy defect risk assessment method, which has the dual advantages of extracting bidirectional semantic information and giving weight to important judgment information and has the capacity to effectively improve the semantic understanding accuracy. The judgment model, which is based on the proposed hybrid mechanism, has the ability to accurately identify and classify defects in emergency obstetric prescreening information. Experimental results have verified the operational superiority of the proposed hybrid model against well-known state-of-the-art techniques.

In future, we are eager to extend the proposed model effectiveness in other emergency areas particularly in the hospitals. Additionally, we will combine the proposed scheme with computer aided design to develop an automated system with maximum possible accuracy and precision ratio.