early prediction of organ failures in patients with acute ...segmentations were obtained through the...

Research ArticleEarly Prediction of Organ Failures in Patients with AcutePancreatitis Using Text Mining

Jiawei Luo1 Lan Lan 1 Dujiang Yang2 Shixin Huang3 Mengjiao Li1 Jin Yin1

Juan Xiao 4 and Xiaobo Zhou5

1West China Biomedical Big Data Center West China HospitalWest China School of Medicine Sichuan UniversityChengdu 610041 China2Department of Gastrointestinal Surgery West China HospitalWest China School of Medicine Sichuan UniversityChengdu 610041 China3School of Communication and Information Engineering Chongqing University of Posts and TelecommunicationsChongqing 400065 China4Department of Cardiovascular Medicine -e Affiliated Hospital of Southwest Medical University Luzhou 646000 China5School of Biomedical Informatics University of Texas Health Science Center at Houston Houston 77030 USA

Correspondence should be addressed to Lan Lan lanlscueducn and Juan Xiao 879214513qqcom

Received 8 December 2020 Revised 12 January 2021 Accepted 27 March 2021 Published 12 May 2021

Academic Editor Pengwei Wang

Copyright copy 2021 Jiawei Luo et al is is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

It is of great significance to establish an assessment model for organ failures in the early stage of admission in acute pancreatitis(AP) And the clinical notes are underutilized To predict organ failures for AP patients using early clinical notes in hospital earlytext features obtained from the pretrained Chinese Bidirectional Encoder Representations from Transformers model and at-tention-based LSTM were combined with early structured features (laboratory tests vital signs and demographic characteristics)to predict organ failures (respiratory cardiovascular and renal) in 12748 AP inpatients in West China Hospital SichuanUniversity from 2008 to 2018 e text plus structured features fusion model was used to predict organ failures compared to thebaseline model with only structured features e performance of the model with text features added is superior to the model thatonly includes structured features

1 Introduction

Organ failure is a serious complication of patients with acutepancreatitis (AP) Acute renal failure is one of the mostcommon causes of death in patients with severe AP [1]Acute renal failure in the setting of AP has been shown tohave a 10-fold increase in mortality in a study of 563 patients[2] Approximately one-third of severe pancreatitis patientsdevelop acute lung injury and acute respiratory distresssyndrome that account for 60 of all deaths within the firstweek [3] Patients with AP associated with congestive heartfailure (CHF) have significantly higher mortality in com-parison with those without CHF [4] erefore it is of greatsignificance to establish an assessment model for organfailures in the early stage of admission in AP

ree organ systems should be assessed to define organfailure respiratory cardiovascular and renal based on theclassification of APmdash2012 [5] In this study organ failuresincluded respiratory failure circulatory failure and renalfailure Previous studies [6 7] used a single indicator or alinear model of several indicators for predicting the risk oforgan failure within 12 hours of admissioneir sample sizewas small and their generalization ability was relativelyweak Additionally their methods did not make full use ofthe patientrsquos clinical manifestations and medical history

e electronic medical record (EMR) of the hospitalinformation system contains structured and unstructureddata which does not include bio-omics data [8ndash10] Inte-gration with clinical narrative would be highly useful be-cause rich information is buried in unstructured text It is of

HindawiScientific ProgrammingVolume 2021 Article ID 6683942 7 pageshttpsdoiorg10115520216683942

great value for the mining and utilization of unstructureddata For clinical notes (an example in West China HospitalSichuan University shown in Figure S1) the clinician andpatientrsquos history of present illness (HPI) may contain in-formation that is not extracted from structured data such asthe onset of AP recurrence time of AP symptoms thedescription of early symptoms before admission and cli-nicianrsquo speculations about certain complications orconditions

erefore we will make full use of clinical notes not justextracting some of them manually as features We will useAP patientrsquos HPI early laboratory tests early vital signs anddemographic data to assess the risk of organ failures between72 hours after admission and discharge Pretrained ChineseBidirectional Encoder Representations from Transformers(BERT) model [11] will be used to transform the sentenceinto a feature matrix and the characteristics of the sentencewill be captured by an attention-based long short-termmemory (LSTM) [12] so that a unified representation of thetext information can be obtained for all patients

2 Materials and Methods

21 Comparison to Recent Works on Medical RecordsNguyen et al [13] developed a deep neural net based onconvolutional neural networks named Deepr for repre-senting EMR and predicting unplanned readmission Deeprsequences the EMR into a ldquosentencerdquo or equivalently asequence of ldquowordsrdquo Each word represents a discrete objector event such as diagnosis procedure or any derived objectsuch as time-interval or hospital transfer eir ldquowordrdquo isdifferent from our ldquowordrdquo Our word is extracted from thetext sentence of clinical notes rather than combining dis-crete objects together

Soysal et al [14] developed a toolkit CLAMP for buildingcustomized clinical natural language processing pipelinesAlthough the CLAMP default pipeline achieved good per-formance on named entity recognition and conceptencoding it has limitations in handling other languages suchas Chinese

Krishnan and Kamath [15] used unstructured physiciannotes modeled using hybrid word embeddings to generatequality features which were used to train and build a deepneural network model to predict disease groups based onICD code Because there are too many types of diseasesinvolved it is not possible to obtain important character-istics specific to a certain disease

Yuwono et al [16] proposed a novel neural networknamed convolutional residual recurrent neural network thatlearns to diagnose acute appendicitis based on doctorsrsquo free-text emergency department notes and optional real-valuedfeatures (from the structured fields) without any featureengineering e effectiveness of their method in the di-agnosis or prediction of other diseases or adverse outcomesneeds to be further explored

Akbilgic et al [17] used a text mining approach (hybridmodel) on preoperative notes to obtain a text-based riskscore to predict death within 30 days of surgery in childrenwhich significantly improved the performance of C-statisticfrom 076 to 092 when text-based risk scores were includedin addition to structured data

erefore this study fully considers the disease char-acteristics of AP collects and analyzes relevant features asmuch as possible and uses the attention mechanism tooutput important words which leaves room for the treat-ment of AP We propose two strategies including thebaseline model of only structured features and text plusstructured features fusion model

22 OurMethods To implement this task (Figure 1) we willdo the following

(1) Extract text information of HPI laboratory testsvital signs and demographic data within the first72 h of admission

(2) Clean HPI including but not limited to punctuationmarks and meaningless symbol then make wordsegmentation

(3) By using the pretrained BERT model for Chineseconvert words into word vectors which are thenconverted into uniform representation feature vec-tors by attention-based LSTM

(4) Combine structured features (laboratory tests vitalsigns and demographic data) and unstructuredfeatures (HPI) and feed them into softmax functionto predict organ failures in AP

(5) Use minimized cross entropy to estimate theparameters

e input of the model is Xi Si Yi1113864 1113865 where i is the indexof the sample X is the structural variables such as vital signsand demographic S is the text description and Y is theoccurrence of a specific target event such as organ failure Si

can be expressed as a string of Chinese words For eachChinese word the corresponding Chinese word is firstconverted into the form of one-hot coding according to thedictionary and then each word is converted into a wordvector by using the word embedding matrix of the pre-trained BERT model as shown in formula (1) en theword vector of each word is input into the attention-basedLSTM network according to the order of sentences e lasthidden layer is taken as the input of the next layer which isrecorded as h(i)

u e structured information Xi is mapped tothe hidden layer vector h

(i)X through the neural network and

then h(i)X and h(i)

u are spliced to get the vector h(i)f Finally

h(i)f is used as the input of FCk and Y

(i)k is used as the output

for the specific target event e formula of attention-basedLSTM is defined as follows Based on the existing standard

2 Scientific Programming

LSTM network when the last hidden layer outputs atten-tion weight is added which is recorded as e(i)

m h(i)m represents

hidden layer vector mapped from the word vector of a word

h(i)X FC Xi( 1113857 (1)

u(i)m Wlowast onehot w

(i)m1113872 1113873

h(i)u ATTLSTM u

(i)1 u

(i)2 u

(i)m1113960 11139611113872 1113873

h(i)f concat h

(i)X h

(i)u1113960 11139611113872 1113873

Y(i)k FCk h

(i)f1113872 1113873

h(i)m LSTM u

(i)m1113872 1113873

h(i)u h

(i)last σ 1113944

M

m1e

(i)m lowast h

(i)m

⎛⎝ ⎞⎠

e(i)m

h(i)last ⊙ h

(i)m

1113936Mm1 h

(i)last ⊙ h

(i)m1113872 1113873

(2)

Different from other ways of using the pretrained BERTmodel we did not directly use the prediction task of organfailure as a subtask of the BERT model but used the em-bedding matrix of the BERT model to convert the sparseword vector represented by one-hot encoding to a dense realvector of degree 200 erefore in this study a sentence is amatrix which can also be regarded as a sequence of vectorsen we input this vector sequence into the attention-basedlong short-termmemorymodel in turn and take out the finalhidden layer with a length of 10 neurons as the featurerepresentation of the text information

23 Data e patients diagnosed with AP based on ICDcode from the hospital information system of West ChinaHospital Sichuan University from 2008 to 2018 were in-cluded ere were 15813 AP patients included initially Weextracted AP patientsrsquo demographic data laboratory testsand vital signs within 72 hours of admission HPI of APpatients also was extracted Respiratory failure was definedas the partial pressure of oxygen in blood gas analysis was

less than 60mmHg or the use of a ventilator Circulatoryfailure was defined as diastolic blood pressure was less than60mmHg or systolic blood pressure was less than 90mmHgand the use of vasoactive drugs Kidney failure was definedas creatinine was greater than 177 umolL e researchprotocol was approved by the ethics review board of WestChina Hospital Sichuan University and the need for in-formed consent was waived owing to the retrospectivenature of the study

24 Data Preprocessing e features with a missing rate ofmore than 20 in the structured data and patients without acomplete description of HPI were deleted A total of 12748patients were used for prediction finally We used micepackage of R software to perform multiple linear interpo-lation on structured data Before and after linear interpo-lation (see Figure S2) there is no statistical difference in thedistribution of the data (Pgt 005)

25Word Embedding Word embedding can be representedby a sparse one-hot vector or a dense real vector Wordvectors represented sparsely not only require large storagespace and computing resources but are also difficult to reflectthe correlation between words erefore since densevectors solve the above two problems at the same time densereal vectors are more commonly used In the classificationand prediction tasks of natural language processing that isthe issue of sequence to sequence the quality of the wordvector often has a great impact on the prediction ability ofthe model ere are currently a large number of wordvectors that have been trained based on continuous bag-of-words (CBOW) and skip-gram models [18] is studyadopted an analogical reasoning task on Chinese [19] Inorder to ensure the accuracy of Chinese medical wordsegmentation Tsinghua Open Chinese Lexicon [20] wasapplied

26 Training and Testing e data then was further dividedinto a training dataset and testing dataset according to the

Vital signs

Demographicinformation

Laboratory tests

Hospitalinformation

systemText

description ofthe patientrsquos

history ofpresent illness

Full connectionlayer

Attention-basedLSTM encoder

10

10

20

Fullconnection

layer

PretrainedBERT model

200

y1

y2

y3

Figure 1 Schematic diagram of prediction organ failures in AP using text mining

Scientific Programming 3

proportions of 70 and 30 Because of class imbalance andthe importance of positive identification in medicine weused 1 minus ri as the weight of the i-th class to make up for theproblems caused by a class imbalance in the cross-entropyloss function [21] We used a pretrained word vector net-work with a word vector length of 60 e batch size was setto 500 and the learning rate was set to 0005 e number ofhidden layer neurons of LSTM was set to 200 Models wereoptimized using a gradient descent approach e perfor-mance of training was an average of 1000 epochs In ourmodel mixed features including structured features and textfeatures were as input and combined attention mechanismcompared to the model that used only structured features asinput PyTorch framework was adopted to implement theexperiment on a Dell T640 GPU server

3 Results and Discussion

ere were 12748 AP patients with average age of 4758years and 609 male in this study Respiratory failurecirculatory failure and renal failure accounted for 142216 and 45 respectively Early vital signs and earlylaboratory tests are shown in Table 1

Table 2 contains the results of four models model 1 weuse structured features as input to predict three organ failuresrespectively model 2 we use text features as input to predictthree organ failures respectively model 3 we use mixed

features including structured features and text features asinput combined with the attention mechanism to predictthree organ failures respectively model 4 we use mixedfeatures as input and combine with the attention mechanismto predict three organ failures together as multitask From thetraining dataset perspective model 2 performed best throughthe prediction of organ failures From the testing datasetperspective the accuracies of model 3 and model 4 werehigher than model 1 and model 2 to predict respiratoryfailure circulatory failure and renal failure which shows thatadding text features can effectively help improve the accuracyof predicting organ failures in AP

In the classification task a good classifier should have agood effect on the judgment of each class so when eval-uating the effect of the classifier the recall and specificityshould be considered comprehensively e specificity ofmodel 3 is the highest to predict respiratory failure andcirculatory failure It shows that model 3 can effectivelylearn some characteristics of AP patients without organfailures Although the accuracy precision and specificity ofrenal failure are the highest in model 1 the recall of model 1is lower than that of model 3 andmodel 4e highest recallof respiratory failure comes from model 3 Based on thecomprehensive evaluation using the performance matrixthe performance of the model with text features added issuperior to the model that only includes structured featuresor text features e top 30 important Chinese word

Table 1 Baseline structured features and organ failures of AP patients

CharacteristicsDemographic characteristicsAge y mean (SD) 4758 (1501)Male n () 7764 (609)

Early vital signsRespiratory rate mg (μl)(hmiddotg) mean (SD) 2046 (222)Pulse ratemin mean (SD) 8570 (1519)

Early laboratory testsSerumASTALT mean (SD) 124 (117)Cholesterol mmolL mean (SD) 457 (231)High-density lipoprotein mmolL mean (SD) 087 (042)Low-density lipoprotein mmolL mean (SD) 191 (100)Amylase IUL mean (SD) 30487 (55537)Carbon dioxide combining power mmolL mean (SD) 2220 (358)Calcium mmolL mean (SD) 206 (024)Glutamyl transpeptidase IUL mean (SD) 13990 (21464)Albumin gL mean (SD) 3571 (598)Albuminglobulin mean (SD) 147 (036)White blood cell count 109L mean (SD) 1096 (513)Alanine aminotransferase IUL mean (SD) 6257 (11328)

Whole bloodRBC distribution width CV mean (SD) 1420 (148)RBC distribution width SD fL mean (SD) 4694 (527)Monocytes mean (SD) 550 (219)Monocytes 109L mean (SD) 057 (031)

Organ failuresRespiratory failure n () 1806 (142)Circulatory failure n () 2752 (216)Renal failure n () 579 (45)

SD standard deviation


segmentations were obtained through the attentionmechanism from model 4 which is shown in Table S1

Compared to the model of Krishnan and Kamath [15](Table 3) our proposed model performs better than it

Mentula et al [6] used 351 AP patientsrsquo some laboratorytests and APACHE II score within 12h of admission to predictorgan failure through logistic regression eir results showed082 of recall to predict respiratory failure and renal failurewithout testing which is lower than the predicted performanceon the training dataset of respiratory failure and renal failurewith text features added in this study (0850 and 0894 resp)Khanna et al [22] used 72 AP patientsrsquo various scores and a fewlaboratory tests to predict organ failure with a maximum recallof 1 using procalcitonin and a minimum recall of 0652 usingCTseverity index It is difficult to believe that only one score ora single indicator can be used to achieve such a good predictioneffect eir findings need to be further verified

Koutroumpakis et al [23] used 1612 AP patientsrsquothree laboratory tests and APACHE II score to predict

persistent organ failure e best recall was 0684 usingadmission APACHE II score and the lowest was 0249using admission creatinine Although these results seemreasonable they have not been tested and are lower thanthe results of the training dataset after adding text featuresin this study Hyland et al [24] developed a machinelearning method to predict circular failure in ICU pa-tients Although the predicted performance of AUC of094 was obtained the data used was routinely collectedstructured data In addition ICU data is monitored in realtime so the onset and duration of organ failure can beobserved and a time series model can be used However inroutine inpatients laboratory tests and vital signs aremonitored irregularly is is why this study did not useearlier data and time series models to predict persistentorgan failure

In addition to predicting single organ failure we also usedpredictive multitasking to output the results of three organfailures simultaneously after adding text features In the

Table 2 e performance of early predicting organ failures in AP using our proposed model

Model Performance Respiratory failure Circulatory failure Renal failureModel 1

Training

Accuracy 0727 (0033) 0668 (0034) 0786 (0048)Recall 0762 (0023) 0680 (0036) 0758 (0033)

Precision 0446 (0039) 0374 (0040) 0141 (0039)Specificity 0776 (0044) 0727 (0026) 0832 (0015)

Testing

Accuracy 0646 (0027) 0684 (0022) 0829 (0046)Recall 0649 (0034) 0664 (0023) 0681 (0036)


Model 2

Training

Accuracy 0744 (0039) 0820 (0013) 0871 (0007)Recall 0866 (0011) 0769 (0035) 0905 (0036)


Testing

Accuracy 0659 (0030) 0752 (0033) 0823 (0042)Recall 0619 (0025) 0723 (0020) 0519 (0022)


Model 3

Training

Accuracy 0659 (0008) 0820 (0029) 0803 (0027)Recall 0850 (0022) 0784 (0035) 0894 (0013)


Testing

Accuracy 0687 (0019) 0771 (0035) 0789 (0024)Recall 0664 (0014) 0732 (0031) 0708 (0017)


Model 4

Training

Accuracy 0772 (0047) 0784 (0036) 0838 (0032)Recall 0738 (0044) 0792 (0044) 0825 (0024)


Testing

Accuracy 0689 (0030) 0687 (0011) 0655 (0044)Recall 0543 (0010) 0688 (0016) 0854 (0026)


Model 1 input of structured features only Model 2 input of text features only Model 3 input of mixed features combined with attention mechanism Model4 input of mixed features and use of attention mechanism with multitask Standard deviation is in the bracket


multitask prediction model the loss function of each task canbe regarded as the constraint of other tasks e predictionresults of the three organ failures can be obtained in a shortertime which may meet different clinical needs We do not wantthemodel to be limited to the learning of the target task but canadapt to multiple task scenarios which can greatly increase thefunctional capability of the model (generalization)

Since it is a single-center study patients only comefromWest China Hospital Sichuan University which is alarge general hospital with 4300 beds in China Doctors inWest China Hospital Sichuan University may describethe patientrsquos current medical history in more detail thandoctors in other hospitals erefore the addition of textinformation will increase the predictive ability of themodel However in the same way when used in otherhospitals the text information of the current medicalhistory may be different from the text information of thisstudy so it should be careful when using the proposedmodel

4 Conclusions

We performed single-task and multitask prediction of organfailures in AP by the joint representation of structuredfeatures and text features According to our best knowledgethis is the first time to use clinical notes to predict organfailures in AP Our methods achieve superior accuracycompared to traditional techniques and uncover the un-derlying structure of the disease and intervention space

Data Availability

e datasets generated during andor analyzed during thecurrent study are available from the corresponding authoron reasonable request

Conflicts of Interest

e authors declare that they have no conflicts of interest

Table 3 e performance of early predicting organ failures in AP using the model of Krishnan and Kamath [15]


Training

Accuracy 0585 (0046) 0526 (0047) 0640 (0067)Recall 0624 (0033) 0537 (0050) 0616 (0046)


Testing

Accuracy 0506 (0038) 0546 (0032) 0682 (0064)Recall 0507 (0048) 0526 (0033) 0538 (0050)


Model 2

Training

Accuracy 0600 (0055) 0685 (0018) 0738 (0009)Recall 0732 (0016) 0626 (0049) 0762 (0050)


Testing

Accuracy 0518 (0042) 0610 (0046) 0678 (0058)Recall 0480 (0035) 0586 (0028) 0381 (0031)


Model 3

Training

Accuracy 0525 (0012) 0680 (0040) 0663 (0038)Recall 0712 (0031) 0641 (0050) 0758 (0018)


Testing

Accuracy 0550 (0026) 0628 (0050) 0650 (0034)Recall 0529 (0020) 0591 (0044) 0572 (0024)


Model 4

Training

Accuracy 0626 (0065) 0641 (0050) 0696 (0046)Recall 0593 (0062) 0646 (0062) 0686 (0034)


Testing

Accuracy 0548 (0042) 0552 (0015) 0509 (0061)Recall 0409 (0014) 0551 (0023) 0714 (0036)




Authorsrsquo Contributions

Jiawei Luo and Lan Lan contributed equally

Acknowledgments

is work was supported by the Postdoctoral ResearchProject West China Hospital Sichuan University (no2019HXBH039) the 1 3 5 Project for Disciplines of Ex-cellence West China Hospital Sichuan University (noZYJC18010) and the Center of Excellence-InternationalCollaboration Initiative Grant (no 139170052)

Supplementary Materials

e supplementary materials contain Figures S1 and S2 andTable S1 (Supplementary Materials)

References

[1] H Li Z Qian Z Liu X Liu X Han and H Kang ldquoRiskfactors and outcome of acute renal failure in patients withsevere acute pancreatitisrdquo Journal Of Critical Care vol 25no 2 pp 225ndash229 2010

[2] P Kes Z VuCICEviC I RatkoviC-GusiC A Fotivec andP Kes ldquoAcute renal failure complicating severe acute pan-creatitisrdquo Renal Failure vol 18 no 4 pp 621ndash628 1996

[3] C J Shields D C Winter and H P Redmond ldquoLung injuryin acute pancreatitis mechanisms prevention and therapyrdquoCurrent Opinion in Critical Care vol 8 no 2 pp 158ndash1632002

[4] H Mehta I Shah M Pahuja et al ldquoOutcomes of acutepancreatitis in patients with heart failure insights from thenationwide inpatient samplerdquo Journal of Cardiac Failurevol 25 no 8 pp S57ndashS58 2019

[5] P A Banks T L Bollen C Dervenis et al ldquoClassification ofacute pancreatitis--2012 revision of the Atlanta classificationand definitions by international consensusrdquo Gut vol 62no 1 pp 102ndash111 2013

[6] P Mentula M L Kylanpaa E Kemppainen et al ldquoEarlyprediction of organ failure by combined markers in patientswith acute pancreatitisrdquo British Journal of Surgery vol 92no 1 pp 68ndash75 2005

[7] M Sporek P Dumnicka A Gala-Bladzinska et al ldquoAngio-poietin-2 is an early indicator of acute pancreatic-renalsyndrome in patients with acute pancreatitisrdquo Mediators ofInflammation vol 2016 7 pages 2016

[8] C Liu J Chyr W Zhao et al ldquoGenome-wide association andmechanistic studies indicate that immune response contrib-utes to alzheimerrsquos disease developmentrdquo Frontiers in Ge-netics vol 9 p 410 2018

[9] Z Ji D Wu W Zhao et al ldquoSystemic modeling myeloma-osteoclast interactions under normoxichypoxic conditionusing a novel computational approachrdquo Scientific Reportsvol 5 no 1 Article ID 13291 2015

[10] Z Ji W Zhao H K Lin and X Zhou ldquoSystematically un-derstanding the immunity leading to CRPC progressionrdquoPLoS Computational Biology vol 15 no 9 Article IDe1007344 2019

[11] Y Cui W Che T Liu et al ldquoPre-training with whole wordmasking for Chinese BERTrdquo 2019 httpsarxivorgabs190608101

[12] D Bahdanau K Cho and Y Bengio ldquoNeural machinetranslation by jointly learning to align and translaterdquo 2016httpsarxivorgabs14090473

[13] P Nguyen T Tran N Wickramasinghe and S Venkateshldquo$mathtt Deepr$ a convolutional net for medical recordsrdquoIEEE Journal of Biomedical and Health Informatics vol 21no 1 pp 22ndash30 2017

[14] E Soysal J Wang M Jiang et al ldquoClamp-a toolkit for ef-ficiently building customized clinical natural language pro-cessing pipelinesrdquo Journal of -e American MedicalInformatics Association vol 25 no 3 pp 331ndash336 2018

[15] G S Krishnan and S Kamath ldquoHybrid text feature modelingfor disease group prediction using unstructured physiciannotesrdquo 2019 httpsarxivorgabs191111657

[16] S Yuwono H Ng and K Ngiam ldquoLearning from the ex-perience of doctors automated diagnosis of appendicitisbased on clinical notesrdquo in Proceedings of the 18th BioNLPWorkshop and Shared Task pp 11ndash19 Florence Italy August2019

[17] O Akbilgic R Homayouni K Heinrich M Langham andR Davis ldquoUnstructured text in EMR improves prediction ofdeath after surgery in childrenrdquo Informatics vol 6 pp 1ndash112019

[18] T Mikolov K Chen G Corrado and J Dean ldquoEfficientestimation of word representations in vector spacerdquo 2013httpsarxivorgabs13013781

[19] S Li Z Zhao R Hu W Li T Liu and X Du ldquoAnalogicalreasoning on Chinese morphological and semantic relationsrdquo2018 httpsarxivorgabs180506504

[20] A K Khanna S Meher S Prakash et al ldquoComparison ofranson glasgow moss sirs bisap Apache-II ctsi scores IL-6crp and procalcitonin in predicting severity organ failurepancreatic necrosis and mortality in acute pancreatitisrdquo HPBSurgery vol 2013 10 pages 2013

[21] E Koutroumpakis B U Wu O J Bakker et al ldquoAdmissionhematocrit and rise in blood urea nitrogen at 24 h outperformother laboratorymarkers in predicting persistent organ failureand pancreatic necrosis in acute pancreatitis a post hocanalysis of three large prospective databasesrdquo AmericanJournal Of Gastroenterology vol 110 no 12 pp 1707ndash17162015

[22] S L Hyland M Faltys M Huser et al ldquoEarly prediction ofcirculatory failure in the intensive care unit using machinelearningrdquo Nature Medicine vol 26 no 3 pp 364ndash373 2020

[23] S Han Y Zhang Y Ma et al ldquoTHUOCL Tsinghua openChinese lexiconrdquo Journal of Chinese Linguistics vol 20205 pages 2020

[24] Y S Aurelio G M de Almeida C L de Castro andA P Braga ldquoLearning from imbalanced data sets withweighted cross-entropy functionrdquo Neural Processing Lettersvol 50 no 2 pp 1937ndash1949 2019


great value for the mining and utilization of unstructureddata For clinical notes (an example in West China HospitalSichuan University shown in Figure S1) the clinician andpatientrsquos history of present illness (HPI) may contain in-formation that is not extracted from structured data such asthe onset of AP recurrence time of AP symptoms thedescription of early symptoms before admission and cli-nicianrsquo speculations about certain complications orconditions

erefore we will make full use of clinical notes not justextracting some of them manually as features We will useAP patientrsquos HPI early laboratory tests early vital signs anddemographic data to assess the risk of organ failures between72 hours after admission and discharge Pretrained ChineseBidirectional Encoder Representations from Transformers(BERT) model [11] will be used to transform the sentenceinto a feature matrix and the characteristics of the sentencewill be captured by an attention-based long short-termmemory (LSTM) [12] so that a unified representation of thetext information can be obtained for all patients

2 Materials and Methods

21 Comparison to Recent Works on Medical RecordsNguyen et al [13] developed a deep neural net based onconvolutional neural networks named Deepr for repre-senting EMR and predicting unplanned readmission Deeprsequences the EMR into a ldquosentencerdquo or equivalently asequence of ldquowordsrdquo Each word represents a discrete objector event such as diagnosis procedure or any derived objectsuch as time-interval or hospital transfer eir ldquowordrdquo isdifferent from our ldquowordrdquo Our word is extracted from thetext sentence of clinical notes rather than combining dis-crete objects together

Soysal et al [14] developed a toolkit CLAMP for buildingcustomized clinical natural language processing pipelinesAlthough the CLAMP default pipeline achieved good per-formance on named entity recognition and conceptencoding it has limitations in handling other languages suchas Chinese

Krishnan and Kamath [15] used unstructured physiciannotes modeled using hybrid word embeddings to generatequality features which were used to train and build a deepneural network model to predict disease groups based onICD code Because there are too many types of diseasesinvolved it is not possible to obtain important character-istics specific to a certain disease

Yuwono et al [16] proposed a novel neural networknamed convolutional residual recurrent neural network thatlearns to diagnose acute appendicitis based on doctorsrsquo free-text emergency department notes and optional real-valuedfeatures (from the structured fields) without any featureengineering e effectiveness of their method in the di-agnosis or prediction of other diseases or adverse outcomesneeds to be further explored

Akbilgic et al [17] used a text mining approach (hybridmodel) on preoperative notes to obtain a text-based riskscore to predict death within 30 days of surgery in childrenwhich significantly improved the performance of C-statisticfrom 076 to 092 when text-based risk scores were includedin addition to structured data

erefore this study fully considers the disease char-acteristics of AP collects and analyzes relevant features asmuch as possible and uses the attention mechanism tooutput important words which leaves room for the treat-ment of AP We propose two strategies including thebaseline model of only structured features and text plusstructured features fusion model

22 OurMethods To implement this task (Figure 1) we willdo the following

(1) Extract text information of HPI laboratory testsvital signs and demographic data within the first72 h of admission

(2) Clean HPI including but not limited to punctuationmarks and meaningless symbol then make wordsegmentation

(3) By using the pretrained BERT model for Chineseconvert words into word vectors which are thenconverted into uniform representation feature vec-tors by attention-based LSTM

(4) Combine structured features (laboratory tests vitalsigns and demographic data) and unstructuredfeatures (HPI) and feed them into softmax functionto predict organ failures in AP

(5) Use minimized cross entropy to estimate theparameters

e input of the model is Xi Si Yi1113864 1113865 where i is the indexof the sample X is the structural variables such as vital signsand demographic S is the text description and Y is theoccurrence of a specific target event such as organ failure Si

can be expressed as a string of Chinese words For eachChinese word the corresponding Chinese word is firstconverted into the form of one-hot coding according to thedictionary and then each word is converted into a wordvector by using the word embedding matrix of the pre-trained BERT model as shown in formula (1) en theword vector of each word is input into the attention-basedLSTM network according to the order of sentences e lasthidden layer is taken as the input of the next layer which isrecorded as h(i)

u e structured information Xi is mapped tothe hidden layer vector h

(i)X through the neural network and

then h(i)X and h(i)

u are spliced to get the vector h(i)f Finally

h(i)f is used as the input of FCk and Y

(i)k is used as the output

for the specific target event e formula of attention-basedLSTM is defined as follows Based on the existing standard



m h(i)m represents


h(i)X FC Xi( 1113857 (1)


(i)m1113872 1113873

h(i)u ATTLSTM u

(i)1 u

(i)2 u

(i)m1113960 11139611113872 1113873

h(i)f concat h

(i)X h

(i)u1113960 11139611113872 1113873

Y(i)k FCk h

(i)f1113872 1113873

h(i)m LSTM u

(i)m1113872 1113873

h(i)u h

(i)last σ 1113944

M

m1e

(i)m lowast h

(i)m

⎛⎝ ⎞⎠

e(i)m

h(i)last ⊙ h

(i)m

1113936Mm1 h

(i)last ⊙ h

(i)m1113872 1113873

(2)







Vital signs


Laboratory tests

Hospitalinformation

systemText





10

10

20

Fullconnection

layer


200

y1

y2

y3

























Training

Accuracy 0727 (0033) 0668 (0034) 0786 (0048)Recall 0762 (0023) 0680 (0036) 0758 (0033)


Testing

Accuracy 0646 (0027) 0684 (0022) 0829 (0046)Recall 0649 (0034) 0664 (0023) 0681 (0036)


Model 2

Training

Accuracy 0744 (0039) 0820 (0013) 0871 (0007)Recall 0866 (0011) 0769 (0035) 0905 (0036)


Testing

Accuracy 0659 (0030) 0752 (0033) 0823 (0042)Recall 0619 (0025) 0723 (0020) 0519 (0022)


Model 3

Training

Accuracy 0659 (0008) 0820 (0029) 0803 (0027)Recall 0850 (0022) 0784 (0035) 0894 (0013)


Testing

Accuracy 0687 (0019) 0771 (0035) 0789 (0024)Recall 0664 (0014) 0732 (0031) 0708 (0017)


Model 4

Training

Accuracy 0772 (0047) 0784 (0036) 0838 (0032)Recall 0738 (0044) 0792 (0044) 0825 (0024)


Testing

Accuracy 0689 (0030) 0687 (0011) 0655 (0044)Recall 0543 (0010) 0688 (0016) 0854 (0026)






4 Conclusions


Data Availability






Training

Accuracy 0585 (0046) 0526 (0047) 0640 (0067)Recall 0624 (0033) 0537 (0050) 0616 (0046)


Testing

Accuracy 0506 (0038) 0546 (0032) 0682 (0064)Recall 0507 (0048) 0526 (0033) 0538 (0050)


Model 2

Training

Accuracy 0600 (0055) 0685 (0018) 0738 (0009)Recall 0732 (0016) 0626 (0049) 0762 (0050)


Testing

Accuracy 0518 (0042) 0610 (0046) 0678 (0058)Recall 0480 (0035) 0586 (0028) 0381 (0031)


Model 3

Training

Accuracy 0525 (0012) 0680 (0040) 0663 (0038)Recall 0712 (0031) 0641 (0050) 0758 (0018)


Testing

Accuracy 0550 (0026) 0628 (0050) 0650 (0034)Recall 0529 (0020) 0591 (0044) 0572 (0024)


Model 4

Training

Accuracy 0626 (0065) 0641 (0050) 0696 (0046)Recall 0593 (0062) 0646 (0062) 0686 (0034)


Testing

Accuracy 0548 (0042) 0552 (0015) 0509 (0061)Recall 0409 (0014) 0551 (0023) 0714 (0036)






Acknowledgments




References



























m h(i)m represents


h(i)X FC Xi( 1113857 (1)


(i)m1113872 1113873

h(i)u ATTLSTM u

(i)1 u

(i)2 u

(i)m1113960 11139611113872 1113873

h(i)f concat h

(i)X h

(i)u1113960 11139611113872 1113873

Y(i)k FCk h

(i)f1113872 1113873

h(i)m LSTM u

(i)m1113872 1113873

h(i)u h

(i)last σ 1113944

M

m1e

(i)m lowast h

(i)m

⎛⎝ ⎞⎠

e(i)m

h(i)last ⊙ h

(i)m

1113936Mm1 h

(i)last ⊙ h

(i)m1113872 1113873

(2)







Vital signs


Laboratory tests

Hospitalinformation

systemText





10

10

20

Fullconnection

layer


200

y1

y2

y3

























Training

Accuracy 0727 (0033) 0668 (0034) 0786 (0048)Recall 0762 (0023) 0680 (0036) 0758 (0033)


Testing

Accuracy 0646 (0027) 0684 (0022) 0829 (0046)Recall 0649 (0034) 0664 (0023) 0681 (0036)


Model 2

Training

Accuracy 0744 (0039) 0820 (0013) 0871 (0007)Recall 0866 (0011) 0769 (0035) 0905 (0036)


Testing

Accuracy 0659 (0030) 0752 (0033) 0823 (0042)Recall 0619 (0025) 0723 (0020) 0519 (0022)


Model 3

Training

Accuracy 0659 (0008) 0820 (0029) 0803 (0027)Recall 0850 (0022) 0784 (0035) 0894 (0013)


Testing

Accuracy 0687 (0019) 0771 (0035) 0789 (0024)Recall 0664 (0014) 0732 (0031) 0708 (0017)


Model 4

Training

Accuracy 0772 (0047) 0784 (0036) 0838 (0032)Recall 0738 (0044) 0792 (0044) 0825 (0024)


Testing

Accuracy 0689 (0030) 0687 (0011) 0655 (0044)Recall 0543 (0010) 0688 (0016) 0854 (0026)






4 Conclusions


Data Availability






Training

Accuracy 0585 (0046) 0526 (0047) 0640 (0067)Recall 0624 (0033) 0537 (0050) 0616 (0046)


Testing

Accuracy 0506 (0038) 0546 (0032) 0682 (0064)Recall 0507 (0048) 0526 (0033) 0538 (0050)


Model 2

Training

Accuracy 0600 (0055) 0685 (0018) 0738 (0009)Recall 0732 (0016) 0626 (0049) 0762 (0050)


Testing

Accuracy 0518 (0042) 0610 (0046) 0678 (0058)Recall 0480 (0035) 0586 (0028) 0381 (0031)


Model 3

Training

Accuracy 0525 (0012) 0680 (0040) 0663 (0038)Recall 0712 (0031) 0641 (0050) 0758 (0018)


Testing

Accuracy 0550 (0026) 0628 (0050) 0650 (0034)Recall 0529 (0020) 0591 (0044) 0572 (0024)


Model 4

Training

Accuracy 0626 (0065) 0641 (0050) 0696 (0046)Recall 0593 (0062) 0646 (0062) 0686 (0034)


Testing

Accuracy 0548 (0042) 0552 (0015) 0509 (0061)Recall 0409 (0014) 0551 (0023) 0714 (0036)






Acknowledgments




References
















































Training

Accuracy 0727 (0033) 0668 (0034) 0786 (0048)Recall 0762 (0023) 0680 (0036) 0758 (0033)


Testing

Accuracy 0646 (0027) 0684 (0022) 0829 (0046)Recall 0649 (0034) 0664 (0023) 0681 (0036)


Model 2

Training

Accuracy 0744 (0039) 0820 (0013) 0871 (0007)Recall 0866 (0011) 0769 (0035) 0905 (0036)


Testing

Accuracy 0659 (0030) 0752 (0033) 0823 (0042)Recall 0619 (0025) 0723 (0020) 0519 (0022)


Model 3

Training

Accuracy 0659 (0008) 0820 (0029) 0803 (0027)Recall 0850 (0022) 0784 (0035) 0894 (0013)


Testing

Accuracy 0687 (0019) 0771 (0035) 0789 (0024)Recall 0664 (0014) 0732 (0031) 0708 (0017)


Model 4

Training

Accuracy 0772 (0047) 0784 (0036) 0838 (0032)Recall 0738 (0044) 0792 (0044) 0825 (0024)


Testing

Accuracy 0689 (0030) 0687 (0011) 0655 (0044)Recall 0543 (0010) 0688 (0016) 0854 (0026)






4 Conclusions


Data Availability






Training

Accuracy 0585 (0046) 0526 (0047) 0640 (0067)Recall 0624 (0033) 0537 (0050) 0616 (0046)


Testing

Accuracy 0506 (0038) 0546 (0032) 0682 (0064)Recall 0507 (0048) 0526 (0033) 0538 (0050)


Model 2

Training

Accuracy 0600 (0055) 0685 (0018) 0738 (0009)Recall 0732 (0016) 0626 (0049) 0762 (0050)


Testing

Accuracy 0518 (0042) 0610 (0046) 0678 (0058)Recall 0480 (0035) 0586 (0028) 0381 (0031)


Model 3

Training

Accuracy 0525 (0012) 0680 (0040) 0663 (0038)Recall 0712 (0031) 0641 (0050) 0758 (0018)


Testing

Accuracy 0550 (0026) 0628 (0050) 0650 (0034)Recall 0529 (0020) 0591 (0044) 0572 (0024)


Model 4

Training

Accuracy 0626 (0065) 0641 (0050) 0696 (0046)Recall 0593 (0062) 0646 (0062) 0686 (0034)


Testing

Accuracy 0548 (0042) 0552 (0015) 0509 (0061)Recall 0409 (0014) 0551 (0023) 0714 (0036)






Acknowledgments




References




























4 Conclusions


Data Availability






Training

Accuracy 0585 (0046) 0526 (0047) 0640 (0067)Recall 0624 (0033) 0537 (0050) 0616 (0046)


Testing

Accuracy 0506 (0038) 0546 (0032) 0682 (0064)Recall 0507 (0048) 0526 (0033) 0538 (0050)


Model 2

Training

Accuracy 0600 (0055) 0685 (0018) 0738 (0009)Recall 0732 (0016) 0626 (0049) 0762 (0050)


Testing

Accuracy 0518 (0042) 0610 (0046) 0678 (0058)Recall 0480 (0035) 0586 (0028) 0381 (0031)


Model 3

Training

Accuracy 0525 (0012) 0680 (0040) 0663 (0038)Recall 0712 (0031) 0641 (0050) 0758 (0018)


Testing

Accuracy 0550 (0026) 0628 (0050) 0650 (0034)Recall 0529 (0020) 0591 (0044) 0572 (0024)


Model 4

Training

Accuracy 0626 (0065) 0641 (0050) 0696 (0046)Recall 0593 (0062) 0646 (0062) 0686 (0034)


Testing

Accuracy 0548 (0042) 0552 (0015) 0509 (0061)Recall 0409 (0014) 0551 (0023) 0714 (0036)






Acknowledgments




References




























Acknowledgments




References


























early prediction of organ failures in patients with acute ...segmentations were obtained through the...

Documents