AKIHelper: Acute Kidney Injury Diagnostic Tool Using KDIGO Guideline Approach Issariya Uboltham

Nakornthip Prompoon

Department of Computer Engineering Faculty of Engineering, Chulalongkorn University, Bangkok 10330, Thailand [email protected]

Department of Computer Engineering Faculty of Engineering, Chulalongkorn University, Bangkok 10330, Thailand [email protected]

Abstract— Acute Kidney Injury (AKI) is common and harmful disorder in hospitalized patients. It is associated with poor outcomes such as a decrease chance of survival, longer hospital stays and an increase progression of chronic kidney disease. To diagnosis AKI, the KDIGO clinical practice guideline has been published for providing standardized criteria of AKI definition and the recommendation in medical pathway. Moreover, early detection of AKI in patient at risk can also improve the outcomes. This paper presents an approach to assist the doctor in diagnosis and decision making process. First, the risk factors of AKI were identified using data mining approach based on Decision Tree classification technique. Simple Cart and J48 were selected as the algorithms for this process. Second, a concept of tool requirements and design named “AKIHelper” is presented. This tool is created based on KDIGO guideline which is expected to use for diagnosis and staging severity of AKI. Keywords— Acute Kidney Injury; AKI; KDIGO; E-Health; Healthcare; Risk Factors; Diagnostic Tool; Data Mining; Decision Tree; Classification

I.

INTRODUCTION

Acute Kidney Injury (AKI) is a common disorder with a rapid reduction of kidney function over hours, days or weeks. The result of the treatment is unsatisfied and the cause is unclear, this is accounted for more than 10% of all hospitalized patients and more than 2/3 of all Intensive Care Unit (ICU) admissions [1]. Besides this, AKI leads to a consequence of other acute illness, such as Fluid Overload, Electrolyte Abnormalities, Impaired Innate Immunity and Progression of Chronic Kidney Disease (CKD). These can also cause the extension of hospital stay, increasing mortality rate and higher cost of care. To date, serum creatinine and urine output criteria have been developed and used to define the development of AKI. Level rise of serum creatinine compared with the baseline and duration of oliguria can be used to define the stage of disease. Early detection of AKI in patients at risk can be helpful in improving outcomes of the treatment. Unfortunately, the diagnosis and monitoring are often neglected or missed during hospital/ICU admission. Moreover, the measurement and documentation of creatinine level or urine output are either

978-1-5090-0806-3/16/$31.00 copyright 2016 IEEE ICIS 2016, June 26-29, 2016, Okayama, Japan

Wirichada Pan-ngum Mahidol-Oxford Tropical Medicine Research Unit (MORU) Faculty of Tropical Medicine, Mahidol University Bangkok 10400, Thailand [email protected]

inaccurate or delayed [2]. These explain why early detection of AKI and risk identification are often failed. The clinical practice guidelines for defining AKI have been published in various definitions such as RIFLE, AKIN (Acute Kidney Injury Network) and KDIGO (Kidney Disease Improving Global Outcomes). However, the incidence rate of AKI is still increasing because it is not a single disease. The progression of AKI comes from multiple clinical conditions that make the diagnostic of this disease become more complex. Furthermore, the implementation of these guidelines is not successful as it should be. This is because of the difference in viewpoints from multidisciplinary stakeholders, barriers in knowledge and a lack of understanding how to apply these guidelines in clinical practice [3]. This paper proposes an approach to identify the risk factors of AKI using data mining techniques. The result will be used for providing information in case of monitoring and caring of focused patients. Furthermore, the paper presents an idea to develop tool for staging AKI severity based on KDIGO guideline named “AKIHelper” that AKI risk factors retrieved from data mining process is also planned to be a part of it. This tool is hoped to systematically identify stage of disease, assist the clinical decision making processes, improve the outcome or clinical care practice and reduce hospital mortality rate as well as medical expenses. The rest of the paper is structured as follows. In Section II, background of KDIGO guideline and data mining in healthcare are presented. Section III briefly reviews related works of AKI. Section IV describes the proposed approach to identify the risk factors of AKI and presents a concept to develop tool based on KDIGO guideline, including evaluation. Section V discusses the limitation of the approach. Finally, section VI provides conclusion and plan for future work. II.

BACKGROUND

A. KDIGO Guideline for AKI KDIGO (Kidney Disease Improving Global Outcomes) was established in 2003 as non-profit global organization managed by the National Kidney Foundation. The objectives of KDIGO are to develop and implement evidence-based

clinical practice guideline in kidney disease. This organization has Global Network members made up of volunteers from every part of the globe and also has a respected record of achievement and accomplishment to improve the care and outcomes for patients worldwide. For AKI, KDIGO provides the guideline for clinical practice with recommendations [4] to assist decision making, caring for patient at risk and for selecting treatment to improve patient survival and preserve or recover kidney function. B. Data Mining in Healthcare Nowadays, data mining becomes necessity for healthcare industry in many aspects e.g. predicting disease, diagnosis treatment planning and healthcare resource management. There are massive volumes of information from this sector that was produced and collected on daily basis activities such as personal information of patient, laboratory results, underlying disease and so on. With the use of implementing information technologies, the large amounts of data can be automatically classified, processed and extracted into the useful patterns which help to get interesting knowledge or hidden pieces of information. Different from standard practices of data mining which simply begins with obscure dataset, data mining in healthcare is initiated from the hypothesis. The discovered information that is not in accordance with patterns and trends will be focused [5] for investigation purpose. Well-known data mining techniques which are successfully applied in healthcare include Artificial Neural Networks, Decision Tree, Genetic Algorithms and Nearest Neighbor method [5,6]. The information which is retrieved from data mining process will be used for decision making process, prognoses, improve clinical practice and provide treatment recommendation for patients in healthcare organizations III.

RELATED WORKS

AKI has been studied by many researchers in different area such as risk, definition and efficiency of tool applied. N. Srisawat et al. [7] studied variation in risk and mortality of AKI by cohort study of the 15,132 critically ill patients admitted to ICU at 6 hospitals in 4 countries and used KDIGO to define AKI. The result was the incidence and outcomes for AKI varied after adjusting for age, sex, baseline severity of illness, and also associated with incremental mortality risk across centers. Likewise, Kai Singbartl et al. [8] study, the detrimental effects of AKI were exposed that they were not limited to classical symptoms such as fluid overload and electrolyte abnormalities but also caused higher rates of infection, CKD (Chronic Kidney Disease) progression and impaired innate immunity. The study has also shown that the death cases among ICU patients, who died of AKI, often involved various factors. Even small changes in serum creatinine concentrations can increase the risk of death. Dilhari DeAlmeida et al. [9] proposed an approach to develop and extract AKI from Electronic Medical Record (EMR) and presented the methodology to evaluate the effect of electronic alerts (e-alerts) system of AKI. The study collected data on AKI e-alerts in university medical center in November

2013 with a sample of 100 non-ICU patients who were analyzed for possible AKI. AKI was defined per KDIGO guideline. The result was found that all cases examined were true cases of AKI with 100% Specificity and 97% Sensitivity. F.Perry Wilson et al. [10] assessed the effectiveness of an automated electronic alert system for AKI to determine injury severity reduction and outcomes improvement. The study was designed to a single-blind, parallel-group and randomized controlled trial. The 23,664 patients were recruited from the hospital of the University of Pennsylvania in Philadelphia between September 2013 and April 2014. After exclusion criteria, 2,393 patients were randomized as 1,201 eligible participants were assigned to AKI alert group and 1,192 were assigned to the usual care group. The result of this study was concluded that the randomized controlled study did not show any advantages of an electronic alert system, nor did it improve the clinical outcomes among patients in a hospital. The application of data mining in healthcare has been developed for various purposes. Asmaa S. Hussein et al. [11] proposed the recommender systems for Chronic Disease Diagnosis (CDD) using data mining techniques in the hybrid way such as multiple classifications and Unified Collaborative Filtering method. The first approach based on decision tree algorithms was applied for monitoring cases by generating disease risk diagnosis prediction. The second approach based on learning classification which was suggested to be used for increasing accuracy of medical advices recommendation. The results were exposed that the system can assist the physicians and patients to decrease chronic disease risk factor and to provide high accuracy of medical advices recommendation which will be benefit for patients. Feixiang Huang et al. [12] applied data mining techniques to predict hypertension from 8 diseases. The 9,862 cases of patient medical record were selected and studied. The approach adopted Under-sampling technique to generate training data sets and used data mining tool Weka to generate the Naive Bayesian and J-48 classifiers. The experimental results were shown that there was a little improvement of the ensemble approach over pure Naive Bayesian and J-48 in prediction performance such as accuracy, sensitivity, and F-measure. These related works were studied and used as a reference and guideline about how to analyze patient data received from hospitals with an appropriate methods and way to design and develop “AKIHelper” tool which is planned to use in specific area, hospitals in Thailand. It will be a challenge of this study because the current standard approach to diagnosis AKI in Thailand is done manually without the aid of systems or tools. IV.

PROPOSED APPROACH

This section describes the proposed study which aims to assist the doctor for decision making process or to provide the information and recommendations for improving the clinical outcomes. The approach is divided into 3 sections. Section A provides an approach to identify AKI risk factors by applying Classification tree technique of data mining such as J48 and Simple Cart to be used for patient dataset processing. Section B presents an idea for developing diagnostic tool named “AKIHelper” using KDIGO guideline. Lastly, section C

describes the method for evaluating this approach using 2x2 contingency table. The conceptual framework is illustrated in Fig. 1 which is separated into two sections, A and B. Each of the section is explained further in description below.

Collect Data

Transform Data

Select Data

Section A

TABLE I.

INFORMATION GROUPS AND ITS ATTRIBUTES

Type

Clean Data Dataset Integrate Data

Apply Data Mining Technique

characteristics, underlying diseases, types of ICU, laboratory results and outcomes that contain their own attributes. These attributes are considered for identifying risk factors as shown in Table 1.

Patient Information

Preprocessing Data

Identify the result (AKI Risk factors)

Underlying Disease

Section B Study KDIGO Guideline

Design Tool

Develop Tool AKIHelper

Evaluate the Result

Create Staging Model

Section C

Fig. 1. Conceptual framework of proposed approach

A. Identify the Risk Factors of AKI 1) Setting: Data used to perform data mining analysis was collected from 14 hospitals in regions across Thailand by registration in web-based format where demographic, clinical and laboratorial data were recorded. 2) Dataset: Sequence of data collection is everyday for the first 7 days and then weekly collected on day 14, 21 and 28 of ICU admission date. The initial dataset included 2,480 records during April to December 2014 of patient admission. A sample of patient data is shown in Fig. 2.

Types of ICU

Laboratory Investigations Outcomes

Attribute Birth Date Age Gender Weight (kg.) Height (cm.) Length of ICU stay (days) Diagnosis (At ICU admission) Hypertension Diabetes Mellitus Coronary Artery Disease Cerebrovascular Disease Malignancy Chronic Kidney Disease (CKD) Medical Surgical Coronary Care Unit Mixed Glasgow Coma Scores (GCS) APACHE II SOFA Scores AKI (Y/N) Stages of AKI (1-3)

To identify the risk factors of AKI based on Table 1. attributes, classification data mining technique including J48 and Simple Cart are applied in data mining process. a) J48: is a tree based learning approach which uses divide and conquer approach to construct univariate decision tree by spliting a root node into a subset of two partitions till leaf nodes (target node) occur in a tree. [13] Random samples of 100 patient records were applied with J48 technique and the result of execution are represented in Fig. 3. It was found that Chronic Kidney Disease (CKD) patients is associated with a high risk of AKI. In addition, it was indicated that the patient with Sofa (Sepsis-related Organ Failure Assessment) score less than or equal to 11 and age greater than 82 years old with no hypertension also increases the risk of AKI.

Fig. 2. A sample of patient data

3) Data Mining Process: Dataset was prepared for using in this process. All data of the patients who were under 15 years old or with end stage renal disease (ESRD) on chronic dialysis and unnecessary data such as name, patient number and reimbursement were cut-off. The remaining data were categorized into patient information groups including baseline

Fig. 3. Result of sample patient dataset applied with J48 decision tree

b) Simple Cart (Classification and Regression Tree): is a classification technique for generating the binary decision tree. This algorithm based on learning approach. The results can be either classification or regression trees, depending on categorical or numeric dataset[14]. Fig. 4 represents the result from applying Simple Cart technique with a sample of 100 patient records that the patient who was admitted to Medical ICU type has a high risk of AKI.

In Fig. 5, severity staging model approach is presented. It was created based on KDIGO guideline which will focus on changing in serum creatinine level or urine output. The proportional increase indicates the AKI stage. The attributes such as baseline, admission time and urine output were mapped with column in the records.

Input Patient

CART Decision Tree ICUType=(CCU)|(Mixed)|(Surgical): No(52.0/7.0) ICUType!=(CCU)|(Mixed)|(Surgical): Yes(22.0/7.0) Number of Leaf Nodes: 2 Size of the Tree: 3

[ESRD] [Not ESRD] [Diagnose by Creatinine Level]

[Diagnose by Urine Output [Weight Available]

[Weight N/A]

Check Time of Creatinine

Fig. 4. Result of sample patient dataset applied with Simple Cart

The result from this section will be used for monitoring patients at risk to increase the outcome and be applied to the tool for providing information purpose. B. Developing Diagnostic Tool To develop diagnostic tool “AKIHelper”, understanding the KDIGO guideline is required because it will be used as the main structure of the tool, for example, definition, severity staging and recommendations. The approach for developing the tool consists of 2 steps, as in (1) AKI Severity Staging Model and (2) Tool Development Approach. Each of step is presented seperately in the following subsections. 1) AKI Severity Staging Model: The model was created based on KDIGO guideline which provide the definition and the staging of AKI [4] as described in Table 2. and Table 3. respectively.

[ Admission time]

Choose Lower Value (colAE VS (colBI))

Estimate Weight by Height Male = Height - 100, Female = Height - 105

Obtain Reference Creatinine (colBG)

Divide Urine Output per Day (colCK) by Weight

Compare Creatinine Day 1 – 7, 14, 21, 28 with colBG

Obtain Urine Output per Kilogram

Each Time Determine Status of AKI (0 – 3 Using the Guideline)

Fig. 5. Model for staging AKI severity using KDIGO guideline TABLE II.

DEFINITION OF AKI (NOT GRADED)

Definition Increase in SCr by ı0.3 mg/dl (ı26.5 lmol/l) Increase in SCr to ı1.5 times baseline Urine volume Medical Record System

Record Data

Stage Severity of AKI

Doctor

View Recommendation Medical Staff View Risk Factors

LIMITATIONS

Although the approach is proposed for assisting doctor in various expectations. One limitation of this tool is that it can only be used in the part of diagnosis process for staging and provide guideline information only, it is not able to handle and manage all of the clinical processes especially decision making, for example, the severity of AKI that AKIHelper can detect may be missed if there are multiple serum creatinine results within a short time period. In this case, using such tool is not effective. Doctor’s consideration and judgement on the patient treatment will be required. In addition, the risk factors may various across the areas due to source of patient data, location setting, differences in clinical practice, process of care or residual confounding from unmeasured factors. Therefore, using different dataset before applying data mining techniques would be effective to identify the risk factors. CONCLUSION

This paper presents the proposed approach to identify risk factors using Decision Tree classification of data mining technique. Simple Cart and J48 were selected as the algorithms for this process. The findings from the study will be used to improve monitoring and treatment pathway in patients at risk. Next, an idea for developing diagnostic tool named AKIhelper was presented. The model for building tool was created based on KDIGO guideline for staging the severity of AKI. The completed AKIhelper is expected to assist the doctor in decision making processes and improve quality of medical outcome which may result in a reduction in mortality and medical expenses.

Fig. 7. Example user interface of AKIHelper

C. Evaluation To evaluate the proposed approach, 2x2 contingency table is planned to use as depicted in Table 4.

A further study is needed to provide better understanding of the association between undisclosed factors and the increased risk of AKI and to develop tool to be plugged in with the electronic medical record system of the hospital or clinic.

2 X 2 CONTINGENCY TABLE

ACKNOWLEDGMENT

Actual Result Disease

Non-Disease

Positive

a True Positive

b False Positive

Negative

c False Negative

d True Negative

Test Result

V.

VI.

Fig. 6. Use case diagram of AKIHelper

TABLE IV.

proportion of negatives that are correctly identified. Accuracy value is used to show the precision of the system.

With this table, sensitivity, specificity and accuracy value can be measured as equations below. Sensitivity = a / (a + c) Specificity = d / (b + d) Accuracy = (a + d) / (a + b + c + d)

(1) (2) (3)

Sensitivity value is used to measure the proportion of positives that are correctly identified (rule out the disease). On the contrary, specificity value is used to measures the

This paper was supported by Dr. Nattachai S., Excellent Center for Critical Care Nephrology, Department of Medicine, Chulalongkorn University, Thailand and Department of Critical Care Medicine, University of Pittsburgh School of Medicine, USA and Dr. Wirichada P., Mathematical and Economic Modelling group, Mahidol-Oxford Tropical Medicine Research Unit, Mahidol University for advices and assistance with data collection. REFERENCES [1]

[2]

Dealmeida, D.; Al-Jaghbeer, M.; Abdelhak, M.; Kellum, J., "A Study to Evaluate the Effectiveness of the Currently Utilized Acute Kidney Injury (AKI) Alert: A Use Case Example for a Learning Health System," in System Sciences (HICSS), 2015 48th Hawaii International Conference on , vol., no., pp.3125-3131, 5-8 Jan. 2015. K. Kashani and V. Herasevich, "Sniffing out acute kidney injury in the ICU: do we have the tools?," Curr Opin Crit Care, vol. 19, pp. 531-6, Dec 2013.

[3]

[4]

[5]

[6]

[7]

[8]

[9]

M. James, E. Dixon, D. Roberts, R. Barry, C. Balint, A. Bharwani, et al., "Improving prevention, early recognition and management of acute kidney injury after major surgery: results of a planning meeting with multidisciplinary stakeholders," Canadian Journal of Kidney Health and Disease, vol. 1, p. 20, 2014. Kidney Disease: Improving Global Outcomes (KDIGO) Acute Kidney Injury Work Group. KDIGO Clinical Practice Guideline for Acute Kidney Injury. Kidney inter., Suppl. 2012; 2: 1–138. B. Milovic and M. Milovic, "Prediction and decision making in Health Care using Data Mining," Kuwait Chapter of the Arabian Journal of Business and Management Review, vol. 1, p. 126, 2012. Muhamad, H.; Muhamad, A.; Husain, W.; Rashid, N., “Data Mining for Medical Systems: A Review,” in Proc. of the International Conference on Advances in Computer and Information Technology - ACIT 2012, pp.17-22 N. Srisawat, F. E. Sileanu, R. Murugan, R. Bellomod, P. Calzavacca, R. Cartin-Ceba, et al., "Variation in risk and mortality of acute kidney injury in critically ill patients: a multicenter study," Am J Nephrol, vol. 41, pp. 81-8, 2015. Singbartl, Kai, and John A. Kellum. "AKI in the ICU: definition, epidemiology, risk stratification, and outcomes." Kidney international 81.9 (2012): 819-825. Dealmeida, D.; Al-Jaghbeer, M.; Abdelhak, M.; Kellum, J., "A Study to Evaluate the Effectiveness of the Currently Utilized Acute Kidney

[10]

[11]

[12]

[13] [14]

Injury (AKI) Alert: A Use Case Example for a Learning Health System," in System Sciences (HICSS), 2015 48th Hawaii International Conference on , vol., no., pp.3125-3131, 5-8 Jan. 2015 F. P. Wilson, M. Shashaty, J. Testani, I. Aqeel, Y. Borovskiy, S. S. Ellenberg, et al., "Automated, electronic alerts for acute kidney injury: a single-blind, parallel-group, randomised controlled trial," The Lancet, vol. 385, pp. 1966-1974. Hussein, A.S.; Omar, W.M.; Xue Li; Ati, M., "Efficient Chronic Disease Diagnosis prediction and recommendation system," in Biomedical Engineering and Sciences (IECBES), 2012 IEEE EMBS Conference on , vol., no., pp.209-214, 17-19 Dec. 2012 Huang, Feixiang; Wang, Shengyong; Chan, Chien-Chung, "Predicting disease by using data mining based on healthcare information system," in Granular Computing (GrC), 2012 IEEE International Conference on , vol., no., pp.191-194, 11-13 Aug. 2012 Neeraj et al., International Journal of Advanced Research in Computer Science and Software Engineering 3(6), June - 2013, pp. 1114-1119 Kalmegh, Sushilkumar. "Analysis of WeKA Data Mining Algorithm REPTree, Simple Cart and Random Tree for Classification of Indian News." International Journal of Innovative Science, Engineering & Technology (IJISET) 2.2 (2015): 438-446.