Early detection of acute kidney injury with Bayesian networks Harry Cruz12 , Bastien Grasnick1 , Henriette Dinger 1 , Frank Bier2 , and Christoph Meinel1 1 Hasso Plattner Institute, Potsdam 2 Fraunhofer Institute for Cell Therapy and Immunology, Potsdam Harry.FreitasDaCruz@hpi.de (corresponding author) Abstract diagnosed for AKI too late, leading to more com- plications and higher mortality. In fact, a study Acute kidney injury (AKI) is a major in the UK found that 60% of post-admission AKI health issue, affecting large numbers of cases were avoidable (Stewart et al., 2009). An au- patients worldwide. It is associated with tomated, early detection of high-risk patients may an increase in complications and poor lead to a faster response of physicians, reducing prognostics if diagnosis is delayed. Med- complications associated with AKI. ical guidelines are routinely employed to Our objective was therefore to develop a proof classify different AKI stages, but guidance of concept for early detection of AKI. For this pur- on the early detection of AKI risk is lim- pose, we created and trained a Bayesian network ited. In this paper, we present a Bayesian model on the basis of real patient data. A Bayesian Networks (BN) proof of concept to predict network is a probabilistic graphical model, con- the likelihood of AKI onset based on lon- sisting of random variables and their influences on gitudinal patient data, such as serum cre- one another. Data for training and validating the atinine values, demographics and comor- model was obtained from the anonymized Multi- bidities. Data for training and validating parameter Intelligent Monitoring in Intensive Care the model was obtained from the Multi- II (MIMIC II) database (Lehman et al., 2011). In parameter Intelligent Monitoring in Inten- this paper, we will provide the background needed sive Care (MIMIC II) database. We de- and present the methods involved in developing scribe the problem domain, data acquisi- the model, including data acquisition and prepa- tion and preparation, model developed, re- ration, results obtained and further discussion. sults obtained and pertaining limitations. We demonstrate that our model can pre- 2 Background dict the onset of the disease with an accu- This section deals with the necessary background racy of up to 87% (area under the curve of for the remainder of this paper. This includes 0.87) in the cohort under analysis. an elucidation of risk factors related to AKI, 1 Introduction Bayesian networks fundamentals as well as related work. Acute Kidney Injury (AKI) affects a large portion of the elderly population and has a high risk of 2.1 Risk factors for AKI death, as there is no trivial treatment once it breaks As a starting point, we needed to identify factors out (Statistisches Bundesamt, 2014). After on- which predispose patients to AKI from medical set, the patient may even need dialysis and/or re- literature sources. These are, among others, cre- nal replacement therapy (kidney transplant). Cur- atinine values taken from blood or urine samples. rently, detection of kidney injury requires contin- Furthermore, comorbidities, such as heart failure uous monitoring of creatinine and other lab values or diabetes and personal background, including (Harty, 2014). In particular, when many patients age, gender and ethnicity, have to be considered. must be monitored at once, it is hard for physi- Since we are in the domain of kidney diseases, de- cians to keep track of subtle changes in blood mea- hydration plays an important role too (Lopes and surements which might be indicative of AKI. As Jorge, 2013; Kellum et al., 2012). In detail, the a consequence, a significant portion of patients is relevant factors are: • laboratory values (serum creatinine, urine output, estimated gloremular filtration rate (eGFR) value) • comorbidities (heart failure, chronic kidney disease, tumor disease, diabetes, obesity, hy- pothyroidism, paralysis, hypertension, pul- monary circulation, valvular disease, peptic ulcer, deficiency anemia, renal failure) • personal background (age, gender, ethnic- ity, admission type, that is emergency or elec- tive) Figure 2: AKIN classification for AKI The clas- sification system works with the help of serum cre- 2.2 Acute Kidney Injury classification atinine and urine output values. There are three Currently, two main guidelines are used in possible categories: risk, injury and failure. medicine for the classification of AKI: RIFLE (Risk, Injury, Failure, Loss) and AKIN (Acute Kidney Injury Network). Both help physicians patients, while RIFLE guideline is better suited establish severity of kidney injury based on the for patients in advanced stage of renal function serum creatinine and urine output of a patient. Fig- loss, while Since both guidelines are well-proven ures 2 and 1 (Cruz et al., 2009) show an overview in practice, they will serve as an additional output of the two classifications depending on the creati- variables for the model (Lopes and Jorge, 2013; nine and urine values. Kellum et al., 2012). 2.3 Bayesian networks Howard et al. define a Bayesian network as “an annotated directed graph that encodes proba- bilistic relationships among distinctions of interest in an uncertain-reasoning problem” (Howard and Matheson, 1983). In general, a Bayesian network consists of multiple random variables and their conditional dependencies modeled as probability functions. This way, based on evidence provided for one or more given variables, the probability of the other random variables can be calculated after Figure 1: RIFLE classification for AKI The RI- the network was trained (Horný, 2014). FLE classification uses serum creatinine and urine An example is given in Figure 3 adapted from output values. It consists of five classes: risk, in- (Barbini et al., 2013), where the probabilistic de- jury, failure, loss of kidney function and end-stage pendencies of a simple Bayesian model with four kidney disease. dichotomous variables (true or false) is shown. It follows from the model that A and B (having a pri- The RIFLE classification (Bellomo et al., 2004; ori associated probalities) exert influence on C. In Ricci et al., 2011) predates AKIN. It consists of turn, this effect is modeled by a conditional prob- five stages: risk, injury, failure, loss of kidney ability table on the children node. As such, the function and end-stage kidney disease (ESKD). probability of event C occurring given that A oc- In comparison to that, the AKIN classification curred but not B is given by XAB . Nodes C and (Ricci et al., 2011; Mehta et al., 2007) uses only D are independent of each other (conditional inde- three stages: risk, injury, and failure (Lopes and pendence). Jorge, 2013; Kellum et al., 2012). Since AKIN is A Bayesian network is developed either by us- based on RIFLE, it is more widely used nowadays. ing expert knowledge and building the network AKIN performs better for detecting early stage manually or letting the network be built directly from the data by a specific algorithm, an approach Further, Król et al. developed an approach to known as structure learning. Once the structure predict chronic kidney disease (CKD). They did of the network is learned, it can be further manip- not build a technical system but “an algorithm for ulated using expert knowledge. Moreover, when the diagnostic procedure” (Król et al., 2009). For the structure is set, the probability functions rep- this purpose, they did an investigative survey with resenting the conditional dependencies can also be 2471 randomly chosen people involved. As a re- learned from data. This is referred to as parameter sult, they found different factors that encourage a learning. CKD. Among others, these factors are the male The most significant drawback of Bayesian net- gender, diabetes and hypertension. works is that the accuracy depends highly on the Specifically applying Bayesian networks, chosen structure of the model. If this is done Onisko et al. present a model based on dynamic negligently, the resulting model can fail to show BNs for predicting the risk of cervical cancer, existing results (false negatives) or show incor- using hospital data and expert knowledge. The au- rect results (false positives). A common way to thors were able to categorize patients in different avoid this is to iteratively establish dependencies risk categories (Onisko et al., 2004). In a similar among variables, usually based on expert knowl- approach, Nachimuthu et al. used BNs for early edge (Heckerman et al., 1995). detection of sepsis(Nachimuthu and Haug, 2012). Similarly, Ward et al. offer a framework for the development of Bayesian networks in the partic- ular example of sepsis. They build their model based on knowledge gained from literature, hospi- tal data as well as expert knowledge. Their result- ing model provides a base for a correct prediction. Since the data set is rather small, a further evalua- tion is planned to support their result (Ward et al., 2014). In an approach analogous to these works, we de- veloped a model based on Bayesian networks for Figure 3: Example of a Bayesian network This estimating the risk of developing AKI. We were Bayesian network shows four random dichoto- also able to use hospital data and an expert con- mous variables and their probabilistic relation- sultation for the development. To the best of the ships. authors’ knowledge, this is the first work explic- itly utilizing a Bayesian network model for AKI 2.4 Previous work prediction. Machine learning has been widely utilized in the 3 Model development medical domain in several instances. Particu- larly in Nephrology, Legrand et al. evaluated 3.1 Methodology the post-operative AKI risk of patients suffering For the model development, we utilized two ma- from infective endocarditis after undergoing car- chine learning tools, Weka and GeNIe and com- diac surgery. They applied super learning, a tech- pared their accuracy to control for possible tool nique to choose the optimal regression algorithm, bias in the results. Further, we extracted the comparing ten different models by using cross- needed data from the MIMIC II database, which validation. Targeted maximum likelihood estima- was preprocessed for tool input. We created two tion was used to obtain the following most impor- data sets for cross-validation, one with 6000 en- tant risk factors: multiple surgery, pre-operative tries and another with 9000 entries (50% more). In anemia as defined by a baseline hemoglobin level the first iteration, the AKI literature laid out in sec- <10 g/dl, transfusion requirement during surgery, tion 2.1 formed the basis for the development of an the use of a nephrotoxic agent: vancomycin, initial model (1st iteration model). This model was aminoglycoside or contrast iodine; and the inter- then augmented and corrected after an expert con- action between vancomycin and aminoglycoside. sultation session with nephrologists at the Charité (Legrand et al., 2013). hospital in Berlin (2nd iteration model). We then compared the two models, as well as the differ- was generating a comma-separated values file by ent tools and analyzed the results obtained. The querying the database tables for preprocessing. following sections will provide further details into 3.3.2 Data preprocessing this procedure. Besides demographic information, the MIMIC II 3.2 Tools utilized database provides information about several risk factors. Furthermore, there are tables for medica- In an effort to avoid bias in the results possi- tion and laboratory events which were used for the bly introduced by differing algorithm implemen- model as well. In order to train and evaluate them, tations, we chose to develop and test the model in we decided to choose a cross-validation approach. two widely available Bayesian network modelling This enables training and evaluation within the tools, Weka and GeNIE. same data set. Apache Weka is a Java toolkit for different For this purpose, we extracted two different data kinds of data mining algorithms. It allows the clas- sets. The first one consists of 6000 patients, the sification, clustering and visualization of data sets second of 9000. Table 1 shows the distribution of (Kumar and Sahoo, 2012). One of the main ad- patients with and without AKI in the two data sets. vantages of Weka is the very powerful capabilities They contain information about the demographics for Bayesian networks (Bouckaert, 2008). For net- of a patient, their comorbidities, the latest crea- work structure learning, an estimator as well as a tinine value changes and an indication whether a search algorithm can be set as parameters in the patient was diagnosed with AKI or not. tool. For the purposes of this paper, we chose the algorithm K2 as it has the best performance among Entries AKI No AKI the search algorithms implemented in Weka (The 6000 50% (Stage 1, 2 & 3) 50% University of Waikato, 2008). 9000 33% (Stage 1, 2 & 3) 67% GeNIe is the user interface of SMILE, a C++ library for the development of graphical decision Table 1: Distribution of the two data sets models (Druzdzel, 1999). Therefore, in compar- For use in our experiments, we needed to pre- ison to Weka, it is limited to Bayesian decision process the data. This included a discretization of models and has no possibility for other data min- continuous values, as well as the computation of ing algorithms. Since it is the most generic ap- auxiliar values derived from available information. proach and suitable for most applications we de- We computed the estimated Gromerular Filtration cided to use the Bayesian search as the algorithm Rate (eGFR) according to the existing guidelines of choice for GeNIE. In effect, heuristic algo- (NIDDK, 2016), since this rate is an important in- rithms such as Tree Augmented Naive Bayes are dication of overall renal function. Next, we calcu- only recommended for large scale projects (Deci- lated the increase of serum creatinine for each pa- sion Systems Laboratory, 2016). tient across multiple measurements and used this 3.3 Model data value for categorizing the severity of kidney in- jury according to the AKIN guideline (Mehta et 3.3.1 Data acquisition al., 2007). The data thus preprocessed can be fed The accuracy of the developed model depends into the tools and offers the necessary information highly on the underlying data. For training pur- to enable risk prediction and result validation. poses, a real dataset consisting of patients affected by AKI and those not affected by it was needed. 3.4 Model input and output This set was obtained from the MIMIC II database The prepared data set is the basis for the Bayesian from PhysioNet (Lehman et al., 2011) which con- network. This means that the risk factors pre- tains data from intensive care units (ICU) from sented in section 2.1 (lab values, comorbidities hospitals in the United States. We utilized the con- and demographics) are the input variables for tained information about disease indications, de- training and running the model. The resulting out- mographics, lab results (most importantly creati- put are the probabilities for the presence of AKI nine value measurements) and comorbidities. AKI as well as the classification stages of RIFLE and is represented by the ICD (International Classi- AKIN as inferred from the provided input. It en- fication of Diseases) code 584.9. The final step ables the user to employ it for decision support with the same or other data sets. For illustration the feedback, we removed the node for RIFLE. purposes, a graphical representation for the second Furthermore, we added the nodes for the new risk iteration model is provided in . factors. Finally, we appended more dependencies between the factors and AKI as well as AKIN. 3.5 First iteration model In the first iteration, the model included the in- put variables as indicated in the AKI literature, encompassing laboratory values, patient demo- graphics and comorbidities. These random vari- ables are the input nodes. Each node has its own probability as well as possible posterior proba- bilities stored which define its impact on defin- ing AKI. These probabilities were automatically trained from the data set obtained from the MIMIC II database. The model concentrates on the AKI as well as the two classification guidelines RIFLE and AKIN. For the sake of brevity, we will not provide a graphical representation of the first iter- ation model. The structure, however, will be clear from analyzing the second iteration model, which already incorporated expert feedback. 3.6 Second iteration model In the next step, we discussed the model in the first iteration with nephrologists from the Charité hos- pital in Berlin. The following main insights were gained: Figure 4: Second iteration model Based on the main insights from the expert consultation, this 1. RIFLE guideline is not used in practice any model was developed. The RIFLE guideline was more since it is an older classification system. removed. Furthermore, new risk factors and de- Moreover, AKIN is based on RIFLE and is pendencies were added. Comorbidities are shown thus the only classification system needed; in gray. 2. There are further influence factors which need to be considered. These factors are weight, urethritis and medication history; 4 Results 3. The time of comorbidities has an influence on This section evaluates how accurate the developed making the correct diagnosis. Diseases that models were. To this extent, the used data set is are years ago have less impact than more re- presented. Moreover, we compare the accuracy cent diseases; values for the two developed models. We show that both improving the model with expert knowl- 4. Physicians normally do not trust any sys- edge as well as increasing the data size increases tems but only themselves. A CDSS needs the accuracy of the model. Our best result was an to provide demonstrable value. Additionally, accuracy of 87% for predicting the occurrence of nephrologists mostly do not need such a sys- AKI. tem since they recognize the symptoms them- selves based on their experience. A better 4.1 Accuracy of the models use case is the ICU, where the physicians are not kidney experts and are overwhelmed with Table 2 shows the obtained results depending on monitoring data. both the utilized tool and the data set used (6000 or 9000). It stands out that the 2nd iteration model Based on the discussion at the Charité, we de- consistently performs better than the first one. velop a second model, shown in Figure 4. As per This demonstrates that expert knowledge is help- ful in improving model performance in such a spe- accuracy. As such, the higher the volume of data cialized scenario as kidney disease. Moreover, it available, the better the results achieved. This was shows the flexibility of Bayesian networks, which demonstrated by increasing in 50% the data vol- allows to integrate such expert knowledge. Fur- ume (from 6000 to 9000). This finding suggests thermore, noticeable discrepancies between the that more robust prediction models can be devel- different tools can be observed. Overall, we oped for the medical domain by tapping into larger achieved a top accuracy of 87% when using Ge- databases. NIe. The results also show discrepancies concern- ing the tools used (Weka and GeNIe). This can 1st iteration 2nd iteration be accounted for by differences in algorithm im- Dataset GeNIe Weka GeNIe Weka plementation and configuration parameters. This 6000 67% 58% 83% 76% fact underscores the need for comparison not only 9000 73% 72% 87% 83% among algorithms, but also among different tools Table 2: Accuracy of the developed models as and configurations, since the details of algorithm per the different tools and datasets implementation can greatly vary. In this paper, while variations were present, the results were A more detailed view of the accuracy can be largely consistent, except for a much poorer result seen by analyzing the receiver operating charac- from Weka when dealing with the smaller dataset. teristic (ROC) of the best performing experiment. The comparably small data set due to limited The ROC curve describes the relation between true hardware capacity was the biggest drawback of positives (TP) and false positives (FP). The perfect our experiments. Furthermore, the limited scope result would be a TP value of 100% and a FP value of this work lead to the decision of concentrate of 0%. Figure 5 shows the curve of the 2nd iter- on one algorithm per tool. Indeed, different algo- ation model (after expert feedback). The curve is rithms show different advantages which we were based on the larger data set (9000 entries) and refer not able to consider for this paper. As such, a to computed AKI patients. more robust analysis should include a comparison of different algorithms, tools and configuration pa- rameters. Going one step further, instead of solely in- creasing the data volume, another promising di- rection to follow would be to increase data vari- ety, including more relevant data, such as urethri- tis as a comorbidity and complete disease history. As the experts consulted suggested, this might im- prove the accuracy as well. Unfortunately, this data could not be obtained from the data source available but it must be included if this this model is to be used in practice. In the beginning, we showed related works that developed a CDSS for various diseases (Onisko et al., 2004; Nachimuthu and Haug, 2012; Ward et al., 2014). These works employ dynamic Bayesian Figure 5: ROC curve of the 2nd iteration model networks, which consider the time component as Curve shows the performance of the Bayesian the involved random variables change. The ap- classifier for the AKI output variable. Area under proach presented in this work is concerned with the Curve (AUC) = 0.8743 a static view. This represents a possible weakness in comparison to other similar works and must be addressed in the future. 5 Discussion Finally, while decision support systems show The results obtained show that for the cohort under much promise in improving healthcare delivery, analysis increasing data volumes lead to improved evidence towards their efficacy in clinical settings is lacking, leading to skepticism among medical Bayesian network on 9000 data entries from ICUs professionals. Particularly in the field of Nephrol- in the US, obtained from the MIMIC database. ogy, a controlled randomized trial (CRT) has been Using information about demographics, comor- conducted by Wilson et al. testing an early warn- bidities and creatinine values, a satisfactory ac- ing alert system for AKI. The CRT which yielded curacy for predicting the risk of an AKI was ob- no demonstrable positive outcomes for patients tained. The results show that a further develop- (Wilson et al., 2015). Their algorithm was based ment and improvement of such a model by inte- on the mere detection of creatinine thresholds and gration of expert knowledge leads to improved ac- the authors of this study encouraged new trials curacy values. However, such initiatives are fre- with more sophisticated algorithms. Such expe- quently met with skepticism by the medical com- riences strengthen the need for making CRT a munity. Randomized controlled trials are needed standard procedure for a prospective CDSS. Even to assess the benefits and potential risks for pa- though such procedures are costly, if benefits can tients and doctors, along with full integration with be factually demonstrated, medical acceptance can medical workflows, so that they can be convinced be increased. of the potential advantages of such a system. 5.1 Further work 7 Acknowledgements Since the results show that larger data sets tend Author H. Cruz was kindly supported by a PhD to deliver more accurate results, the tests should grant from CAPES Foundation, Ministry of Edu- be repeated with more representative data sets that cation of Brazil, Brası́lia, Brazil. might also contain new input variables like urethri- tis as a comorbidity and the history of diseases. Furthermore, in a practical application, the ma- References chine learning system has to be easily modifiable. Emanuela Barbini, Pietro Manzi, and Paolo Barbini. One approach is to develop a Clinical Decision 2013. Bayesian Approach in Medicine and Health Support System specifically for this purpose. An- Management. In Current Topics in Public Health. other one is to use the existing tools, but perform InTech, may. more experiments with changing parameters. Rinaldo Bellomo, Claudio Ronco, John A Kellum, For actual use in practice, integration with care Ravindra L Mehta, and Paul Palevsky. 2004. Acute delivery workflows and the hospital information renal failure - definition, outcome measures, an- system itself is needed. The user interface of such imal models, fluid therapy and information tech- nology needs: the Second International Consensus a system has to be clearly structured and free of Conference of the Acute Dialysis Quality Initiative unnecessary information, so that physicians can (ADQI) Group. Critical care (London, England), readily see the most important facts about a pa- 8(4):R204–12. tient. Further, an alert system powered by the Remco Bouckaert. 2008. Bayesian network classifiers clinical decision support system must be imple- in Weka for version 3-5-7. Technical report. mented, so that care team can be notified timely in case of AKI. Still, it remains to be seen how many Dinna N Cruz, Zaccaria Ricci, and Claudio Ronco. physicians would actually use such a system, so 2009. Clinical review: RIFLE and AKIN–time for reappraisal. Critical care (London, England), acceptance is a possible barrier. Similarly, the ex- 13(3):211. pected benefits for the patients must also be mea- sured in terms of outcomes (rate of complication, Decision Systems Laboratory. 2016. GeNIe Docu- mortality, prognosis, subjective well-being, etc.) mentation. Therefore, a pilot study considering both physi- Marek J Druzdzel. 1999. SMILE: Structural Model- cian acceptance and patient outcomes should be ing, Inference, and Learning Engine and GeNIe: A conducted. Development Environment for Graphical Decision- Theoretic Models. Proceedings of the Sixteenth Na- 6 Conclusion tional Conference on Artificial Intelligence (AAAI- 99), pages 342–343. We developed a proof of concept for a machine John Harty. 2014. Prevention and management of learning model that can be used in a CDSS in the acute kidney injury. The Ulster medical journal, domain of AKI. For this purpose, we trained a 83(3):149–57. D Heckerman, D Geiger, and D Chickering. 1995. Symposium proceedings / AMIA Symposium. AMIA Learning Bayesian networks: the combination of Symposium, 2012:653–62, jan. knowledge and statistical data. Machine Learning, 20, 197-243, 243:197–243. NIDDK. 2016. Estimating Glomerular Filtration Rate (GFR). Michal Horný. 2014. Bayesian networks: A Technical report. Technical Report 5. A Onisko, MJ Druzdzel, and RM Austin. 2004. Appli- cation of Dynamic Bayesian Networks to Cervical Ronald A. Howard and James E. Matheson. 1983. Cancer Screening. Intelligent Information Systems, Readings on the principles and applications of de- pages 1–10. cision analysis. Sdg Decision Systems, Menlo Park. Zaccaria Ricci, Dinna N Cruz, and Claudio Ronco. John a Kellum, Norbert Lameire, Peter Aspelin, 2011. Classification and staging of acute kidney in- Rashad S Barsoum, Emmanuel a Burdmann, Stu- jury: beyond the RIFLE and AKIN criteria. Nature art L Goldstein, Charles a Herzog, Michael Joan- reviews. Nephrology, 7(4):201–8, apr. nidis, Andreas Kribben, Andrew S Levey, Alison M MacLeod, Ravindra L Mehta, Patrick T Murray, Sar- Statistisches Bundesamt. 2014. Todesursachen in aladevi Naicker, Steven M Opal, Franz Schaefer, Deutschland. Miet Schetz, and Shigehiko Uchino. 2012. KDIGO Clinical Practice Guideline for Acute Kidney Injury. J Stewart, G Findlay, N Smith, K Kelly, and M Mason. Kidney international supplements, 2(1):1–138. 2009. Adding insult to injury. National Confidential Enquiry into Patient Outcome and Death, s10-I(1):4. Ewa Król, Bolesław Rutkowski, Piotr Czarniak, Ewa Kraszewska, Sławomir Lizakowski, Radosław Szu- The University of Waikato. 2008. Search algorithms. bert, Stanisław Czekalski, Władysław Sułowicz, and Andrzej Wicek. 2009. Early detection of chronic Logan Ward, Mads L Mogensen, Mical Paul, Leonard kidney disease: Results of the PolNef study. Ameri- Leibovici, and Steen Andreassen. 2014. A Bayesian can Journal of Nephrology, 29(3):264–273. Approach to Model-Development : Design of Con- tinuous Distributions for Infection Variables. In Yugal Kumar and G. Sahoo. 2012. Analysis of Bayes , 19th IFAC World Congress, pages 5653–5658. Neural Network and Tree Classifier of Classification Technique in Data Mining using WEKA. Computer F Perry Wilson, Michael Shashaty, Jeffrey Testani, Science & Information Technology (CS & IT), pages Iram Aqeel, Yuliya Borovskiy, Susan S Ellenberg, 359–369. Harold I Feldman, Hilda Fernandez, Yevgeniy Gitel- man, Jennie Lin, Dan Negoianu, Chirag R Parikh, Matthieu Legrand, Romain Pirracchio, Anne Rosa, Peter P Reese, Richard Urbani, and Barry Fuchs. Maya L Petersen, Mark Van der Laan, Jean-Noël 2015. Automated, electronic alerts for acute kidney Fabiani, Marie-paule Fernandez-gerlinger, Isabelle injury: a single-blind, parallel-group, randomised Podglajen, Denis Safran, Bernard Cholley, and Jean- controlled trial. The Lancet, 385(9981):1966–1974. Luc Mainardi. 2013. Incidence, risk factors and prediction of post-operative acute kidney injury fol- lowing cardiac surgery for active infective endo- carditis: an observational study. Critical care (Lon- don, England), 17(5):R220, jan. Li-wei Lehman, George Moody, Thomas Heldt, and Tin H Kyaw. 2011. Multiparameter Intelligent Monitoring in Intensive Care II (MIMIC- II): A public-access intensive care unit database. Critical Care, 39(February 2010):952–960. José António Lopes and Sofia Jorge. 2013. The RIFLE and AKIN classifications for acute kidney injury: A critical and comprehensive review. Clinical Kidney Journal, 6(1):8–14. Ravindra L. Mehta, John A. Kellum, Sudhir V. Shah, Bruce A. Molitoris, Claudio Ronco, David G. Warnock, Adeera Levin, and $author firstName $au- thor.lastName. 2007. Acute Kidney Injury Net- work: report of an initiative to improve outcomes in acute kidney injury. Critical Care, 11(2):R31. Senthil K Nachimuthu and Peter J Haug. 2012. Early detection of sepsis in the emergency department us- ing Dynamic Bayesian Networks. AMIA ... Annual