Design and Prototypical Development of a Web Based Decision Support System for Early Detection of Sepsis in Hematology Andreas Wicht Gerrit Meixner Ulrike Klein Heidelberg University Hospital, German Research Center for Heidelberg University Hospital, Institute of Medical Biometry Artificial Intelligence (DFKI) Department of Hematology and Informatics Trippstadter Strasse 122 and Oncology Im Neuenheimer Feld 305 Kaiserslautern, 67663 Germany Im Neuenheimer Feld 410 Heidelberg, 69120 Germany Gerrit.Meixner@dfki.de Heidelberg, 69120 Germany Andreas.Wicht@med. Ulrike.Klein2@med. uni-heidelberg.de uni-heidelberg.de ABSTRACT decades, the practical impact remains low for several Physicians do not always make optimal decisions. reasons, e.g. [3]: Computer based clinical support systems are intended to - Systems failed to cover an entire medical domain provide clinicians with decision aids, but their practical - Poor practicability and integration into the clinical impact remains low. We introduce a software architecture workflow which might overcome key barriers and present the - Poor availability of digital patient data prototypical implementation of a web based knowledge - Poor acceptance module for early detection a life-threatening medical Within the ESGOAB1 project, which will be described in condition, sepsis. more detail in the next section, we are trying to overcome the weaknesses mentioned above. In this paper we will Keywords briefly describe the ESGOAB software architecture, which Clinical decision support, Knowledge-based systems in provides an electronic health record (EHR) and is designed medicine, hematology, sepsis, fever, web-based application, to provide the base to interact with knowledge modules. We knowledge maintenance, ESGOAB. focused on two specific clinical challenges: supporting the INTRODUCTION physicians’ order entry process (CPOE) and supporting The information overload physicians are confronted every early detection of sepsis (a serious and life-threatening day with makes it impossible for them to keep up with all medical condition) on patients with hematological the information and knowledge that would be potentially underlying diseases. In this paper we will describe the useful in making optimal clinical judgments. Empirical second challenge. We introduce our conceptual design of studies have shown that physicians do not always make the sepsis knowledge module and present the currently optimal decisions [17] [6]. Clinical decisions are often implemented web-based prototype. made under time pressure, without having all information Project Background and knowledge needed in the right place at the right time. The ESGOAB project is a 2-year public funded joined Computer-assisted clinical decision support systems research project between two scientific partners (CDSS) are intended to provide “clinicians, patients or (Heidelberg University Hospital, DFKI) and two industrial individuals with knowledge and person-specific or partners (COPRA System GmbH, Dosing GmbH). population information, intelligently filtered or presented at appropriate times, to foster better health processes, A survey at the Department of Hematology and Oncology better individual patient care, and better population health. of the Heidelberg University Hospital has shown a range of CDS interventions include alerts, reminders, and order sets problems concerning various applications of Information […]” [11]. and Communications technologies (ICT) within the hospital Although research has been done in the field of CDSS for (response rate = 70.5 %, 36 of 51 of the medical personnel) 1 Copyright © 2011 for the individual papers by the papers' ESGOAB = „Entwicklung einer Softwareumgebung zur Generierung authors. Copying permitted only for private and academic von organisationsspezifischen Anwendungen zum Behandlungs- prozessmanagement“; english: Development of a Software Environment purposes. This volume is published and copyrighted by for Generation of Organization-Specific Applications for Treatment the editors of EICS4Med 2011. Process Management 69 [20]. The study revealed problems concerning e.g., time stress, and optimization of clinical decisions, which consuming searches, redundant data entries, use of various altogether may improve quality of care. software applications to perform the various tasks, different Related work user interfaces, and only marginal decision support. Research in the field of medical knowledge-based systems However we found a promising openness towards CDSS has been done since the early 1960s. Many different [20]. In general the staff was open-minded towards new research approaches have been explored, but yet the degree Information Technology (IT) systems (88 % indicated to be of impact of clinical decision support systems still remains “rather open-minded” or “open-minded”), concerning low [5]. Only few knowledge-based systems are widely CDSS, the potential benefit was assessed by the majority used day-to-day, such as automated electrocardiogram (72 %) as “rather high” or “high”, as well as prospects of (ECG) interpretation [18]. However, systems utilizing a success (53 %). At least 47 % rated reliability as well as broad variety of individual patient data had to fail due to acceptance as “rather high” or “high”. poor availability of digital data. Providing an EHR The ESGOAB project aims at encapsulating various data eliminates this obstacle [3]. sources like e.g., hospital information system (HIS), Research in the field of monitoring and analyzing vital laboratory data or drug information systems (see Figure 1) signs in Intensive Care Units (ICU) for early warning of into one integrated software system which provides one patient deterioration or sepsis were done by [15] and [12]. consistent user interface. The second main aim is the However, including microbiological findings as well as the development and integration of knowledge bases into the accurate handling of the specifics of the hematological ESGOAB system. patients remained unconsidered. The ESGOAB system is based on a 3-layer software Clinical Background architecture concept. A data collector (layer 1) is The treatment of patients with hematological diseases has responsible to capture and gather data from various existing advanced enormously in the past years. Nevertheless such sub systems which are illustrated below the three layers. infections pose a serious life-threatening risk for these Adapters for each of the sub systems are transforming data immunocompromised patients. Beside various other clinical into defined structures. The knowledge carrier (layer 2) and laboratory parameters, fever is an essential factor, analyzes incoming data from the data collector layer or which indicates a manifest or beginning infection. handles requests triggered by user interactions. The Therefore a refined assessment of the body temperature is knowledge carrier consists of various knowledge bases needed. The responsible physician has to distinguish (e.g., about drug information) which are connected between innocent fever as immunologic reaction, fever of following a modular concept. If there are e.g., new blood unknown origin, fever caused by bacteremia or the onset of values these will be evaluated by the knowledge carrier and a severe sepsis. Hereby assists the combination of lab- according hints or warnings will be immediately displayed values, microbiological findings and vital signs. The by the visualizer (layer 3) following a defined alerting emphasis and valuation of the combination of single-values concept, taking into account the importance, severity etc. of and the experience of the doctor partly determine the the alert [4]. treatment course and the outcome of the patient. Clinical studies demonstrated that the survival probability of Layer 3 Visualizer Patient X patients with sepsis depends most essentially on the period …. Dynamic Desktop … of time between diagnosis and start of effective antibiotic treatment [7]. Sepsis is not only a problem of hematological Layer 2 patients. It’s rather a challenge for the population. Severe Workflow Model Knowledge Carrier sepsis is considered to be the most common cause of death Decision Model in non-coronary critical care units. Approximately 150.000 persons die annually in Europe and more than 200.000 in Layer 1 the United States [1]. The problematic nature of a timely Data Collector Meta Data Model recognition is not that data is missing, but it is detached Adapters from one another, generated at various places and different times. The responsible physician has to link the separated information for the plurality of patients. Ward rounds, HIS Lab PACS … Sub Systems printouts with lab values, calls from the microbiologist and signs from monitors serve as instruments for this Figure 1: 3-layer architecture of the ESGOAB system connection. This work has to be done by the medical At the end of the ESGOAB project we expect an increase of personnel even in the future, but the model we present in efficiency of tasks and processes at the Department of this article offers the convincing advantage of automated Hematology and Oncology at the Heidelberg University joining of relevant data and usable presentation, resulting in Clinic (due to time-savings), a reduction of performance efficient and faster decisions. 70 APPROACH the flexibility to use the same implementation (objects) in Conceptual Requirements two kinds of settings: When deploying and operating knowledge-based systems, a 1. Interactive data retrieval with the user (module weak point is often poor practicability, in particular in terms execution) of maintenance. Either the knowledge model is implemented statically, there is no way experts of the 2. Running in background through a web service certain domain (in the present case hematologists) can retrieving data from the EHR modify the model on their own. This leads to the fact that Knowledge Model each adaptation has to be done by a software engineer. Or Sepsis accompanies with several symptoms, such as fever, the user interfaces do not provide intuitive means to modify increased heart rate, low blood pressure etc. Further the knowledge base; thus users have to be instructed and the important parameters are signs of infections such as specific system becomes error-prone. Between designing and using blood values and microbiological findings. The sequence of knowledge-based systems, a long-lasting cyclic process of appearance and the severity of these manifestations differ modeling, testing, adapting, and retesting of the core engine from patient to patient. We have to deal with fuzzy and has to be passed through. While operating knowledge-based uncertain information. However, some signs are more systems the focus shifts towards maintenance issues. important than others and certain value ranges are Maintaining the knowledge within the system is critical to supporting sepsis more than other diagnoses. So the idea successful delivery of decision support [3]. In this context was to design a decision model, which balances between a practicability plays an important role. Being aware of this set of differential diagnoses and specifies the one which can issue, easy knowledge maintenance was an important goal. be explained best by the observed findings. The set The clinical expert should be able to modify the underlying covering model is a potentially useful approach, which was knowledge model without extensive training. It has to be introduced by [13], as well as the more abstract view on simple and intuitive to use. Furthermore it should be multiple diagnose problems by [9]. Our approach is based possible to test the constructed or adapted model right on the set covering model, extended by the possibility to away. Beyond that, the knowledge model should be generic, define parameters which contradict certain diagnoses since so it might be useful for other diagnostic problems. we experienced a further need for accuracy. Knowledge Engineering Process Two sub modules based on a rule engine were required to Knowledge Engineering is the systematic approach for the handle two specific problems: development of knowledge based systems. The process may 1. Interpretation of microbiological findings: Presence of be divided into two main phases (Figure 2): knowledge an infection or suspected contamination? acquisition and knowledge operationalization [14]. It should be noted, that typically the process of acquisition 2. Interpretation of white blood cell count (leukocytes): and operationalization is not a linear process but rather a May we take this parameter into account? cyclic process characterized by continuous, iterative Implementation refinement. Initially, the basic idea was demonstrated using a prototype realized in Microsoft Excel (Figure 3). Taking advantage of Knowledge Acquisition Knowledge Operationalization quick implementation possibilities, this prototype helped us to refine our knowledge model. Capturing Design Knowledge Model Structuring Implementation Formalization Information Model Figure 2: The Knowledge Engineering Process Knowledge acquisition is the process of capturing, structuring and formalizing knowledge. The result of the acquisition phase is a knowledge model and an information model which both serve as a base for the system design of an implementation [5]. Sources of knowledge are typically Figure 3: Prototype developed with Microsoft Excel domain experts, medical literature and patient data (result presentation in the upper left corner, value repositories. Our approach was based on expert opinion and selection in the mid-area) literature research. The decision to proceed with the development of a web- Information Model based application was made due to the following main A specification of the kinds of information that were reasons: required was created, including the data format and the - Easy Accessibility from all clients via the clinic’s taxonomy. The resulting information model – which will be intranet; no need to install software on clients implemented as an object-oriented data model – provides us - Quick deployment of new versions 71 We used a XAMPP [2] installation on Windows including an Apache 2.2.14 web server and a MySQL database system. We used the scripting language PHP and the Ajax toolkit xajax [19]. The application is based on the Model-View-Controller (MVC) design pattern; the PHP code architecture follows an object-oriented approach. The application is implemented using the open source relational database management system MySQL, with use of the InnoDB storage engine. An initial database model was designed using the database-modeling tool MySQL Workbench [10]. The model was refined iteratively during the implementation of the application. Figure 5: Screenshot of a weight table (columns: diagnoses, rows: parameters and values) WEB-BASED PROTOTYPE The process of creating a new model facilitates the Application Structure following steps: The application is composed of three modules: Sepmod, Leukomod and Mibimod (Figure 4). Sepmod represents the 1. Add diagnoses generic core model, which implements the weight model as 2. Add parameters specified before and interacts with the sub modules 3. Add value ranges for each parameter Leukomod and Mibimod. Leukomod is based on a rule engine, which is responsible for the white blood cells’ 4. Create weight relations between value and diagnosis assessment. Mibimod is also based on a rule engine and 5. Define Equivalencesets performs the assessment of the microbiologic findings. 6. Define Minimalsets These steps are described in more detail below. Step 1: Firstly we need to add diagnoses, by providing its name and optionally a description. For each diagnose added, the table will be expanded by one column. Example definition: Name: SIRS Range: Systemic Inflammatory Response Syndrome Step 2: In the second step, symptoms or “parameter” can be defined, specifying its name, type, unit, value range, validity, importance. The type tells us, which kind of data we deal with, e.g. integer, float or special medical Figure 4: Structure of the application classifications like the anatomic therapeutic classification The web-front-end provides the following two main (ATC) code. The value range defines the valid value range sections: for the parameter and is used to perform plausibility checks. - Maintenance: Provides editors for each module for Regarding the process of diagnosis, an important and construction and modification of the knowledge base. relevant issue is always the time context. How long can I rely on a measured value? The answer depends on each - Module-Execution: Interactive tool, which requests for parameter. In the present model we therefore define for data or selection of options by the user and presents each parameter a time frame (validity), entering a numerical the results. value for the absolute time (minutes, hours or days) within The interactive tool is basically designed for testing this parameter remains valid or in other words we can purposes. The impact of adaptions of the knowledge base assume that the measured value may be used for the can be explored right away. diagnostic assessment. If a parameter exceeds the defined Knowledge Maintenance Section time frame, it will be ignored and treated, as it would be not available. Alternatively we can define a relative time frame Sepmod In this section the user can create a new and edit existing such like “Valid until next measurement”. Some parameters knowledge models. Created models can be loaded, saved, may have a more significant importance than other. Thus and deleted. The core element is the weight table for each parameter the importance can be defined, choosing (Figure 5). from given weights “very important”, “fairly important”, “important”. 72 Example definition: least a certain set of parameters is available. In the section minimalset the clinical experts can define, which Name: bodytemperature parameters they consider as being essential for performing a Type: float profound assessment. It is possible to define more then one Unit: °C set. Range: 30.0 to 43.0 Leukomod Validity: 6 hours Leukomod is based on a rule engine. The rules were Importance: very important defined by clinical experts. In the current version they are implemented statically and can be activated or deactivated Step 3: In the next step, for each parameter, several values through the web-front-end in the section leukomod/rules. have to be defined. The values are specified by its name Further we can adapt various parameters (like thresholds) of and its value range. the rules. For the assessment, if the white blood cells may be included, the underlying disease is of high importance. A Example definition: dynamic list of diagnosis codes (ICD, International Name: fever Statistical Classification of Diseases and Related Health Problems) can be maintained in the section diagnoses. Range: 38.0 to 43.0 Mibimod Step 4: Adding diagnoses, parameter and its values results Mibimod is based on a rule engine. A dynamic list of germs in a count(diagnoses) x count(values) matrix. For each pair which support the suspicion of a contamination can be of a value and a diagnosis we can now define a weight maintained in this section. The rules of this module are relation between a specific value and a diagnosis by currently implemented statically. clicking on the corresponding button. A weight relation is Module Execution Section the symbol for the strength a specific value supports a For testing purposes the web-front-end provides a section to diagnosis. The currently implemented model supports five run the modules. The current version has one section to run different weights, ranging from H0 (value does not support each of the modules separately and one, which encapsulates the diagnosis or even contradicts) to H4 (value strongly all three of them. These execution modules are supports the diagnosis or is even essential) depending on implemented as interactive tools, requesting each parameter the currently used weight model which can be defined in a step-by-step. separate section. For each weight symbol the value can be selected through a slide control. For each input field a check for plausibility is performed while entering data, considering two main issues Step 5: In the section called equivalencesets we define sets of previously defined parameters, which are clinically - Valid characters (depending on the data type) equivalent (Figure 6). In other words all of these parameters - Valid value range (as defined in the model). support a specific context diagnosis, such as low blood To keep track of entered values, we show breadcrumbs pressure. horizontally across the top of the input form (Figure 7). Figure 7: Entering values using the interactive mode Figure 6: Screenshot of the equivalenceset dialog The user can easily go back to previously entered values and change them if necessary. Each breadcrumb item shows In this case we would define that low blood pressure exists the name of the parameters as well as the entered value and if at least one of these blood pressure parameters takes a unit. Once all values are entered, the system performs the specific value (and a corresponding high weight) or in this assessment and presents the results, giving explanations and special case we also assume low blood pressure if a visualization of the results (Figure 8). Presenting results in vasopressors (substances which result in an increase in an appropriate way means finding a balance between (1) blood pressure) are given what we can specify through a list simple, clear and aggregated representations, supporting of ATC codes. e.g. quick task completion and (2) comprehensive and Step 6: A decision model might be sophisticated but its detailed representations, which are needed for making accuracy highly depends on the available parameters. So it comprehensible decisions. only makes sense to perform a diagnostic assessment if at 73 Elsevier Academic Press, Amsterdam; Boston, 2007. 6. Kohn, L.T. (Ed.). To err is human. National Academy Press, Washington, DC, 2000. 7. Kumar, A., Roberts, D., Wood, K.E., et al. Duration of hypotension before initiation of effective antimicrobial therapy is the critical determinant of survival in human Figure 8: Result presentation (diagram and textual septic shock. Critical care medicine 34, 6 (2006), explanation) 1589-1596. DISCUSSION & FUTURE WORK 8. Lo, H.G., Matheny, M.E., Seger, D.L., Bates, D.W., We identified two important and promising factors that will and Gandhi, T.K. Impact of non-interruptive help overcome key barriers limiting more widespread use of medication laboratory monitoring alerts in ambulatory CDSS: Firstly we did not experience the “physician care. Journal of the American Medical Informatics resistance” using decision support systems; they are rather Association 16, 1 (2009), 66–71. convinced of their usefulness as described in the introduction. Secondly we can take advantage of the 9. Miller, R.A. A Heuristic Approach to the Multiple software infrastructure we presented which has the potential Diagnoses Problem. Proceedings of the 6th to improve clinical processes and to integrate and interact Conference on Artificial Intelligence in Medicine in with sharable knowledge modules conceived to support the Europe, Springer-Verlag (1997), 187–198. physicians in their decisions. We also introduced a web 10. MySQL Workbench visual tool. Available at based prototype of a knowledge module for early detection http://www.mysql.com/products/workbench/. of sepsis. We believe to contribute important elements to 11. Osheroff, J.A., Teich, J.M., Middleton, B., Steen, E.B., lift computer-aided decision support into widespread Wright, A., and Detmer, D.E. A roadmap for national practice. action on clinical decision support. Journal of the However, our approach still needs to be refined and American medical informatics association 14, 2 evaluated. It has to be shown that our knowledge modules (2007), 141-145. are able to provide accurate and traceable support, 12. Pilz, U., and Engelmann, L. Früherkennung von characterized by high sensitivity. A test concept as well as septischen Krankheitsverläufen durch Einsatz eines test scenarios will be defined. Further alerting concepts will Expertensystems. Intensivmedizin und Notfallmedizin be tested in practice since empirical studies have shown that 43, 7 (2006), 575-579. too many generated alerts could lead to “alert fatigue”, whereas “non-interruptive” alerts only have low impact and 13. Reggia, J.A., Nau, D.S., and Wang, P.Y. Diagnostic are not effective [8]. Expert Systems Based on a Set Covering Model. International Journal of Man-Machine Studies, ACKNOWLEDGMENTS (1983), 437-460. This work was funded by the German Federal Ministry of 14. Spreckelsen, C., and Spitzer, K. Wissensbasen und Education and Research under grant number Expertensysteme in der Medizin. Vieweg + Teubner, 01 IS 09027 C. The responsibility for the content of this Wiesbaden, 2008. publication lies with the authors. 15. Tarassenko, L., Hann, A., and Young, D. Integrated REFERENCES monitoring and analysis for early warning of patient 1. Angus, D.C., Linde-Zwirble, W.T., Lidicker, J., deterioration. British Journal of Anaesthesia 97, 1 Clermont, G., Carcillo, J., and Pinsky, M.R. (2006), 64-68. Epidemiology of severe sepsis in the United States: analysis of incidence, outcome, and associated costs of 17. Warner, H.R., Sorenson, D.K., and Bouhaddou, O. care. Critical care medicine 29, 7 (2001), 1303-1310. Knowledge engineering in health informatics. Springer, New York; Berlin; Heidelberg, 1997. 2. Apache Friends. XAMPP distribution. Available at http://www.apachefriends.org/xampp.html. 18. Willems, J.L., Abreu-Lima, C., Arnaud, P., et al. Evaluation of ECG interpretation results obtained by 3. Bates, D.W., Kuperman, G.J., Wang, S., et al. Ten computer and cardiologists. Methods of Information in commandments for effective clinical decision support: Medicine 29, 4 (1990), 308-316. making the practice of evidence-based medicine a reality. Journal of the American Medical Informatics 19. xajax Project. xajax framework. Available at Association 10, 6 (2003), 523–530. http://www.xajax-project.org/. 4. Durieux, P. Electronic Medical Alerts – So Simple, So 20. Zuehlke, D., Meixner, G., and Klein, U. Smart Medical Complex. New England Journal of Medicine 352, 10 Software Systems For Dummies? The Case For A (2005), 1034-1036. User-Centered Systems Design. Proceedings of the 3rd International Conference on Health Informatics 2010 5. Greenes, R.A. (Ed.). Clinical decision support. (Valencia ESP, 2010), 350-354. 74