Data summarization method for chronic disease tracking

Journal of Biomedical Informatics 69 (2017) 188–202

Contents lists available at ScienceDirect

Journal of Biomedical Informatics journal homepage: www.elsevier.com/locate/yjbin

Data summarization method for chronic disease tracking Dejan Aleksic´ a, Petar Rajkovic´ b,⇑, Dušan Vucˇkovic´ b, Dragan Jankovic´ b, Aleksandar Milenkovic´ b a b

University of Nis, Faculty of Science and Mathematics, Department of Physics, P.O. Box 224, 33 Visegradska, 18000 Nis, Serbia University of Nis, Faculty of Electronic Engineering, Laboratory for Medical Informatics, 14 Aleksandra Medvedeva, 18000 Nis, Serbia

a r t i c l e

i n f o

Article history: Received 23 January 2017 Revised 10 March 2017 Accepted 17 April 2017 Available online 19 April 2017 Keywords: Medical information system History of disease summarization Chronic disease management System usage scenarios

a b s t r a c t Objectives: Bearing in mind the rising prevalence of chronic medical conditions, the chronic disease management is one of the key features required by medical information systems used in primary healthcare. Our research group paid a particular attention to this specific area by offering a set of custom data collection forms and reports in order to improve medical professionals’ daily routine. The main idea was to provide an overview of history for chronic diseases, which, as it seems, had not been properly supported in previous administrative workflows. After five years of active use of medical information systems in more than 25 primary healthcare institutions, we were able to identify several scenarios that were often end-user-action dependent and could result in the data related to chronic diagnoses being loosely connected. An additional benefit would be a more effective identification of potentially new patients suffering from chronic diseases. Methods: For this particular reason, we introduced an extension of the existing data structures and a summarizing method along with a specific tool that should help in connecting all the data related to a patient and a diagnosis. The summarization method was based on the principle of connecting all of the records pertaining to a specific diagnosis for the selected patient, and it was envisaged to work in both automatic and on-demand mode. The expected results were a more effective identification of new potential patients and a completion of the existing histories of diseases associated with chronic diagnoses. Results: The current system usage analysis shows that a small number of doctors used functionalities specially designed for chronic diseases affecting less than 6% of the total population (around 11,500 out of more than 200,000 patients). In initial tests, the on-demand data summarization mode was applied in general practice and 89 out of 155 users identified more than 3000 new patients with a chronic disease over a three-month test period. During the tests, more than 100,000 medical documents were paired up with the existing histories of diseases. Furthermore, a significant number of physicians that accepted the standard history of disease helped with the identification of the additional 22% of the population. Applying the automatic summarization would help identify all patients with at least one record related to the diagnosis usually marked as chronic, but ultimately, this data had to be filtered and medical professionals should have the final say. Depending on the data filter definition, the total percentage of newly discovered patients with a chronic disease is between 35% and 53%, as expected. Conclusions: Although the medical practitioner should have the final say about any medical record changes, new, innovative methods which can help in the data summarization are welcome. In addition to being focused on the summarization in relation to the patient, or to the diagnosis, this proposed method and tool can be effectively used when the patient-diagnosis relation is not one-to-one but many-to-many. The proposed summarization principles were tested on a single type of the medical information system, but can easily be applied to other medical software packages, too. Depending on the existing data structure of the target system, as well as identified use cases, it is possible to extend the data and customize the proposed summarization method. Ó 2017 Elsevier Inc. All rights reserved.

⇑ Corresponding author at: University of Nis, Faculty of Electronic Engineering, Laboratory for Medical Informatics, 14 Aleksandra Medvedeva, Lab 534, 18000 Nis, Serbia. E-mail addresses: [email protected] (D. Aleksic´), [email protected] (P. Rajkovic´), [email protected] (D. Vucˇkovic´), [email protected]. ac.rs (D. Jankovic´), [email protected] (A. Milenkovic´). http://dx.doi.org/10.1016/j.jbi.2017.04.012 1532-0464/Ó 2017 Elsevier Inc. All rights reserved.

1. Introduction and motivation Describing chronic medical conditions and documenting them properly is one of the key Electronic Health Record (EHR) functionalities enabling an effective primary care [1,2]. When the data have

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

been collected from different sources, or by following different procedures, the data aggregation must be introduced [3] and accompanied by the adequate tracking tools. The ways of tracking chronic diseases constitute an important part of sustainable healthcare and this is classified as a long-term healthcare tracking [4]. As it is recommended in [5], HER-based system should collect data about various events related to chronic (and generally to any other) diseases and allow users to continuously track the health status of the patient. In addition, different trending tools and summarizing methods can be developed to ease medical professionals’ operations and allow a further data mining [6,7]. The data summarization is especially important when new patients suffering from chronic diseases need to be identified [8]. Its importance is further emphasized by the fact that, in many countries, the percentage of patients with chronic diseases is around one half of the total population [9]. This large number of patients makes chronic diseases an important social issue. From the medical point of view, such prevalence categorizes certain diseases as epidemic. A closer look reveals that one of the most frequent chronic diseases is hypertension. It affects 35–45% of the world population [9]. A similar situation is noted in the Republic of Serbia, where the most frequently diagnosed chronic disease is high blood pressure [10]. Over the past few years, after the introduction of the Medical Information System (MIS) to primary healthcare in the Republic of Serbia, opportunities for the chronic disease tracking have emerged [11]. Primary healthcare centers are organized on municipal level, and almost all of them have the MIS in place. Installed MISs are successfully used for obtaining data and reporting them to interested parties, such as the Ministry of Healthcare (MoH) and medical insurance funds (HMIFs). Once uploaded, data are processed and the national level statistics are generated. Also, medical institutions create their own reports and statistics. From a medical practitioner’s point of view, the use of the MIS in daily operations offers several functionalities that can help in the chronic disease tracking. The MIS users are informed about certain guidelines during their training that, if properly followed, produce very clear patient-level reports with abundant information. Unfortunately, not all users consider the mentioned guidelines important enough and consequently, several different data collection routines have been observed. These aberrations reduce the transparency of patient-level reports, and data are being scattered across different categories. Management-level reports that require extraction of full data summaries for a specific patient can then produce false positives. This is in line with the findings presented in [12]. As it has been stated in [12], the introduction of the EHR-based MIS into primary healthcare has its positive and negative effects. In this case, the positive side is the fact that the volume of the data related to chronic conditions are properly collected and registered. This contributes to the creation of a valid general statistics report from a medical institution. Now, an additional effort is needed to improve the data structure and upgrade personal-level reports. The development group managed to pinpoint five distinctive routines that the MIS users had in the data entry process. Some of them were in line with general guidelines, but others were the product of the non-envisaged software use, generating the data that could not be easily linked to the main history of disease (HoD) for a specific diagnosis. This is why an algorithm was defined to help with the data summarization and enable all existing aggregation tools to display a much cleaner data spectrum. The proposed method will not introduce any changes into the existing data, but rather build an appropriate data structure around it. In this way, no data will be changed or structural integrity corrupted, which should ensure the end-user’s satisfaction with the quality of reports.

189

The data collected in the Nis Primary Healthcare Center were used for this research. Its MIS had more than 600 active users in the monitoring period, covering the population of more than 200,000 patients from the Nis metro area. The Medical Information Systems research group from the Faculty of Electronic Engineering, Nis, Serbia, had been actively involved in the MIS development and deployment since 2002. In 2008, the development of the EHRbased MIS [11] started. This was predominantly aimed at primary and ambulatory care centers [13]. From the beginning of 2016, the system has been in use by 31 healthcare institutions, with user bases ranging from 20 to 600 medical professionals per institution. 2. Related work It is common knowledge that patients suffering from chronic diseases usually visit their doctor more often [14]. Since their data are collected from several departments, and since the MIS users have different data entering routines, there are instances where some of the data related to some chronic disease are not properly structured and not easily accessible. As it has been presented in [15], the data related to chronic diseases must contain an additional set of internal relations in order to increase the chances of being accepted by medical professionals. It is important to mention that doctors spend a lot of time reviewing medical charts. According to [18], the time needed to do this varies from a few minutes to half an hour. When data are not properly structured and displayed, this time increases, thus raising the chances of overlooking some of the important facts. As it has been stated in [18] ‘‘properly structured data leads to a reduction in treatment time vs. increased unproductive documentation-gathering time”. Although this conclusion is valid for nephrologists, it is a rather generally applicable statement. Therefore, we decided to extend the existing data structures by introducing a minimal set of entities and relations. Consequently, the end-users will get one additional set of overviews and reports that will eventually help them when important decisions regarding the medication process should be made. Applying the summarization not only to a single patient, but rather to specific social groups, can trigger preventive actions and health campaigns. The importance of such campaigns is discussed in [16,17]. Notwithstanding the fact that the collected data can give meaningful results, the weak spot is the data collection based on questionnaires and surveys which can be ‘‘polluted” by subjective answers. This fact induced us to move on in a different direction. Our data summarization approach is based only on the analysis of the data stored in the MIS, and thus, the obtained results are expected to be more objective. In the Primary Care Center Nis, we identified the doctors that effectively used the available features and those that checked patients’ printed records and used the MIS to enter data and print one-page-long summaries after the visit. The data summarization proposed in the next section is intended to improve their daily routine. Considering the available reference materials, there are many articles describing different data summarization methods and tools often connected to chronic diseases. Some state-of-the-art summarizing tools are presented in [19,20]. The tool presented in [19] named HARVEST is a Web tool basically designed as a table view. The tool [19] supports also the timeline style overview. The concepts of table-based and timeline overviews are also applied in our summarization tool. Unlike HARVEST, the summarization engine developed by our research group is part of the Windows-based rich client. The main summarization overview is also a table view, but supported by a multi-leveled, strong filtering support. Our timeline is, by default, focused on the main events, such as a therapy change. Thanks to that approach, our summarization tool does not have the data cloud

190

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

problem, which is present in HARVEST as a consequence of a higher number of displayed events. The main advantage of HARVEST is a much higher level of generalization. HARVEST is able to parse HL7 messages, whereas our summarization tool is a homebrew solution customized only for one specific MIS and developed by our research group. MedWISE [20] presents another cutting-edge reporting tool. Compared to HARVEST and the home-brew reporting tool, its advantage lies in a high level of configurability. Users can define their own layout using widgets. Widgets are active components aimed at retrieving data, and users are, thus, allowed to define layouts. This approach could lead to users being flooded with different, non-distinctive data in the form of different reports, and could also generate a potentially high network traffic volume. For this reason, we decided to offer only a default set of summarization overviews and reports, which created a balance between the amount and representativeness of the presented data. When defining a summarization strategy, our research and development team started with the AORTIS as the theoretical model [21]. Our intention was to cover each of the phases – aggregation, organization, reduction, transformation, interpretation and synthesis, and develop them according to the guidelines. To this end, the work presented in [22] was marked as important since it classified summarization approaches and pointed out ‘‘open challenges critical to the implementation and use of a robust EHR summarization system”. The main goal of introducing an additional summarizing method was to improve the readiness of the extracted data and give the opportunity to the users of our system to have an effective tool, which is easy to use and contains a well-balanced set of configurable properties. The work presented in [23] gave us a good overview of the existing standards and helped us avoid common weak spots. In developing a summarization routine, we had two main objectives: To gather the already collected data and to allow the medical practitioner to have a better overview of medication processes. This is aimed at registered patients suffering from chronic diseases; To help in identifying potentially new patients suffering from chronic diseases and schedule all necessary diagnostics and therapeutic treatments. Besides helping physicians with an earlier identification of new patients suffering from a chronic disease, the results generated by our summarization tool can be also used in further clinical studies. In this way, medical researchers will get a tool that can help them in organizing more detailed studies, such as the one presented in [24]. We plan to continue our research by introducing the summarization not only of the structured data, but also of the free text entered as the medical history. To achieve this goal, the largescale participation of medical professionals is needed. As it is stated in [25], ‘‘to realize this, it is necessary to build models of how physicians read and understand unstructured free-text in medical records.” Usually, summarization methods and tools are focused on the entire EHR of a patient, dealing with a large number of records and offering to the user a layout full of various medical data. In an attempt to satisfy the need of our users and keep the user interface simple and clear, our summarization method tends to be a bit more effective when maintaining its focus mainly on a specific diagnosis for a selected patient. The presented data is then initially filtered, and only the ‘‘main events”, such as a therapy change or important procedures, are displayed. The user can later, as with other reporting tools, apply or remove additional filters in order to tailor the content to be viewed. Another advantage of our approach is a routine that can help physicians in identifying patients that potentially suffer from a chronic disease. Reporting

tools are usually focused on the data retrieval and display, but when implementing a certain notification subsystem, they commonly suggest missed medical procedures and alert values outside the reference range. 3. Materials and methods As mentioned before, the research is based on the data collected in the Nis Primary and Ambulatory Healthcare Center. The data were collected from GPs along with thirty different specialists treating more than 200,000 inhabitants of the town of Nis. Together with therapeutic and laboratory departments, its MIS registers more than 2.5 million medical services on an annual basis. The MIS has been in use since mid-2011. This research takes into account the data collected from January 1st 2012 to December 31st 2015 (Table 1). A significant percentage of the registered data was related to chronic diseases. For example, almost one quarter of all medication prescriptions were related to high blood pressure (IDC-10 code I10). In order to support an effective tracking of chronic diseases, in a format of a HoD, we introduced a basic set of functionalities followed by the intended usage scenario. After analyzing the statistics for the mentioned period, we realized that there was a significant number of records related to chronic diseases, but a comparatively low percentage of them were structured in the way suggested. Since the end-users did not follow the suggested mode of use, we drilled down the retrieved data and identified the representative data collection scenarios. Besides the suggested usage scenario, we identified four additional usage scenarios that brought the data to the EHR, but were not connected to the supposed HoD structure. For this reason, we proposed the extension of the data structure that would allow a later data summarization. The extended data structure was followed by the appropriate summarizing method and tools. 3.1. Intended usage scenario for chronic disease registration and tracking Upon a doctor’s admission of a patient for either a medical examination, or therapeutic treatment or diagnostics session, all of them will be registered as a separate item, based on the medical services provided (according to the Serbian MoH). The information on a patient collected during one of medical treatments can be grouped into items called ‘‘visits” (Fig. 1). For each visit, the medical practitioner can define the primary diagnosis. Several visits can be linked together as one primary diagnosis, hence creating the HoD for that patient. In the past, when medical documentation was kept only in the paper form within the primary health system, the HoD was tracked only in some specialist departments. Now, with the help of electronic health records (EHR), general practitioners can track the HoD for any disease. This is particularly important for chronic diseases. In order to mark one HoD as chronic, the medical professional just needs to tick a single check box when entering the data about a new visit. By marking the initial visit in the HoD as chronic, the user chart opens some additional features. When one diagnosis is marked as chronic, the users will get a special view of the chronic diagnosis with the timeline of medication and all connected documents. This could be helpful for many physicians, but, within the existing system of medical care, just few of them use actively the mentioned feature. The research presented in [12] shows a systematic review of the general impact that the EHR-based systems have on the primary healthcare. The bottom line is that MIS systems improve the working environment, but not significantly. Some of the system’s

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

191

Table 1 Overview of an overall data volume by the most important items. Item

Overall

In 2015

In 2014

In 2013

In 2012

Number of active users Number of active patients Registered patient visits Prescriptions Requests for specialist examination Various registered medical services

669 209,809 9,871,575 5,523,545 957,336 18,087,511

476 185,974 2,684,592 1,692,605 295,310 5,371,467

458 184,681 2,668,763 1,586,891 248,503 5,215,960

495 187,026 2,719,009 1,392,569 238,008 4,887,458

411 172,705 1,799,211 851,480 175,515 2,612,626

Fig. 1. Visit form (1 – administrative data, 2 – main diagnosis, 3 – additional diagnoses, 4 – basic medical data, 5 – anamnesis, 6 – advice and overall therapy, 7 – prescribed therapies, 8 – requests for further medical examinations).

features can be evaluated only as potentially effective. Due to the ‘‘lack of understanding of potential benefits [12],” they are just not utilized to their full capacity. This is exactly the case with the support of the chronic disease tracking. We evaluated the initially offered functionality and offered the data summarization tool as an additional benefit. We wanted not only to improve the chronic diagnosis tracking and identifying, but also to enrich the toolset offered to the users with the option of defining and storing the data aggregation definition. 3.2. Statistics on the usage of current MIS features for chronic diseases Analyzing the extracted data, the total of 63,525 initial visits marked as chronic were identified. These visits were connected

to 21,929 patients. The total of 155,235 chained visits were linked to the mentioned initial visits related to chronic diseases. This indicated that, after the HoD for chronic disease was created, only 2–3 later visits were about to be linked to the initial one. This average indicated that the existing functionality was just sporadically used and was not accepted in the desired extent by the users. Another discouraging fact, in support of the decision to introduce a summarizing method, was the actual number of physicians using the existing chronic disease-related feature. A total of eight doctors created 50% of all records, while the other 27 users generated the remaining 30% of records. Considering the yearly statistics shown in Table 2, the number of newly registered patients was the highest in the first year of the information system’s active use. In the years to follow, the number of newly registered patients slightly rose

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

192

Table 2 Overview of the usage of the chronic diagnosis feature with a number of newly registered patients, initial and chained visits per year. Category Number of Initial Chained Chained in Chained in Chained in Chained in

new registered patients

2015 2014 2013 2012

Overall

2015

2014

2013

2012

21,929 63,525 155,235 42,764 40,101 42,363 30,007

4998 13,376 13,341 13,341 n/a n/a n/a

4699 14,524 20,148 8658 11,490 n/a n/a

4425 11,868 23,089 4837 9206 9046 n/a

16,175 23,757 98,657 15,928 19,405 33,317 30,007

from 4425 to almost 5000. The number of initial visits was also the highest in 2012 and varied in the following years. On the other side, the number of chained visits was the lowest in the first year, but it rose to 42,363 in 2013 and remained steady (40,101 in 2014 and 42,764 in 2015). Nevertheless, the doctors that accepted this functionality created detailed histories of diseases that could be used for further clinical studies. In 2015, there were still open chronic HoDs from 2012, when almost 16,000 visits were recorded, which further emphasizes the fact that there is no standing regulation on how long one chronic visit should remain open. We can safely say that there is a usual practice of opening the HoD and then keeping it open indefinitely. In some cases, records were kept open for the time periods of one to two years. Table 3 displays the statistics on five diagnoses most often marked as chronic. Looking at the number of visits marked as chronic, it is evident that the total number of visits marked as chronic is in each of the five cases significantly lower than the total number of visits attributed with the same diagnosis (at least six times). This is solid ground for developing an EHR summarization method that should help medical practitioners in mutually linking all of the medical records related to the same disease. 3.3. Identified usage scenarios The main items used for structuring medical data are visits. They are considered as top entities, and they can be chained together forming the HoD. A visit can include a set of defined medical services. Each medical service then functions as an envelope containing all defined medical documents and prescriptions. It is important to say that a diagnosis can be assigned to any of previously defined entities. The users can assign one diagnosis to each visit, but then they create prescriptions or a document related to some other diagnosis. The common scenario is the following: a patient having some respiratory infection comes to the doctor’s and starts describing symptoms related to hypertension, the doctor then registers the visit and associates it with i.e. J20 Acute Bronchitis, but at the same time the doctor creates, under the same visit, a request for further examinations connected to I10. Analyzing the MIS software usage, we identified five different ways in which the data about chronic diseases were entered into the system. Medical professionals used to follow some of the following scenarios (Fig. 2): Scenario 1: Flagging the initial visit as ‘‘chronic” and using the HoD with additional functionalities dedicated to chronic diagnoses. Scenario 2: Creating the HoD for a diagnosis, but without marking it as chronic. Scenario 3: Creating independent visits associated with a chronic diagnosis. Scenario 4: Creating requests for a further medical examination under the visit associated with a different diagnosis. Scenario 5: Creating prescriptions for a chronic diagnosis under the visit associated with another diagnosis.

The first of the above-mentioned scenarios represents the desired case in which all HoDs to be joined are marked as chronic. The resulting HoD will contain all created medical records from joined items without any addition or reduction. Also, it will cover a longer period of time, enabling a medical practitioner to have a long-term overview of the patient’s condition and correlation with the medication scheme changes. The downside of this approach, from the technical point of view, is that the HoD assigned to one diagnosis can contain sub-items (such as prescriptions) related to other diagnoses. The example is shown in the top-left sub-figure of Fig. 2. The items not related to the main diagnosis are painted yellow1. The very existence of the items associated with other diagnoses proves that the derived HoD can be ‘‘polluted” with non-directly related items. For this reason, an additional option to clear the derived diagnosis is introduced. It removes the items that are attributed to other diagnoses from the joined HoD. On the other side, some medical practitioners tend to be unclear regarding the derived diagnosis, since they think that displaying data about other illnesses is valuable. Currently, we are not able to identify which is the desired system behavior for our users, but we expect to get more comprehensive results in the future. The next two scenarios (number 2 and 3) represent the situations in which the doctor creates the HoD, or just a single visit record, for diagnostic purposes, but does not mark them as ‘‘chronic”. The medication is associated with a diagnosis which is usually assumed as chronic (such as I10 in the example). In this case, in the regular MIS usage circumstances, including both HoDs and the independent visits into the summary would be the same as for the scenario 1. The complete data structures behind single visits and HoDs will be just joined to the summary HoD. The interesting fact for a future research is that HoDs are mostly created and maintained by a single medical practitioner, while introducing the single visits (scenario 3, top-right on Fig. 2) in many cases indicates that the visit was conducted by a different doctor (mostly in cases when the patient’s physician is unavailable for some reason). In both of the scenarios, 2 and 3, it is possible to find documents and/or prescriptions related to different diagnoses just as in the scenario 1. Creating independent visits for a diagnosis is an old routine that many general practitioners still follow even though they have already switched to the EHR. Usually, they create one visit for one diagnosis and mark it as finished immediately after all data have been entered (scenario 3 in the list). If the patient already has an associated chronic diagnosis, then the mentioned cases will be included. But, when the patient does not have any chronic diagnosis associated, it is not completely correct just to include visits in the chronic disease’s HoD. For example, if the patient has three visits associated with the I10 diagnosis, it must be checked what the time span between the visits is and, using some of the heuristic methodologies, offer the doctor a selection of potential chronic diagnoses. 1 For interpretation of color in Fig. 2, the reader is referred to the web version of this article.

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

193

Table 3 Statistics for five most frequent chronic diagnoses – I10 Hypertensio arterialis essentialis (primaria), E11 Diabetes mellitus ad insulino independens, E10 Diabetes mellitus ab insulino dependens, I20 Angina pectoris, I49 Arrhytmiae cordis aliae. Category

I10

E11

E10

I20

I49

Initial chronic visits Chained chronic visits Total visits marked as chronic Total visits Requests for specialist examinations Total prescriptions

20,892 51,230 72,122 531,483 25,465 2,464,136

3727 10,178 13,905 112,1504 5106 335,726

1770 5517 7287 76,602 2495 130,674

1737 4366 6103 34,402 2777 222,888

1070 1693 2763 37,907 6194 109,901

Fig. 2. Illustrative scheme of the data created by common MIS usage scenarios – top left: scenario 1 – a medication associated with a chronic diagnosis and defined as HoD and marked as ‘chronic’, bottom left: scenario 2 – a medication associated with a chronic diagnosis and defined as HoD but not marked as chronic, top right: scenario 3 – a single visit associated with a chronic diagnosis, mid right: scenario 4 – a medication/visit associated with an acute diagnosis having a chronic-related document, bottom right: scenario 5 – a medication/visit associated with an acute diagnosis having a chronic-related prescription.

As it has been mentioned, the data collected in the scenarios 1, 2 and 3 can contain ‘‘background noise” consisting of the items registered under the chronic diagnosis, but related to some other medical condition. The set of documents and prescriptions representing ‘‘noise” will be highlighted in the summary, and the user will be able to remove them. It is important to point here that prescriptions will be removed only from the summary, but not from the EHR and they will stay connected to the original visit. On the other hand, the records related to a chronic diagnosis can appear in the visit associated with another diagnosis. These cases are covered in the scenarios 4 and 5 (mid-right and bottom-right in Fig. 2). The system will extract the records related to the chronic diagnosis, create a virtual visit, link the extracted records to the virtual visit and include the virtual visit into the summary HoD. As regards the category of patients not still diagnosed with a chronic disease, a good indicator of a chronic condition could be the recurrent requests for further medical examinations related to some of the chronic diagnoses (scenario 4). Since they can be created under the visit with the additional

diagnosis, they will be linked to the chronic disease HoD. It is similar with medical prescriptions – scenario 5. The criteria for suggesting that a patient have a chronic diagnosis could be the lapse of time between the issuing of the two prescriptions, but in that case, the user should be the only decision maker, with no automation implemented. 3.4. Proposed data structure extension The original database structure dedicated to storing data related to patients’ visits is represented by the entities marked with blue header in Fig. 3. The data table Visit is referred to as a top entity. It contains basic pieces of information such as the main diagnosis and visit type. Under the visit, the set of associated medical services can be listed. The table MedicalService has VisitId as a foreign key which connects each separate medical service to a single parent visit. Furthermore, each medical service acts as a parent entry for a range of different medical service items. The medical service item is any entity created during the process of the medical service

194

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

Fig. 3. Proposed data structure extension to support summarization.

storing data about a treated patient – from temporal diagnoses, via prescriptions and reports, to requests for further examinations and therapeutic treatments. Each of the mentioned entities subordinated to the medical service has a medical service id as a foreign key, storing the id of the medical service under which it has been created. On the other side, users regularly need a possibility to link some of the existing medical items to a newly created visit. In order to support this kind of requests, the table MedicalServiceItem, acting as a junction table between MedicalService and the already mentioned subordinated entities, such are Document or Prescription, is introduced. This means that when a new lab analysis report is created, one medical service item is created by default – having one foreign key set to the prescription’s parent medical service and another foreign key set to the prescription’s own id. Now, this lab report can be used on the next occasion when the hospitalization request gets created. The user can then link the existing lab report to a new hospitalization request by creating a new entry in MedicalServiceItem table. The possibility to link the existing medical service items to new medical services without jeopardizing the existing data structures was the starting point in the summarization structure development. We decided to use as much as possible of the existing data structures with minimal structural updates. This request is also driven by the fact that there are more than 30 running instances of the existing system, and any significant update of the existing structure would lead to unwanted maintenance tasks.

For this reason, our main approach was to extend the data structure and leave the existing tables unchanged. In order to enable the data summarization without the intrusion of the existing database structure, new data tables are introduced (Fig. 3) – SummaryHeader and SummaryHeaderVisit. The top level in the summarization structure is named SummaryHeader. It contains a reference to a diagnosis, patient and medical professional that created the summarization (field CreatedById), as well as the fields for basic timestamps. In this way, each MIS user can define their own summarization for any patient. In addition, references to a diagnosis and patient are set as nullable, which offers the opportunity for the user to create more complex summarizations related to more diagnoses and to more patients at the same time. The connection between the summarization header table and the existing data structure starting with the items of the visit type is a junction table named SummaryHeaderVisit. It has references to the parent summary header and to the connected visit. This allows one of the visits to be included in more than one of the summarization processes. The described process explains the connection of a complete visit with all of the appurtenant entries and with a summary. In some cases, it is necessary to include only some documents or prescriptions into the summary. Since it is not allowed to alter the initially entered data, placeholder items for visits and medical services will be created to ensure a connection to the summary. The existing prescription or document will be connected to a virtual medical service as its own medical service item. Since the table MedicalServiceItem functions as a junction table between

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

the medical service and various medical documents, one document can be chained to more than one medical services (Fig. 3). The described placeholder or virtual items will be assigned a special entry type and will not appear in the regular EHR overviews. When one must remove some document associated with a different diagnosis from a visit assigned to a chronic diagnosis, the copy of the original diagnosis containing only the required documents will be created. The carbon copy of the visit along with all the assigned medical services will be marked as a placeholder (or virtual), and will be visible only from the summary. All of the new data created during the summarization process are stored within the data tables shown in Fig. 3. The primary target for the data created during the summarization process are two new tables, while, occasionally, tables Visit, MedicalService and MedicalServiceitem will be used to store additionally created placeholder entities. As it has been described, new tables consist mostly of ids and timestamps in order to minimize the volume of the needed additional data. Also, indexing on foreign keys helps in improving the performance of select statements and makes them easy to write. For example, the select statement used to retrieve all of the patients identified as potentially chronic for diagnosis I10 after October 24th, 2016 would look like: Select pt.⁄ From SummaryHeader sh join Diagnose d on sh.MainDiagnoseId = d.Diagnoseid join Patient pt on pt.PatientId = sh.PatientId Where d.Code = ‘I10’ and sh.CreatedTs > ‘2016–10–24’. The summarization tool is implemented as a separate system module, so the initial system features are not affected. All of the existing routines will continue to work as before, only new options will become available for the end users. Since it is developed as an extension of the existing MIS [11], it can interact through common architecture with other parts of the system. Currently, the data replication and reporting are supported, but in the near future, many other standard features, such as archiving and exchange with other systems, will be enabled. 3.5. Summarizing method and tool Our summarizing method is envisaged to be used as the basis for a summarizing tool. It could be used as a suggestive mechanism that will propose to the doctor to chain potential records together in a single HoD. An additional option would be splitting HoDs into specific important events, such as therapy changes. Doctors can define HoDs and keep them open as long as they find it fit. At some point, a physician can close an existing HoD and create a new one. This is often the case when the therapy is changed or after some time (six months to a year). We must underline here that the data summarizing will not change the data initially acquired, but it will create a structure around that to help in the summarization process. According to all of the mentioned MIS usage scenarios, summarizing will start with defining a diagnosis for the selected patient. The summarizing tool will then access the database and identify all of the patient’s visits related to the specified diagnosis. Furthermore, all the documents and prescriptions that do not belong to the previously extracted visits, but are associated with a specified diagnosis, will be extracted. The user will be promptly introduced to the overall statistics and has to confirm if they want to proceed with the summarization. Fig. 4 depicts the summarization routine. For extracted visits, the user has to decide whether connected documents associated

195

with diagnoses, other than the one selected, should be added to a summary. If this is the case, then the extracted visit will be just connected to the created summary. If not, the placeholder entities of medical service and visit will be created and associated with the defined summary header. The placeholder entities are derived from the original ones, as shallow copies, containing all of direct properties and will appear only in the summary view. None of the existing data overviews will be influenced by them. After the summarization process has been finished, the summary header will be associated with several entities connecting the summary header and visits. These entities are links to the existing data and users can enable/disable them from the summary, sort them by date of visit, or just arrange them according to their needs. Once the summarization for a patient and diagnosis is established, the user can choose to automatically connect all newly added entities to the summarization. During each successive visit, a document or prescription for the selected patient is added and connected to the summarization, i.e., if there is a link to a correct diagnosis. The summarization tool consists of several software components developed as an extension of the existing MIS (Fig. 5). It includes the previously described data structure extension affecting both the database and data access layer and the required extension module for the base EHR service. Summarization routines and the related business logic are part of the summarization service. It interacts through the EHR business logic with a core system and can trigger alerts, as well as initiate a report creation and invoke necessary data synchronizations with other information systems within the MoH. On the client layer, two additional plugins are implemented. One is the extension for medication overview and visit registration plugins intended to be used by medical professionals. Another is a configuration plugin used to define the behavior of the summarization tool for both the overall system and separate user accounts. The configuration plugin is intended to be used by the medical institution management and technical staff. The summarization tool can run on-demand or as a background process. If started on-demand, it will show results and ask the user for action. As far as newly discovered chronic patients are concerned, it will prompt the user with the question in order to establish the summary for the diagnosis. On the other hand, if it works as a background process, it will analyze the patient’s data while its EHR overview is open and generates an alert if needed. The alert will be generated if some new non-connected visit, document or prescription is discovered. Then, the user can decide if they have to be added to the summary. For newly discovered patients, the alert condition must be carefully defined in order not to generate too many alerts that will lead to false positives. In this case, users can define the number of discovered occurrences and the time lapse between the first and the last discovered entry. For example, it could be inconvenient to raise an alarm for any patient having just one visit connected to the diagnosis I10. But, defining the filter condition to ‘‘more than three requirements” or ‘‘at least two visits in 30 days” should produce better results. Judging by the AORTIS model [21], the initial data aggregation is already supported through the existing data structure with the routine of chaining visits and forming HoDs. Additional data structures are used to support a further data aggregation, and to combine the data stored in the already created HoDs with the additionally linked visits. After the data summarization, timestamps remain the primary key for the data included. While working with data summaries, users have full control of the summarized items and can decide which ones to keep and which ones to remove from the final output.

196

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

Fig. 4. Summarizing routine – A: handling the visits associated with a selected diagnosis, B: handling the documents associated with a selected diagnosis but found in the medication associated with other diagnoses.

4. Results In order to check the possible benefits of the summarizing method, the data collected in the Nis Primary and Ambulatory Care Center between January 1st 2012 and December 31st 2015 were analyzed. The efficiency of the presented method was tested on five diagnoses that were most often defined as chronic in the mentioned database. We checked the results of the actual use of ondemand data summarization initiated by MIS users (Table 4), as well as potential effects if the automatic summarization was used (Tables 5 and 6). Looking solely at the usage of HoDs flagged as a chronic disease (scenario 1), it is visible that only 11,784 patients were affected. Since the total number of registered patients is 209,809 (Table 1), this makes only 5.6% of population. Although this looks very optimistic, it is unfortunately untrue. With this piece of information, a

conclusion can be drawn that the users of the MIS did not recognize all the benefits they could get from the specialized data overview set for chronic diseases. In addition, the flag set on the diagnosis will not be included in any of the statistical reports sent to the MoH. Only the diagnosis itself with the IDC-10 code will be included, and the data will be processed further in the servers of the MoH. 4.1. Initial effects of the data summarization tool use The proposed summarization was released for a limited use from September 1st 2015. It was first introduced for general practitioners. Table 4 shows the overall statistics concerning the initial effects of the summarization items usage for five most commonly diagnosed chronic diseases. Between September 1st 2015 and the end of the same year, a total of 6982 summarizations were defined

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

197

Fig. 5. Components of summarization tool in relation with base MIS system (A – Client Layer, B – Business Logic and Data Access Layer, C – Data Storage Level).

Table 4 Statistics on the on-demand usage of the summarization items. Category

Count

Patients affected Patients that had no HoD before Total number of summarization headers Total links to existing visits Total placeholder visits Prescriptions included Active users

3576 3214 6982 38,905 58,356 106,703 89

affecting 3576 patients. A significant majority of them (3214 or 89.9%) were patients that had never had an HoD created for a chronic diagnosis. This number is close to one third of the number of patients that had proper chronic HoD already defined. In addition, defined summarization headers were connected to 38,905 existing visits and 58,356 placeholder visits that were related to 106,703 prescriptions. This way, the quality of the existing chronic HoDs was improved and the summarization ensured that there were no missing items for newly identified patients. The total number of active users for this new functionality was 89. They represent 18.7% of all the users that were active in 2015. A closer look at the structure of the mentioned users reveals that 81 out of 89 were from the general practice department. Since the number of active users in the general practice department was 155 in 2015, the acceptance rate within this department was 57.4%. Taking into account the fact that the functionality is still on trial and that a general training was not conducted for all potential users, the initial response was satisfactory. 4.2. Effects of the accepted automatic summarization On the other hand, the automatic summarization can identify a much larger number of patients with chronic diagnoses. Since the

automatic summarization can process large amounts of data which can affect more than one half of registered patients, this feature must be configured carefully. From the medical point of view, all existing HoDs can be included in the summarization, but for other scenarios, the medical professional should confirm the results. Fig. 6 shows the flow of the reasoning process for the initial summarization tool configuration. The initial setting for the automatic summarization was to identify the existing HoDs, which were marked as regular HoDs, for a chronic diagnosis from the predefined set (scenario 2). The additional condition was that the identified HoDs contain at least three visits. The total number of registered visits (initial + chained) in this scenario was more than 6 times higher than in scenario 1. The number of patients affected by this scenario was 44,582 (Table 5), or 22.2% of the whole population. Added to the patients identified in scenario 1, this percentage amounted to 27.8%, which was close to the expected level [9,10]. For the rest of the scenarios, the automatic summarization is still questionable. It is not possible to give precise information on how many new chronic patients could be identified. There is still a significant number of medical professionals that register each visit separately, creating the total of 716,442 visits, which is almost equal to the number of all visits present in scenarios 1 and 2 (the sum of all initial and chronic diagnoses for scenarios 1 and 2 is 737,587). The total number of potentially new patients (32,531) shows how important it is to include this set of data into summarization. This number represents one eight of the total population. In order to attain a higher level of certainty so that the patient having a HoD registered for a chronic disease is someone who needs more medical attention, the patients identified here must be more carefully examined. More than half of them (18,526 out of 32,531, or 56.94%) have only one or two registered visits related to some of the mentioned diagnoses. They cannot automatically be categorized as patients with a chronic disease. From the category of people having exactly

198

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

Table 5 Potential effects of the proposed automatic data summarization method in all five usage scenarios and five most common chronic diagnoses – I10 Hypertensio arterialis essentialis (primaria), E11 Diabetes mellitus ad insulino independens, E10 Diabetes mellitus ab insulino dependens, I20 Angina pectoris, I49 Arrhytmiae cordis aliae. Category

Sum

I10

E11

E10

I20

I49

Initial chronic visits – scenario 1 Chained chronic visits – scenario 1 Patients affected – scenario 1

29,196 72,984 11,784

20,892 51,230 8517

3727 10,178 1408

1770 5517 611

1737 4366 792

1070 1693 456

Initial visits – scenario 2 Chained visits – scenario 2 Potential new patients – scenario 2

69,963 565,444 44,582

52,134 406,840 34,489

7939 72,394 5414

3580 33,810 1455

3542 30,911 1902

2768 21,489 1322

Single visits – scenario 3 Potential new patients Patients having > 5 single visits Patients having 3–5 single visits Patients having 1–2 single visits 2 visits – 7 days difference 2 visits – 14 days difference 2 visits – 90 days difference

716,442 32,531 7195 6810 18,526 557 781 2795

472,606 22,500 5442 4161 12,897 415 584 2009

103,440 3276 805 647 1824 60 78 303

72,500 1944 368 539 1037 15 17 91

33,937 2553 359 878 1316 21 40 201

33,959 2258 221 585 1452 46 62 191

Request for further examination – scenario 4 Total number of patients affected Number of patients having > 5 requests Number of patients having 3–5 requests Number of patients having < 3 requests Potential new patients Number of patients having > 5 requests Number of patients having 3–5 requests Number of patients having < 3 requests

55,383 25,163 1612 4855 18,696 3279 133 453 2693

29,427 14,336 727 2534 11,075 802 18 78 706

9329 3819 303 854 2662 289 5 32 252

3070 1025 133 264 628 107 3 13 91

6421 2871 202 551 2118 957 47 151 759

7136 3112 247 652 2213 1124 60 179 885

Total medication prescriptions Medication prescriptions – scenario 5 Total number of patients affected Patients having > 5 prescriptions Patients having 3–5 prescriptions Patients having < 3 prescriptions Potential new patients Patients having > 5 prescriptions Patients having 3–5 prescriptions Patients having < 3 prescriptions

3,263,325 1,567,717 86,327 56,655 12,594 17,078 14,406 5585 2851 5970

2,464,136 1,098,089 54,367 36,578 7919 9870 5060 1989 923 2148

335,726 184,572 9899 7294 1130 1475 896 336 143 417

130,674 45,438 2330 1760 276 294 233 90 45 98

222,888 163,090 12,317 7131 1950 3236 5258 2022 1175 2061

109,901 76,528 7414 3892 1319 2203 2959 1148 565 1246

Table 6 Cumulative estimation of the number of patients that can be identified using the summarization tool in all five identified scenarios and five most common chronic diseases. The values in bold represent minimum and bold italic maximum potential number of patients in the category. Category

Count

Cumulative sum

Percentage

Patients affected – scenario 1 Potential new patients – scenario 2 Potential new patients – scenario 3 Patients having > 5 single visits Patients having 3–5 single visits Patients having 1–2 single visits 2 visits – 7 days difference 2 visits – 14 days difference 2 visits – 30 days difference 2 visits – 60 days difference 2 visits – 90 days difference Potential new patients – scenario 4 Number of patients having > 5 requests Number of patients having 3–5 requests Number of patients having < 3 requests Potential new patients – scenario 5 Patients having > 5 prescriptions Patients having 3–5 prescriptions Patients having < 3 prescriptions

11,784 44,582 32,531 7195 6810 18,526 557 781 1271 2106 2795 3279 133 453 2693 14,406 5585 2851 5970

11,784 56,366 88,897 63,561 70,371 88,897 70,928 71,152 71,642 72,477 73,166 92,176 71,061 71,514 74,207 106,582 76,646 73,912 77,031

5.87 28.07 44.27 31.65 35.04 44.27 35.32 35.43 35.68 36.09 36.44 45.90 35.39 35.61 36.95 53.08 38.17 36.81 38.36

two visits, the number of potential candidates is 2795 (if 90 days between visits is used as a cut-off) – but only 781 of them have two examinations registered in the period of two weeks or shorter. Usually, when patients come for the first time with health problems indicating a chronic disease, the doctor will schedule a follow-up in a week or in 14 days. A further review shows that around 22% of these patients have more than five visits if they are diagnosed with a chronic disease and around 20% have three to five examinations.

73,299 73,752 76,445 82,030 84,881 90,851

36.50 36.73 38.07 40.85 42.27 45.24

The additional two opportunities for categorizing patients as chronic could be recurrent requests for further medical examinations and prescriptions. Requests for further examinations are an important category since everyday use reveals that when a GP defines a request for a specialist or lab analysis it very often happens that the patient visits a medical institution with incompatible software. In some cases, data is entered into the MIS later, but sometimes the doctor just takes the hard copy report from the patient and archives it. The total of 25,163 patients with unrelated

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

199

Fig. 6. Potential effects of additional data summarizations.

requests were identified, while 3279 were potentially newly identified patients with chronic diseases. A study of the medical prescriptions revealed a new category of patients – the patients having many prescriptions connected with chronic diagnoses, but with no HoD or single visit defined. The total of 5585 patients had more than five prescriptions, and the total number of patients from this category was 14,406. Generally, two kinds of patients can be identified here – those with acute problems and older patients whose diagnosis had been made many years before the MIS was introduced. They see all medical problems and medical examinations as an opportunity to ask for a medication related to their chronic disease. In most of the cases like this one, the GP will create only one visit – connected to the acute problem and then just include medical prescriptions related to the chronic diagnosis. 4.3. Cumulative estimation of the number of patients that can be identified using the summarization tool As explained in the previous sections, the definition of the HoD for certain diagnoses (including chronic ones, scenario 2) is used more often and at a satisfactory level, but the option of marking the HoD as chronic is used significantly less. Checking the number of patients affected by a direct usage of the data summarization tool together with these identified by the accepted automatic summarization, we can conclude that the number could approach the expected percentage of patients with chronic diseases. The total of 56,366 patients (scenarios 1 and 2) had a defined HoD for five most frequently diagnosed chronic diseases (28% of the population). This percentage was close to the expected level of 35–45% (for the Republic of Serbia), but there was still a number

of cases where HoD was not defined. Instead, potentially new patients with some chronic diagnoses could be identified by tracking individual visits as well as documents and prescriptions associated with a chronic disease, but defined within the scope of some other diagnosis. The proposed summarizing method should help form more accurate statistics on the number of chronic patients. Table 6 shows an estimate of potentially newly identified patients. It can be said with high accuracy that the patient whose HoDs belong to scenario 1 or scenario 2 has a chronic disease, whereas all others discovered in scenarios 3, 4 and 5 can only be referred to a physician for further tests. In cases like these, the medical practitioner is the only decision maker. The set of data uploaded to the MoH will contain literally all the patients associated with a chronic diagnosis at least once, and in many cases they can be identified as false positives. For example, the number of new potential patients having at least one visit associated with the mentioned diagnoses amounts to 88,897, or 44.3% of all patients. Next, judging by the prescribed medication, there are 14,406 patients that have been excluded from the previous statistic. If they are added to the count, the total number rises up to 103,303 or 51.4% of the population. If patients marked as using scenario 4 are to be added, the total number goes all the way to 106,582 patients or more than 53% of the population. This is definitely a number higher than expected. The analysis of the data collected through different scenarios induces the following conclusions:  As regards scenarios 1 and 2, the situation is very clear and all the patients having a defined HoD with a chronic diagnosis, despite the fact whether it is particularly marked as chronic

200

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

Fig. 7. Overview on number and ratio (in percentages) in overall population of potentially identified patients suffering chronic diseases.

or not, can be positively identified as patients with chronic medical conditions. The number of such patients sums up to 28%.  In scenarios 3, 4 and 5, automatic reasoning could be applied but since a significant number are deviations, medical professionals must be the ones who can definitely confirm the patient’s state. Including the mentioned scenarios, the percentage of identified chronic patients can go up to 50%. Judging only by the numbers obtained from the research of the examined case of the Nis Primary and Ambulatory Care center, the significant potential for detecting patients suffering from chronic diseases can be spotted (Fig. 7). It can be observed that the actual support for chronic diagnoses is currently used only for less than 6% of all patients. Unfortunately, a much higher percentage of patients is affected by some chronic medical problem. The percentage of identified patients was risen to 28% upon the inclusion of the data from scenario 2. A further inclusion of potential patients identified through the next three scenarios could rise this percentage up to 53%. But, there is a point at which a medical professional must decide which patients should be automatically treated as chronic and which must be verified by the doctor; in other words, the system can either automatically set the flag for some patients, or just raise an alarm for others.

5. Discussion A proper identification of patients suffering from chronic diseases along with improving medical professionals’ working environment is important in many respects: social, medical and research ones. Primarily, if a chronic disease is identified in due time, medication can start and the process of the patient’s condition effective monitoring can be introduced [26,27]. Once the chronic diagnosis is made, a proper care and disease control can be established following some of the available models [28]. Also, these patients have to be educated in a certain way in order to care about the quality of their living environment [29]. This is even more important for the patients suffering from more than one chronic disease [30]. If one chronic disease is already identified

in a patient, the summarization method will help in identifying if there is enough data registered to be able to associate the patient with some other chronic disease. In terms of legal requirements, the Serbian MoH officially does not require any special interface for tracking chronic diseases within general practice. Nevertheless, within the MIS, one can easily mark a diagnosis as chronic, and get all the benefits of the enhanced data analysis. In order to help with the unification process, the summarization method has been proposed, followed by the summarization tool. Due to the lack of formalized directions, several different practical approaches were identified. All of the data stored in the patients’ ‘‘visits” is eventually uploaded to the service provided by the MoH where they are further processed and categorized. Unfortunately, no relevant feedback is provided to general practitioners, except for the overall statistical reports. Since overall statistical reports do not provide patient-level data, they cannot be used either for the automatic identification of new chronic patients or for the data aggregation for those already identified. This is another reason why the summarization tool is needed. The good side of the proposed method is that eventually the physician is the one that has to confirm whether a patient can be diagnosed with a chronic disease. The suggestion tool will just create an alarm that should help the doctor to decide about the patient’s status. Furthermore, once established, the summarization for the patient and a certain diagnosis can easily be associated with all newly created entities (visits, documents and prescriptions). When the automatic discovery mode is enabled, the summarization tool will raise alarms when some entity connected to a chronic disease gets discovered. The system can raise an alarm if some visit, document, or prescription associated with a chronic diagnosis is discovered, but then the number of occurrences and the frequency of discovered items must be taken in account. Table 6 shows an estimated number of newly discovered patients in all of the defined scenarios and with additional restrictions in the number of occurrences. Adding these mentioned restrictions is necessary in order to avoid false positives and not to distract the user with too many alerts. In other words, if an alarm is raised for each of the patients having at least one visit related to a chronic disease, more than 32,000 people will be involved. Contrary to this, setting

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

the filter to more than 3 visits will reduce this number to fewer than 14,000. For all newly identified chronic patients since the MIS’s went live, all of the data have been securely stored in the system and they can be retrieved. The problem that could be solved using the proposed summarizing method is the proper categorization and connection of the stored data. The causes of the problem were unforeseen different routines of the MIS usage. Without the proposed summarization, users will not be able to get easy-to-read reports. A default EHR overview shows only active HoDs, as well as completed ones that were marked as ‘‘important” or ‘‘chronic”. In order to see all of the existing medications, a user has to search through the EHR, but the users regard this action as complicated and time consuming. Even though the proposed method tends to be general, it has several limitations at present. It is implemented only for one specific information system, developed by one specific technology, and before its usage can be extended, some additional work is needed. Next, the summarization tool relies only on the data loaded from the database while other types of data sources are not supported. From the prospective of the summarization method, defining a complex relation between diagnoses and patients currently looks like too complex a task for end users without a foreseen usability. Users accepted the summarization mode including one patient and one diagnosis. The result of this summarization helps in tracking chronic diagnoses, but the relation to some medical entities gathered by other diagnoses could not be properly identified. 6. Conclusion In this paper, we presented a data summarizing method and supporting tool defined to improve the transparency of extracted data and allow the users of our system to have an effective data overview per patient and per diagnosis. The introduction of the proposed summarizing method will help medical professionals in retrieving, presenting and analyzing all necessary data for any patient, not only the ones suffering from chronic medical conditions. Furthermore, the proposed method could help in easier identification of new patients with chronic medical conditions. The introduced tool is intended to be easy to use for any category of potential users. The only user action required is to enter the diagnosis IDC-10 code in order to enable the summarization. When the summarization is active in the automatic mode, the user just needs to verify the retrieved data. The proposed summarization principles have been tested on a single type of a medical information system, but could be applied to other medical software applications, too. Depending on the existing data structures of the target information system, the required data extension could be more or less complex. The identified usage scenarios can be used as initial scenarios when analyzing the new target system, and this set can be later either expanded or reduced. Since a new data structure is introduced, using the data summarization in the MIS will increase an overall volume of the collected data. The data structure required for the data summarization contains only a small number of joined tables having only references to the originally stored entries, so the additional data will require a fraction of memory required for regularly stored visits and related entries. It should be mentioned that the previously collected data have not been impacted. The software builds the structure around the data already stored allowing summarization tools to interpret them in more details. It is envisaged that any future extension of the summarizing process should leave the existing data collection routines unchanged. Once the summarization is enabled, all newly created records will be automatically linked. The summarization is enabled by

201

default for all diagnoses marked as chronic or important, but can be enabled for literally any other diagnosis. The proposed method is not intended to change medical practitioners’ routine, but to provide them with some invisible support. The next goal of the proposed method is to help in an early identification of chronic patients. If a patient has a few visits during a certain period of time related to a chronic diagnosis, the system will raise an alarm if these visits occurs in short intervals, and it will suggest the patient be sent to further analyses. In the near future, more complex summarization methods will be introduced. At one point, it should include more diagnoses for a patient, and more patients for a diagnosis. This will help medical practitioners get a better overview of medical records for families and other closely related groups. Also, the summarization of different social groups of people can help in organizing follow-up examinations [17]. One of the next steps in supporting the chronic disease management within our MIS will be the introduction of a more direct access to the patients. Based on the research presented in [31] and on our own experience [32], we plan to introduce an SMSbased notification along with personal health record application connected to our MIS. In that way, the patients once diagnosed with chronic conditions could be more effectively notified for pending examinations and medical procedures. One problem that we were able to identify during the testing is the fact that a significant number of users tend to disable all additional tools treating them as just another task to do on top of the workload they already have. In these cases, an inactive summarization will produce no results. For this reason, the summarization can be enabled on a departmental level, or even a medical facility level and get automatically created and updated in the background. The user that intentionally disables this functionality will not have the summarization enabled, but the option will remain available to other doctors having access to the patient’s EHR. Each user can enable the summarization on the diagnosis or patient level and use it as a regular feature. The proposed diagnosis summary is developed with the intention of providing the medical practitioner with an instant access to the chronic diseases overview – including all visits, medication prescriptions and other connected medical documents related to medical examinations, diagnostics and therapeutic treatments. Even though the overview was developed for chronic diseases, it is applicable to any other diagnosis. Our next research goal regarding the data summarization is to help physicians properly mark newly identified patients and, for those that already have a chronic diagnosis, to aggregate all relevant data. The proposed summarization method and accompanying tools have not been designed to eliminate the role of the doctor, but to alert them when some data related to a chronic disease are not connected to a proper HoD. It will provide a complete overview and necessary statistics, but in the end, the medical professional is the one that should set the diagnosis status. 7. Conflict of interests Medical information system Medis.NET, which extension is presented in the work, is a result of a joint project of the Laboratory of Medical Informatics and Nis Ambulatory and Primary Care Center. As a commercial product, it is sold to other public primary care centers in Republic of Serbia. During the conduct of this study, Dejan Aleksic, Petar Rajkovic and Dusan Vuckovic received no financial compensation, but were allowed to use statistically processed data for other researches. Aleksandar Milenkovic and Dragan Jankovic received personal fees as full members of Laboratory of Medical Informatics.

202

D. Aleksic´ et al. / Journal of Biomedical Informatics 69 (2017) 188–202

References [1] M. Reed, J. Huang, R. Brand, I. Graetz, R. Neugebauer, B. Fireman, M. Jaffe, D.W. Ballard, J. Hsu, Implementation of an outpatient electronic health record and emergency department visits, hospitalizations, and office visits among patients with diabetes, JAMA 310 (2013) 1060–1065. [2] M.E. Tinetti, T.R. Fried, C.M. Boyd, Designing health care for the most common chronic condition-multimorbidity, JAMA 307 (2012) 2493–2494. [3] A. White, M. Danis, Enhancing patient-centered communication and collaboration by using the electronic health record in the examination room, JAMA 309 (2013) 2327–2328. [4] C.-J. Hsiao, E. Hing, T.C. Socey, B. Cai, Electronic health record systems and intent to apply for meaningful use incentives among office-based physician practices: United States, 2001–2011, System 18 (2011) 13–17. [5] A.H. Krist, J.W. Beasley, J.C. Crosson, D.C. Kibbe, M.S. Klinkman, C.U. Lehmann, C.H. Fox, J.M. Mitchell, J.W. Mold, W.D. Pace, et al., Electronic health record functionality needed to better support primary care, J. Am. Med. Informatics Assoc. 21 (2014) 764–771. [6] I. Yoo, P. Alafaireet, M. Marinov, K. Pena-Hernandez, R. Gopidi, J.-F. Chang, L. Hua, Data mining in healthcare and biomedicine: a survey of the literature, J. Med. Syst. 36 (2012) 2431–2448. [7] R. Fang, S. Pouyanfar, Y. Yang, S.-C. Chen, S.S. Iyengar, Computational health informatics in the big data age: a survey, ACM Comput. Surv. 49 (2016) 12. [8] J.S. Ancker, H.O. Witteman, B. Hafeez, T. Provencher, M. de Graaf, E. Wei, You Get Reminded You’re a Sick Person: personal data tracking and patients with multiple chronic conditions, J. Med. Internet Res. 17 (2015). [9] W.H.O., others, A global brief on hypertension: silent killer, global public health crisis, World Health Organization (2016). [10] D. Vukovic´, V. Bjegovic´, G. Vukovic´, Prevalence of chronic diseases according to socioeconomic status measured by wealth index: health survey in Serbia, Croat. Med. J. 49 (2008) 832. [11] P. Rajkovic´, D. Jankovic´, V. Tošic´, A software solution for ambulatory healthcare facilities in the Republic of Serbia, in: E-Health Networking, Appl. Serv. 2009. Heal. 2009. 11th Int. Conf., 2009, pp. 161–168. [12] J.M. Holroyd-Leduc, D. Lorenzetti, S.E. Straus, L. Sykes, H. Quan, The impact of the electronic medical record on structure, process, and outcomes within primary care: a systematic review of the evidence, J. Am. Med. Infor. Assoc. 18 (2011) 732–737. [13] P. Rajkovic, D. Jankovic, A. Milenkovic, Developing and deploying medical information systems for Serbian public healthcare: challenges, lessons learned and guidelines, Comput. Sci. Inf. Syst. 10 (2013) 1429–1454. [14] J.L.S. Sauver, D.O. Warner, B.P. Yawn, D.J. Jacobson, M.E. McGree, J.J. Pankratz, L. J. Melton, V.L. Roger, J.O. Ebbert, W.A. Rocca, Why patients visit their doctors: assessing the most prevalent conditions in a defined American population, Mayo Clin. Proc. (2013) 56–67. [15] C.J. Green, P. Fortin, M. Maclure, A. Macgregor, S. Robinson, Information system support as a critical success factor for chronic disease management: necessary but not sufficient, Int. J. Med. Inform. 75 (2006) 818–828. [16] P.A. Estabrooks, M. Boyle, K.M. Emmons, R.E. Glasgow, B.W. Hesse, R.M. Kaplan, A.H. Krist, R.P. Moser, M.V. Taylor, Harmonized patient-reported data elements in the electronic health record: supporting meaningful use by primary care action on health behaviors and key psychosocial factors, J. Am. Med. Inform. Assoc. 19 (2012) 575–582. [17] B. Jawhari, L. Keenan, D. Zakus, D. Ludwick, A. Isaac, A. Saleh, R. Hayward, Barriers and facilitators to Electronic Medical Record (EMR) use in an urban slum, Int. J. Med. Inform. 94 (2016) 246–254.

[18] D. Reichert, D. Kaufman, B. Bloxham, H. Chase, N. Elhadad, others, Cognitive analysis of the summarization of longitudinal patient records, in: AMIA Annu. Sympos. Proc., 2010, pp. 667–671. [19] J.S. Hirsch, J.S. Tanenbaum, S.L. Gorman, C. Liu, E. Schmitz, D. Hashorva, A. Ervits, D. Vawdrey, M. Sturm, N. Elhadad, HARVEST, a longitudinal patient record summarizer, J. Am. Med. Informatics Assoc. (2014), http://dx.doi.org/ 10.1136/amiajnl-2014-002945. [20] Y. Senathirajah, S. Bakken, D. Kaufman, The clinician in the Driver’s Seat: Part 1 – a drag/drop user-composable electronic health record platform, J. Biomed. Inform. 52 (2014) 165–176. [21] J.C. Feblowitz, A. Wright, H. Singh, L. Samal, D.F. Sittig, Summarization of clinical information: a conceptual model, J. Biomed. Inform. 44 (2011) 688– 699. [22] R. Pivovarov, N. Elhadad, Automated methods for the summarization of electronic health records, J. Am. Med. Inform. Assoc. 22 (2015) 938–947. [23] A. Laxmisan, A.B. McCoy, A. Wright, D.F. Sittig, Clinical summarization capabilities of commercially-available and internally-developed electronic health records, Appl. Clin. Inf. 3 (2012) 80–93. [24] T. Williamson, M.E. Green, R. Birtwhistle, S. Khan, S. Garies, S.T. Wong, N. Natarajan, D. Manca, N. Drummond, Validating the 8 CPCSSN case definitions for chronic disease surveillance in a primary care database of electronic health records, Ann. Fam. Med. 12 (2014) 367–372. [25] M. Kvist, M. Skeppstedt, S. Velupillai, H. Dalianis, Modeling human comprehension of Swedish medical records for intelligent access and summarization systems-Future vision, a physician’s perspective, 2011 . [26] P.A. James, S. Oparil, B.L. Carter, W.C. Cushman, C. Dennison-Himmelfarb, J. Handler, D.T. Lackland, M.L. LeFevre, T.D. MacKenzie, O. Ogedegbe, et al., Evidence-based guideline for the management of high blood pressure in adults: report from the panel members appointed to the Eighth Joint National Committee (JNC 8), JAMA 311 (2014) (2014) 507–520. [27] A.S. Go, M.A. Bauman, S.M.C. King, G.C. Fonarow, W. Lawrence, K.A. Williams, E. Sanchez, An effective approach to high blood pressure control: a science advisory from the American Heart Association, the American College of Cardiology, and the Centers for Disease Control and Prevention, J. Am. Coll. Cardiol. 63 (2014) 1230–1238. [28] N.M. Clark, M. Gong, N. Kaciroti, A model of self-regulation for control of chronic disease, Heal. Educ. Behav. 41 (2014) 499–508. [29] M. Heijmans, G. Waverijn, J. Rademakers, R. van der Vaart, M. Rijken, Functional, communicative and critical health literacy of chronic disease patients and their importance for self-management, Patient Educ. Couns. 98 (2015) 41–48. [30] M.E. Tinetti, A.D. Naik, J.A. Dodson, Moving from disease-centered to patient goals-directed care for patients with multiple chronic conditions: patient value-based care, JAMA Cardiol. 1 (2016) 9–10. [31] M. Siddiqui, M.Y. ul Islam, B.A.I. Mufti, N. Khan, M.S. Farooq, M.G. Muhammad, M. Osama, D. Kherani, A.N. Kazi, A.M. Kazi, Assessing acceptability of hypertensive/diabetic patients towards mobile health based behavioral interventions in Pakistan: a pilot study, Int. J. Med. Inform. 84 (2015) 950–955. [32] P. Rajkovic´, D. Jankovic´, A. Milenkovic´, A software model of mobile notification system for medication misuse prevention, in: E-Health Networking, Appl. Serv. (Healthcom), 2013 IEEE 15th Int. Conf., 2013, pp. 569–574.