Methodology

In the proposed DIAKID system, all drug classes, drug doses and patient characteristics are linked to the patient identity and all data is customized for each patient. While creating the DIAKID system, the following procedure was used:

Finding a dataset, an expert knwoledge and clinical guidelines: A publicly available data set was found to test the proposed system on patients. Then, an expert physician was found to guide the creation of drug lists as well as relations between drugs used in the treatment of T2DM and CKD. Finally, we followed the clinical guidelines [70] while creating the DIAKID ontology and comphrehensive semantic rules.

Organizing the dataset and populating DIAKID ontology: The dataset contains 41 characteristics (age, BMI, hbA1c, etc.) of 249 patients in Excel format [7]. However, the data on the active pharmaceutical ingredients of all drugs used in the treatment of T2DM and CKD were created by the authors was in semantic format (OWL) according to the information obtained from specialist physician and medical guidelines. Therefore, the dataset was organized into lists (patients, drugs, drug-drug interactions and K-raising drugs) in order to populate data in the proposed DIAKID ontology and data sets were obtained and transferred to Protégé. This was achieved by writing scripts with the help of Celfie, where the data in Excel was transferred to the object property, data property and individuals of the ontologies. As a result of the transfer, four ontologies were obtained: DIAKIDDI, DIAKIDPatient, DIAKIDRug and Modified DMTO. All ontologies then were combined under the umbrella of the DIAKID Ontology.

By creating comprehensive SWRL rules; A unique system was created that can recommend drug doses and provide DDI and K-raising drug warnings. In the created system, all drug classes, drug doses and patient characteristics are linked to the patient identity and all data is customized for each patient. Figure 8 is the summary of the DIAKID system that has been created.

Fig. 8

Methodology illustrating the dataset gathering stages and integration of data into the DIAKID ontology

The unique DIAKID system will guide specialist physicians and clinicians when writing prescriptions to patients. When specialists prescribe drugs to patients, the system will give necessary warnings about drug dosage, DDI and K-raising drugs. Ultimately, our aim is to improve the quality of life of patients and reduce the burden of specialist physicians with a busy work schedule by preventing errors caused by incorrect drug dosage, DDI and K-raising drugs.

Structure of the recommendation system

The KDIGO (Kidney Disease Improving Global Outcomes) 2020 Clinical Practice Guideline [70] and specialist physician information were used as criteria for the development and management of clinical practice for T2DM and CKD. CKD drug ontology and DDI ontology were systematically created with the help of a nephrologist. The drug ontology used in the treatment of T2DM was obtained by reusing the DMTO ontology and this ontology did not need to be restructured. We've expanded existing drug classes and added individuals. In addition, with the advice of a specialist doctor, all drug classes and doses were created in Protégé according to CKD stages. DIAKID includes four modules: dataset (raw data), ontology management system, rule editor, and query editor, as shown in Fig. 9. The dataset is the primary core part of DIAKID where core data can be saved as Ontology Web. The ontology management system allows editors to create and update the ontology model. Protégé 5.5, an open-source ontology software, was used to implement the development of DIAKID. The rule editor is often used to edit and run the Semantic Web Rule Language rules of the ontology model.

Fig. 9

The structure of recommendation system for drugs

DMTO and newly created drug-drug interaction, drug dosage determination, and patient ontologies were used to support the rule framework of the DIAKID ontology. Protégé was utilized to manage created ontologies, knowledge base and T2DM-CKD drug suggestions. An ontology is essentially built on three basic elements: classes, attributes, and relationships. Classes are used to represent domain information. Attributes are used to define classes, and relationships are used to describe the connection between classes. The flowchart of our system is as shown in Fig. 10: First, the classes and subclasses are created. Then, properties are created, followed by the creation of individuals based on the created ontologies. To infer the drug warnings, SWRL rules are created and executed. Finally, with SPARQL, inferred axioms are visualized.

Fig. 10

Flowchart of recommended system

Challenges

When we started this research, we encountered several difficulties during knowledge and data gathering. Two of the challenging issues were: (1) Finding physicians who specialized in both T2DM and CKD diseases. After contacting to several specialists in the field, we managed to find specialists and use their expertise to generate extensive SWRL rules for drug dose prediction, DDI warnings and K-raising drugs. In addition, for evaluating the predictions of the proposed system, we asked these specialist doctors to assess the accuracy of the system predictions. (2) The second challenge was finding an open dataset that contains patient data about T2DM and CKD-related test results. After several searches, we managed to find the open dataset [7] and re-purposed the dataset to this work as we explained above. To summarize, the proposed work requires expert knowledge and patient data for its predictions and it can be challenging to obtain them.

DIAKID: Proposed ontology

Ontologies, also known as the knowledge base, which form the basis of many smart applications and systems today; It is a collection of statements written in a language such as RDF and OWL that describe relationships between abstract ideas and set logical rules (SWRL) for reasoning about concepts. Linking or reusing ontologies has enormous benefits in medicine as it supports semantic interoperability between different datasets and applications, improves accuracy, and reduces engineering costs and efforts. At DMTO Ontology, we've expanded the list by adding classes and subclasses to drugs. We are removing it from ontology as the drug Troglitazone in DMTO has been withdrawn from the market [71].

Our ultimate goal in this study is to create a unique system using SW technologies that can provide appropriate warnings during the prescription of drugs for T2DM and CKD patients based on available patient data. This can minimize the risks of DDIs and inappropriate drug dosages that may occur in the treatment processes for T2DM and CKD. As a result, our aim is to increase the quality of life by suggesting a useful system for the patient's health and to suggest a system that can help clinicians in the treatment process. We develop patient ontology and drug ontology for the system that we propose and combine TPs and DDIs in the DIAKID ontology. And we also create the DIAKID Ontology by reusing the DMTO. Therefore, the proposed DIAKID ontology is new and aims to assist patients and physicians by solving the problems of K-raising drugs, DDIs, and drug doses for T2DM and CKD patients.

Reused DMTO ontologies

Class, Object Properties and Data Properties of DMTO: We reused the previously created ontology to develop our own ontology. The DMTO ontology was reused for diabetes diagnosis, treatment, and diabetic DDI. This ontology was more useful and better served our ultimate goal, the treatment and care of diabetic patients with CKD. In our work, we particularly preferred DMTO ontology since it contains diabetes drugs and their active ingredients. The classes, data properties and object properties, created from DMTO for the interactions of drugs used to create TPs of T2DM patients and drugs used in treatment, were integrated into DIAKID.

The main goal of our research is to reveal the interactions of drugs with each other in patients with more than one prescription and to help physicians/doctors in adjusting the drug dose according to the condition of the disease. Our proposed ontology-based system makes it easy for field experts to learn about DDI and also helps to warnings experts to protect patients from the side effects of DDIs and improve people's quality of life. We developed DIAKID by applying the newly created ontologies and reused ontologies together, which we will explain one by one below.

Patient ontology

The patient ontology consists of 41 characteristic features of 249 patients transferred from Excel format to Protégé with the help of Cellfie and stored [72]. Cellfie is a Protégé Desktop add-on for importing spreadsheet data from Excel format into OWL ontologies. By default, it comes with 5.0.0 version of Protégé and the new Cellfie version is automatically updated with Protégé. To transfer data to ontology with Cellfie, open the previously prepared ontology in Protégé, click "Create axiom from Excel workbook…" and select the excel worksheet from "Open File" window. The incoming Cellfie window consists of five main components: "Sheet tab, Worksheet view, Transform rule edit panel, Transformation browser, and Create Axiom". To create the conversion rules, the "Add" button is selected in the conversion rule edit panel, and then an editor dialog opens where the conversion expression can be written. To import new axioms, transformation rules are prepared and continued by selecting the "Create Axioms" button. As seen in Fig. 11, after clicking the “Create Axiom” button, Cellfie automatically generated 10,025 OWL axioms and showed a preview of all of them. These new axioms, created with the help of the Cellfie plugin, can be imported into a new OWL as well as an existing open OWL.

Fig. 11

Patient information transferred and stored in Protégé with the help of Cellfie

We use the Dataset [7] in excel format, which contains data on T2DM and CKD about 250 patients. In order to semantically annotate these patient data [73], we created the Patient ontology. The patients are members (instance) of the Patient class. In addition, the Patient class has several sub-classes such as Dialysis, Patient Gender, etc. To annotate information in the dataset, we also created necessary datatype and object properties (i.e. GFR, SCr, HbA1c, BMI, etc.) of CKD and T2DM patients in order to develop DIAKID for the treatment of these two diseases. Since CKD and T2DM diseases can cause many long and/or short-term complications in humans, the 'CKD Treatment' class has been added to our proposed ontology and used with the DMTO ontology that we have reused. The characteristic features of 249 patients were transferred to the ontology and the OntoGraf of DIAKIDPatient Ontology showing class, subclass, data properties, object properties, instances and relationships is presented in Fig. 12.

Fig. 12

The OntoGraf of DIAKIDPatient Ontology

Drug ontology

This chapter details the design of our proposed drug ontology and how it was developed. There are classes, subclasses, and individual characteristics of advanced drug ontology for both diseases (T2DM and CKD). Classes, subclasses and their individual characteristics are interrelated. For T2DM drugs, we re-used the DMTO ontology as we described in Sect. "Structure of the recommendation system". To our knowledge, there was no ontology for doses and interactions of CKD drugs critical to our research. Therefore, we created a new ontology of CKD drugs and drug doses as well as K-raising drugs using specialist physicians and KDIGO knowledge about CKD drug classes and doses [22, 74, 75]. And we also expanded its scope by adding K-raising drugs to our ontology. We expanded the scope of our study by transferring drug classes, drug active ingredients, drug starting doses and drug doses according to the GFR range to our ontology.

CKD drug classes consist of 7 main classes (Table 4). We used Cellfie to import the drug classes, subclasses and individual characteristics we received from the specialist physician to our ontology. Figure 13 shows the data import process to the Drug Ontology, the main classes, sub-classes and sample Drug individuals. In this process, a total of 823 axioms are created and transferred to our current ontology. The created Drug Ontology for CKD consists of 86 classes and 212 individuals.

Table 4 Information on drug doses according to eGFR in the treatment of CKD and T2DM [70]Fig. 13

Use of Cellfie in the import of drug classes and subclasses with individuals

Table 4 includes drug classes, general names of drugs, maximum and minimum drug doses, and the change of drug doses according to the GFR value range of CKD. Not all drug lists for drug classes are presented in this table, they are abbreviated. CKD directly affects drug absorption, drug distribution, metabolism, and drug excretion [70]. In this disease, the patient's life can be endangered as a result of misuse of drug doses and may cause side effects by rendering the treatment ineffective. Doses of renally excreted drugs should be adjusted according to the patient's functioning level of kidney function, and renal function is calculated as the estimate GFR ratio [22]. The drug ontology was established according to KDIGO for Clinical Practice Guidelines for the Treatment of Diabetes in CKD. The OntoGraf of the DIAKIDrugs Ontology, consisting of classes, subclasses and individuals, are presented in Fig. 14 after all drugs (classes and subclasses) used in CKD Treatment are added to the Ontology along with their doses (individual).

Fig. 14

The OntoGraf of DIAKIDrugs Ontology

Drug-drug interaction ontology

Again, DDI ontology, the knowledge of specialist physicians and KDIGO was used. Provided knowledge was used to create drug classes, subclasses and individual characteristics (instance), and the necessary information was transferred to the Protégé with the help of Cellfie. The DDI Ontology has been defined to warn of the interaction between drugs used for CKD disease and the outcome of this interaction. This work establishes the ontology of drug interactions, adopting the following steps: (1) Defines drug classes and drugs, (2) Analyses and states relationships between these drugs, and (3) creates rules of reasoning for drug interactions for these drugs. After all these stages, the basic concepts and relationships of the DDIOntology is created for DDI (Fig. 15).

Fig. 15

DDI Basic concepts and relationships

DDIs can cause side effects on the patient, and some drugs can increase the K level and put the patient in a difficult situation. Because the K level rises; Respiratory failure, respiratory irregularity, heart rhythm disorder and high amount of toxic substances in the heart can cause health problems that can seriously affect vital functions. As presented in Fig. 16, in order to draw attention to this situation, we added warnings about drugs that increase K levels, as well as DDI, to our system.

Fig. 16

Drugs that increase potassium level and protégé interface

DIAKID ontology

A detailed ontology was created to describe the concepts related to CKD disease/ treatment and the relationships between these concepts. The ontology layer defines concepts related to CKD stages and contains information about the doses of recommended drugs. To create classes of CKD Treatment, the KDIGO 2020 CPG and specialist physician knowledge were used. This ontological approach to staging the disease can more accurately predict the doses of drugs recommended for disease treatment based on data from the dataset for each patient. The patient's GFR category is required for the correct adjustment of drug doses in the management of CKD Patients. In this study, we overhaul the entire data set with SWRL rules and determine the GFR category of each patient. As a result of all re-used and newly created ontology, we form the novel DIAKID Ontology. Figure 17 shows stages of CKD and the doses of drugs recommended for the treatment of the disease for each patient using the proposed ontologies and SWRL rules. We will explain the DDI warning and drug dose recommendation process (Sect. "SWRL rules for drug-drug interaction and dose suggestion") using ontologies and SWRL rules in the coming sections. Figure 18(a) shows the generic classes for the DIAKID ontology and their subclasses. The generated classes have two kinds of properties, "data properties" to denote properties of particular class instances and "object properties" to describe relationships between different classes, as shown in Figs. 18(b, c), respectively. When we combined the newly created drug dose, drug-drug interaction and patient ontology with the DMTO ontology we modified, we obtained the DIAKID ontology, which is very dense with all classes, subclasses, data properties, object properties, individuals and relationships. The state of our ontology in question after all integrations is presented as OntoGrap in Fig. 19.

Fig. 17

The developed DIAKID ontology in the protégé-OWL 5.5. showing patients and their dataset

Fig. 18

The developed new concepts in DIAKID? ontology in the protégé-OWL 5.5

Fig. 19

The OntoGraf of DIAKID Ontology

Summary of ontologies used

To develop our DIAKID System (Fig. 20), we combined the dataset (raw data) consisting of 41 characteristic features of 249 diabetic chronic kidney patients with three new and one modified ontology, using specialist physician knowledge and clinical practice guidelines data to form the proposed DIAKID ontology. In this section, we have presented the summary information about our DIAKID system, which we have created by integrating the new ontologies and the modified ontology. To improve our system, we developed three new ontologies ourselves and modified and reused one existing ontology. Statistical information about newly created ontologies and modified and reused ontology (Ontology Name, Class Subclass, Object Property, Data Property, and Individual) are given in Table 5.

Fig. 20

Structure of DIAKID Ontology

Table 5 Statistics of reused and newly created ontologiesDataset and annotation with DIAKID

After doing research from many sources and web pages, we finally found a dataset that contains both T2DM and CKD patient data together and used it for our research. Our dataset, part of which is shown in Fig. 21, contains data from 249 diabetic patients with CKD at any stage, excluding ESRD (end-stage renal disease), who were followed for 4.2 years [7, 76]. The data we use to build our ontology will be available on request as we explain in Sect. "Methodology".

Fig. 21

249 diabetic patients with and CKD at any stage

We use the proposed DIAKID ontology to annotate the dataset. A total of 41 rules were written to transfer patient data to the DIAKID RDF axioms. A total of 10,025 axioms were transferred to Protégé with the help of Cellfie.

SWRL rules for drug-drug interaction and dose suggestion

While many drugs can have side effects on the patient, there are drugs that can cause side effects when used together. In this part of our study, we focused on DDIs that may cause adverse effects on the patient and adversely affect the patient's health if used together by diabetes patients with CKD. Diabetic patients with CKD cannot use ACE inhibitors and ARA drugs together [77]. Diltiazem and Verapamil from the Calcium Channel Blockers drug group cannot be used together with Beta blockers [78].

In our system, rules have been added to the system to provide the necessary warnings about drugs as shown in Fig. 22; for example, any drug in the ACE inhibitors drug group will have a negative effect on the patient when used together with any drug in the ARA drug group.

Fig. 22

ACE-i and ARA Drug Lists, which drug-drug interacting when used together

In our system, rules have been added to the system to provide the necessary warnings about all the drugs in Fig. 23, which explains that any drug in the Beta Blockers drug group will have a negative effect on the patient when used together with Diltiazem or Verapamil in the Calcium Channel Blockers drug group. Table 6 illustrates a DDI SWRL rule that we added to our system; when Benazepril in the ACE-i group is prescribed together with any drug in the ARA group, our system makes the necessary warnings to specialist physicians. We created SWRL rules to predict and warn of DDIs. The system will warn when any of the Beta Blockers drug group is prescribed together with Verapamil. Likewise, the system will warn again when any of the Beta Blockers drug group is prescribed together with Diltiazem. However, all drugs of the Beta Blocker group can interact with Verapamil and Diltiazem. As we explain in the SWRL rule in Table 7 below, the system makes the necessary warnings to specialist physicians when the drug Diltiazem, which is in the Calcium Channel Blockers drug group, is prescribed together with any of the drugs in the Beta Blockers drug group. A total of 34 SWRL rules have been created for DDI and adverse event warnings.

Fig. 23

Beta Blockers and Calcium Channel Blockers Drug Lists, which DDI when used together

Table 6 Examples of DDI rulesTable 7 Examples of DDI rules

SWRL is an ontology-based rules language and was used to create rules. Drug regulation guidelines are from KDIGO. As shown in Fig. 24, the rules can be explained simply as follows: If the GFR level is < 51 and > 9, which dose and how can the drug be used for CKD according to the GFR and albuminuria category? Side effects of the drug, if the patient has been on dialysis, will the dosages change and what will be the things to watch? The following SWRL rule may conclude that a patient with a GFR between < 51 and > 9 can use Labetalol at 400 mg twice daily. The Knowledge Management System was used as editors to create and update the ontology model. The Rule Editor was used to edit and run the SWRL rules of the ontology model. The graphical communication interface enables users to operate the system easily and get the desired result. We created 325 SWRL rules to improve the proposed DIAKID ontology, providing warnings for DDI and recommending drug doses. Rules were created to give warnings for DDIs, drug doses, CKD stages of the patient, and K levels as follows.

Fig. 24CKD Patient stage rules

According to the feedback we receive from specialist practitioners, GFR value range of the patient determines the drug doses used in the medication of T2DM and CKD. Therefore, 6 SWRL rules were created by looking at the GFR value of each patient to determine the CKD stage of each patient according to the GFR values given in Table 8. Two example rules are described in Table 8. In rule 1, if the patient's GFR is greater than 89, the patient has CKD in Stage G1. As stated in Rule 2 in the same table; If the patient's GFR test result is greater than 59 and less than 90, CKD is in Stage G2.

Table 8 The GFR stage of CKD patientsDDI rules

Diabetic patients with CKD cannot use ACE inhibitors and ARA drugs together, and they cannot also use Calcium channel blockers Diltiazem and Verapamil with Beta blockers [3, 13]. 34 Rules were created using SWRL. In Rule1 in Table 9, the system will give a "hasDrugDrugInteractionWarning" warning when Acebutolol is prescribed with any of the drugs Diltiazem and Verapamil. As seen in Rule 2 in the same table; When Verapamil is prescribed together with any drug from the Beta blocker drug group (Acebutolol, Atenolol, Betaxolol, …), the system will warn "hasDrugDrugInteractionWarning".

Table 9 DDI warnings written with SWRL rulesDrug dosage rules

As a result of the values, we obtained from the dataset [77], we wrote 276 rules by looking at GFR of each diabetic patient with CKD and recommended the drug dose accordingly. Two of these rules are shown in Table 10. In Rule 1, if the patient is a Chronic Kidney Patient, not a Dialysis patient and has a GFR between 9 and 51; The dose to be used for the drug Acarbose is 50 mg tid (three times a day). In rule 2, if the patient is not a Dialysis patient, but has a CKD and GFR is less than 10; Acebutolol dosage is 180 mg or 300 mg once a day.

Table 10 Warnings about drug doses written in SWRL rulesK-raising drug rules

4 Rules were created using SWRL for 34 drugs in 4 different drug classes. Below are a few rules we wrote about ACE-i, ARA, Beta Blocker drug groups and the ability of Spironolactone to increase K levels. In Rule2 in Table 11; When any of the drugs Benazepril, Captopril, Enalapril, Fosinopril, Lisinopril, Pentopril, Perindopril, Quinapril, Ramipril and Trandolapril are prescribed, the system will give a "hasPotassiumLevelIncreaseWarning(?a, true)" warning.

Table 11 SWRL rule warning for drugs that raise potassium levelsRules for T2DM drug dose

Here are some examples of rules we wrote for classes, data property and object property that we imported from DMTO to our own ontology. In Rule1 in Table 12, if the patient has diabetes and his GFR is greater than 29; For lispro_mix dose usage the system will result in "Reduce_Dose_by_25_percent".

Table 12 SWRL rules for T2DM drug dosesEvaluation and resultsEvaluation setup and metrics

The end goal of the proposed system is to accurately predict (a) drug dose, (b) DDI, and (c) K-raising drug warnings for patients having both diabetes and CKD. As we discussed before, we use the publicly available dataset [22] that includes information about patients having both T2DM and CKD. Using the available p