Reducing the Toxicity Risk in Antibiotic Prescriptions by Combining Ontologies with a Multiple Criteria Decision Model

Abstract

We consider the risk of adverse drug events caused by antibiotic prescriptions. Antibiotics are the second most common cause of drug related adverse events and one of the most common classes of drugs associated with medical malpractice claims. To cope with this serious issue, physicians rely on guidelines, especially in the context of hospital prescriptions. Unfortunately such guidelines do not offer sufficient support to solve the problem of adverse events. To cope with these issues our work proposes a clinical decision support system based on expert medical knowledge, which combines semantic technologies with multiple criteria decision models. Our model links and assesses the adequacy of each treatment through the toxicity risk of side effects, in order to provide and explain to physicians a sorted list of possible antibiotics. We illustrate our approach through carefully selected case studies in collaboration with the EpiCURA Hospital Center in Belgium.

Introduction

Adverse drug events are one of the most important causes of mortality in the healthcare context. They claim each year between 700,000 and 1.5 million casualties in the United States¹. From these, antibiotics are the second most common cause of drug related adverse events^2,3,4 and one of the most common classes of drugs associated with medical malpractice claims⁵. Every antibiotic treatment is associated with a risk rate, i.e., an assessment of the degree of risk, which takes into account both the severity and the frequency of undesirable side effects. Determining the best treatments is highly dependent on the considered acceptance risk. The acceptance risk of a particular treatment varies according to the vulnerability of the patient. Therefore, physicians are in need of support that can give them both an overview of the situation, but also a detailed description of the risk and the vulnerability of each patient.

For many hospitals, guidelines constitute the solution to guide antibiotic prescriptions, by linking infection diagnosis to the relevant antibiotic therapies. As an example, our collaborators in the EpiCURA Hospital Center have been using a hundred-page guideline since 2011⁶. Unfortunately, there is evidence showing that little or nothing has really changed regarding the malpractice issue in spite of the existence of guidelines. The main causes of this situation are that a) the physician cannot have a global view of the guidelines or complete explanations for the suggested treatment; b) in their current textual form these guidelines are completely static, making them hard to use and adapt to either the specific needs of the patient or the changing environment context (e.g. usage in emergency rooms).

Using decision support systems (DSS) in a clinical context is mainly about providing recommendations to physicians. Some of these Clinical DSS are using semantic technologies. From this perspective, Bright and coauthors⁷ have developed a formal ontology to structure the empiric antibiotic therapy guidelines of the New-York Presbyterian Hospital (NYP). The guidelines in this study have been explicitly entered in Protégé. This work was able to generate three kinds of prescribing alerts when the guidelines were not respected. Despite its advantages, this approach has serious limitations. Its main drawback is that both the basic data and the relationships between them should be explicitly entered in the system (i.e. the system cannot generalize to new data), making maintainability difficult, if not impossible. Other funded projects around semantic technologies for medical procedures such as REMINE⁸ and PSIP⁹ use data mining techniques to reduce drug adverse effects by taking into account the patient’s medical record.

Despite these efforts there is currently no widely accepted standardized framework that will help physicians in their day to day prescriptions needs, although some researchers have tried to move towards this direction^10,11. The broadest approach has been taken by Doulaverakis and coauthors^12,13 to cover drug to drug interactions and drug to diseases interaction but, apparently, with no consideration of patient vulnerability and the risk of the drug side effects. Other Clinical DSS use Multiple Criteria Decision Aiding¹⁴ (MCDA). For example, the AHP¹⁵ (Analytic Hierarchy Process) method has been used¹⁶ to analyze and compare the treatment options for eradication of H. pylori infection in children. The main limit of this work is the lack of structured objective knowledge. The recommendation is highly sensitive to the stakeholders evaluations and weights.

Clinical DSS have another important task which is to provide physicians with explanations for recommendations. In a previous work¹⁷, we combined semantic technologies with MCDA to sort alternatives (antibiotics) by their suitability to the subject (patient) but we did not consider the toxicity risk as is done by the physicians (assessing the risk level by combining severity and frequency).

Our objective in this work is to build a model for a clinical DSS that links knowledge structures for assessing the antibiotic prescription risk. The method we propose integrates MCDA with ontologies in order to provide to physicians sorted list of antibiotics assessed according to their toxicity risk for a given patient with an infection disease.

We use ontologically structured knowledge about the pharmacological characteristics of antibiotics and an ontology describing the critical criteria of patients. These ontologies are then linked through a set of rules structured in a Majority Rule Sorting model (MR Sort) with Veto¹⁸. This model, which is a simplified version of ELECTRE TRI²², assesses the alternatives by sorting them into ordered categories. This process results in antibiotic prescription recommendations categorized according to the risk of their side effect toxicity. We model the relations between concepts by simple interpretable rules, involving a small number parameters, to guarantee generality and maintainability of the knowledge model.

The main contributions of this paper are:

• A new model for integrating MCDA with ontologies for clinical Decision Support Systems

• An adapted version of the ELECTRE TRI – Majority Rule Sorting model (MR Sort) with Veto that is tailored to prescription recommendation with explanation.

• An experimental validation of the above model for categorization of antibiotics through risk assessment of side-effect toxicity

The rest of the paper is organized as follows: The next section (“The system”) describes our adaptation of the ELECTRE TRI - MR Sort rule, as well as our model for integrating the method with the systematic representation of patient profiles and antibiotic knowledge in ontological structures. Section “Illustration” details our validation of antibiotics categorization through the risk of side-effect toxicity. Finally in the “Conclusion” section, we present further research perspectives.

The system

The Semantic Model Our solution models a patient P that hosts pathogens which cause a bacterial infection. This patient goes to the hospital to seek medical care and requires a treatment by antibiotics. The first step of the decision support model for antibiotics prescription is to gather knowledge.

For this purpose we use an ontology O_P that models the patient and his/her relevant characteristics. These involve gender, age, comorbidities, allergies, and all the necessary patient information in order to assess the efficiency and the hazard of an antibiotic. All these characteristics influence the choice of an antibiotic in a way which is specified in reasoning rules. Indeed, a given antibiotic could suit a pregnant woman but not an old diabetic man, or conversely.

The variable SP_j → {0,1, 2} represents the sensitivity indication of patient to the side effect S_j. A value 0 means “no sensitivity”, a value 1 means “minor sensitivity”, while 2 represents “major sensitivity”.

Our second ontology O_A provides us with the set of side effects S_j of a given antibiotic A_i.

The variable RSA_ij is the risk of side effect S_j caused by antibiotic A_i. It represents the relation between an antibiotic A_i and a side effect S_j with indication of harmfulness. RSA_ij combines the intrinsic (i.e., independent of the antibiotic) evaluation of the side effect severity GS_j → {0,1, 3} with the side effect frequency indicator of the antibiotic FSA_ij → {0,1, 2}. The following combination rules have been established with the help of an infectious diseases specialist.

$$RS{A}_{ij}=\{\begin{array}{c}\begin{array}{cc}0& ifG{S}_j=0andFA{S}_{ij}=0\end{array}\\ \begin{array}{cc}1& \begin{array}{c}ifG{S}_j<3andFA{S}_{ij}=0\\ orG{S}_j=0andFA{S}_{ij}>0\end{array}\end{array}\\ \begin{array}{cc}2& \begin{array}{c}ifG{S}_j=1andFA{S}_{ij}>0\\ orG{S}_j=3andFA{S}_{ij}=0\end{array}\end{array}\\ \begin{array}{cc}3& ifG{S}_j>1andFA{S}_{ij}>0\end{array}\end{array}$$

The values of the severity GS_j are presented in Table 1. Value 0 indicates that side effect S_j is not severe. Value 1 indicates that the side effect Sj is severe. Similarly, value 3 corresponds to a harmful side effect. The frequency values FSAij are presented in Table 2. Value 0 represents a less than 1/10,000 chance that antibiotic A_i has side effect S_j. Similarly, values 1 and 2 represent the chance being between 1/10,000 and 1/100 and more than 1/100, respectively.

Table 1:

Severity

Description of the severity GSj of the side effect j	Level
Not severe	0
Severe	1
Harmful	3

Table 2:

Frequency

Description of the frequency FSAij of side effect Sj due to antibiotic Ai	Level
Does not exist and Very rare (< 1/10,000)	0
Rare and Infrequent (≥ 1/10,000 to < 1/100)	1
Frequent and Very frequent (≥ 1/100)	2

Currently we deal with a total of 45 side effects organized in 8 categories for 60 antibiotics and 22 patient criteria.

The MR Sort model with Veto for assigning objects to ordered categories We use the MR Sort (Majority Rule Sorting) model with Veto (^18,19,21) to classify the antibiotics for a given patient in ordered categories. MR Sort with Veto is a variant of ELECTRE TRI, which is a decision model belonging to the family of outranking methods^22,23. The goal of MR Sort (and ELECTRE TRI)²⁴ is to sort alternatives in ordered categories based on their performance on several criteria. In our context, the alternatives are antibiotics A_i and the criteria are the side effects S_j. The performance of an antibiotic w.r.t. a side effect for a given patient is determined by three elements: the severity GS_j, the frequency FSA_ij and the sensitivity of the patient to the side effect SP_j.

Each category C_k, k = 1,…, K, is associated a lower profile t(C_k), which is a vector of levels on each criterion representing the minimal requirements to belong to category C_k. Actually, not all these requirements are to be fulfilled, but only a sufficiently large majority of them, in order for an alternative to be assigned to this category. The majority condition can be implemented by assigning weights to the criteria and selecting a majority threshold; the condition is fulfilled whenever the sum of the weights of the criteria on which the alternative is at least as good as the profile passes the majority threshold. A simpler version of this rule consists of counting the criteria on which the alternative is at least as good as the profile; the condition is fulfilled whenever this number passes a minimal fixed number of criteria (This is tantamount to assigning equal weights to all criteria). An additional condition for being assigned into category C_k is that none of the alternative performances is “unacceptably bad”. Unacceptably bad performances are determined by veto thresholds on each criterion. Whenever the performance of an alternative on some criterion is worse than the corresponding profile performance by some specified quantity, a veto is activated which precludes assignment into the category.

To sum up, the principle implemented in MR-Sort with Veto in order to assign alternatives into categories is the following: An alternative is assigned to category C_k or better (i.e. C_k+1 up to C_K) if the performances of the alternative are at least as good as these of the profile t(C_k) on a majority of criteria and none of these performances falls below the veto threshold on each criterion.

In this work, we chose to use MR Sort with Veto for two reasons. On the one hand, it gives us structured rules to link the antibiotic knowledge structure to that of the patient. On the other hand, it provides us with an assessment of the suitability of the antibiotics to the patient. In addition, our method assigns the antibiotics (alternatives) in three categories (K = 3): R (“recommended”), P (“possible”) and TBA (“to be avoided”), according to their toxicity risk.

The MR Sort with Veto assignment principles are applied to the context of antibiotic prescription, yielding the following rules:

• An antibiotic is assigned to category R for a given patient if it has only a small number of side effects that the patient is sensitive to and if there is no unacceptable side effect (risk) for the same patient (no veto against R).

• A similar rule applies for an antibiotic being assigned to category P. The number of side effects tolerated is higher than those for category R and the list of unacceptable side effects (risk) can possibly be smaller.

• If none of these conditions are fulfilled, the antibiotic is assigned to category TBA.

Coupling ordered classification with the Semantic Model MR Sort with Veto is combined with the Semantic models to assess one by one each antibiotic A_i (as illustrated in Fig. 1).

Figure 1:

Using MR Sort method with Veto in the Semantic Model to link and to assess antibiotics for a patient by toxicity risk

These assignment principles are implemented using the following mathematical representation:

• A_i, i ∈ 1,…,n denotes the antibiotics that are potentially considered, i.e., antibiotics that cover the germs infecting the patient.

• S_j, j ∈ 1,…, m denotes the possible side effects of an antibiotic.

• CSA_ij, i ∈ 1,…, n, j ∈ 1,…, m is a variable taking its values from {0,1,2}. It represents the level of risk of side effect S_j of antibiotic A_i for the patient (taking into account the sensitivity of the patient to side effect S_j.

  $$CS{A}_{ij}=\{\begin{array}{l}\begin{array}{cc}0& \begin{array}{l}ifS{P}_j=0\hfill \\ orS{P}_j=1andRS{A}_{ij}=0\hfill \end{array}\end{array}\hfill \\ \begin{array}{cc}1& ifS{P}_j>0and1\le RS{A}_{ij}\le 2\end{array}\hfill \\ \begin{array}{cc}2& ifS{P}_j>0andRS{A}_{ij}=3\end{array}\hfill \end{array}$$

• CSA_i is a number associated to antibiotic A_i. It counts the number of risky side effects S_j such that CSA_i,j = 0.

• Two tolerance levels λ_R, λ_P with λ_R < λ_P determine the maximal number of side effects (risk) that are compatible with an assignment in categories R and P, respectively.

• An unacceptable risk (CSA_j > 1) can prevent assignment of A_i to category R or P. The list of unacceptable risks for an assignment to category R (resp. P) is a subset Veto[R] (resp. Veto[P]) of the set of all side effects risks. The vetoes are presented in Table 3.

Table 3:

Vetoes

Antibiotic (GSj, FSAij)	(0,0)	(0,1)	(1,0)	(0,2)	(1,1)	(1,2)	(3,0)	(3,1)	(3,2)
Patient SPj
0	R	R	R	R	R	R	R	R	R
1	R	P	P	P	P	P	P	TBA	TBA
2	P	P	P	P	TBA	TBA	P	TBA	TBA

For a given Patient, the assignment of a suitable antibiotic A_i to the class R, P or TBA is summarized in Table 4. CSA_j counts the number of side effects the patient is sensitive to. The first tolerance level, λ_R, is the maximum number of side effects the antibiotic could have to be in the R (recommended) category. The second tolerance level, λ_P, is the maximum number of side effects the antibiotic could have to be in the P (possible) category. In our application, λ_R was empirically set to 6 and λ_P to 12. Some side effects can induce a high risk on a patient, which explains the usefulness of vetoes (Table 3). The first veto, Veto[P], is put when the antibiotic has an unbearable side effect for the patient. This antibiotic could not be prescribed, even though it only has this side effect. For example, this veto would be raised if the antibiotic contains penicillin (RSA_ij = 3) and if the patient has a major allergy to penicillin (SP_j = 2). With a Veto[P], the considered antibiotic is put in the TBA category. Similarly, a second veto, Veto[R], is put when the antibiotic has a very rare severe impact on the patient health (RSA_ij = 1), which, however, is not unbearable. For instance, an antibiotic which contains penicillin would get this Veto[R] if the patient has a minor allergy to penicillin (SP_j = 1). With the Veto[R], the considered antibiotic A_i is put either in the P (possible) or in the TBA (to be avoided) category, depending on the value of CSA_i.

Table 4:

Rules for assigning antibiotics to classes R, P and TPA

CSAi < λR and no veto [R]	Recommended
CSAi < λP, no veto [P] and not Recommended	Possible
CSAi> λP or veto [P]	To be avoided

Illustration

In order to illustrate the model, we built a scenario through several meetings with practitioners of the EpiCURA Hospital Center⁶ (infectious diseases specialist, microbiologist) through which we were able to tune the sensitivities and the thresholds of our model. For the purpose of illustration, we present the following case:

Bill is a 68 years old man, he is suffering of alcoholism problems and he is diabetic. He comes to the emergency room with an increased temperature of 40.1C and he has an inflammation in his leg (see Fig. 2). It is later revealed, that he got injured 2 days before when he was gardening. The diagnostic yielded is “Erysipelas”.

Figure 2:

Erysipelas around the ankle²⁵

The guidelines (Fig. 3) inform us about the pathogens which cause the infection. In our example, the pathogens in question are Staphylococcus Aureus (MSSA) and Group A streptococci (GAS), as indicated in the second line of Fig. 3. To suggest an appropriate antibiotic, the guidelines distinguish three situations with respect to penicillin: (a) a patient who is not allergic to penicillin, (b) a patient with a minor allergy and (c) a patient with a major allergy. In the latter two cases, it suggests two different antibiotics.

Figure 3:

Guidelines recommendation for Erysipelas

Table 5 gives us the set of side effects (we did not present all the side effects that Bill is sensitive to, because for this use case the set of all antibiotics (Table 6) that can help Bill don’t have them; in terms of frequency, FSA_ij = 0), that Bill is sensitive to. SP_j denotes the sensitivity of Bill to the side effect S_j, j = a,…, h.

Table 5:

Sensitivities of Bill

SPa	SPb	SPc	SPd	SPe	SPf	SPg	SPh
1	2	1	1	1	1	1	1

Table 6:

Frequency, severity and risk of side effects per antibiotic

Antibiotics	Sa		Sb		Sc		Sd		Se		Sf		Sg		Sh
	GSa = 3		GSb = 3		GSc = 0		GSd = 1		GSe = 1		GSf = 3		GSg = 3		GSh = 3
	FSAia	RSAia	FSAib	RSAib	FSAic	RSAic	FSAid	RSAid	FSAie	RSAie	FSAif	RSAif	FSAig	RSAig	FSAih	RSAih
Oxacillin	0	2	0	2	1	1	0	1	1	2	0	2	0	2	0	2
Cefazolin	0	2	0	2	0	0	0	1	0	1	1	3	1	3	0	2
Ceftriaxone	0	2	0	2	0	0	0	1	0	1	1	3	1	3	0	2
Amoxicillin Clavulanic	0	2	0	2	0	0	0	1	1	2	0	2	0	2	0	2
Piperacillin Tazoboctam	0	2	0	2	1	1	0	1	1	2	0	2	0	2	0	2
Vancamycin	0	2	0	2	0	0	0	1	0	1	0	2	1	3	0	2
Moxifloxacin	0	2	1	3	0	0	2	2	0	1	2	3	0	2	1	3
Clindamicin	2	3	0	2	1	1	0	1	0	1	0	2	0	2	0	2

The list shown in Table 6 gives us: The set of antibiotics which are effective against or cover the germs causing Bill’s infection; the set of side effects S_j; the severity evaluation of the side effects GS_j; the frequency FAS_ij and the risk RSA_ij, where i = 1,.., 8 is the set of antibiotics and j = a,.., h is the set of side effects.

We sort this list by suitability to Bill. The output of our system for this case is the following:

• P: Oxacillin, CSA = 2, Veto[R]

• P: Amoxicillin-Clavulanic, CSA = 1, Veto[R]

• P: Piperacillin-Tazoboctam, CSA = 2, Veto[R]

• TBA: Moxifloxacin, CSA = 4, Veto[P]

• TBA: Cefazolin, CSA = 2, Veto[P]

• TBA: Ceftriaxone, CSA = 2, Veto[P]

• TBA: Vancomycin, CSA = 2, Veto[P]

• TBA: Clindamycin, CSA = 2, Veto[P]

Note that, currently, in order to get these results (without using our system), the physician has to manually cross-check and combine several different sections of the guidelines.

More precisely, for a patient that is as much vulnerable as Bill, at least one of the vetoes [R] and [P] are activated and therefore, no antibiotic is assigned to category “R”. For this kind of situation, indeed, we rarely find antibiotics sorted in the “Recommended” category. The best alternatives are usually sorted in the “Possible” category. The advantage of this kind of output is that it provides the prescribing physician with a global view of the sorted treatments. In addition it gives him/her a detailed explanation of the toxicity risk assessment for each particular patient.

To describe in further details our model, we present here a second scenario. For this scenario we consider the case of “Abscess Perirectal” from the guidelines (as seen in Fig. 4).

Figure 4:

Guidelines recommendation for Abscess Perirectal

The guidelines inform us about the pathogens which cause the infection in the second line of Fig. 4. For this example, the pathogens are Enterobacteria, Anaerobes, Enterococcus. To suggest an appropriate antibiotic, the guidelines distinguish three possible contexts with respect to penicillin: (a) a patient who is not allergic to penicillin, (b) a patient with a minor allergy and (c) a patient with a major allergy. In the last two cases, the guidelines suggest the same antibiotic.

The list presented in Table 7 gives the set of antibiotics which are effective against the germs causing the infection (Amoxicillin_Clavulanic, Piperacillin_Tazoboctam, Moxifloxacin). The set of side effects Sj, where j = 1,.., 22 is presented underneath. These side effects can be given by the antibiotics: Amoxicillin_Clavulanic, Piperacillin_Tazoboctam and Moxifloxacin.

Table 7:

Frequency, severity and risk of 22 side effects for 3 antibiotics

Antibiotics	S1		S2		S3		S4		S5		S6		S7		S8
	GS1 = 1		GS2 = 3		GS3 = 0		GS4 = 1		GS5 = 1		GS6 = 3		GS7 = 1		GS8 = 8
	FSAi1	RSAi1	FSAi2	RSAi2	FSAi3	RSAi3	FSAi4	RSAi4	FSAi5	RSAi5	FSAi6	RSAi6	FSAi7	RSAi7	FSAi8	RSAi8
Amoxicillin_Clavulanic	2	2	0	2	0	0	0	1	0	1	0	2	0	1	0	2
Piperacillin_Tazoboctam	2	2	0	2	1	1	0	1	2	2	0	2	2	2	0	2
Moxifloxacin	2	2	1	3	0	0	2	2	1	2	2	3	1	2	1	3
	S9		S10		S11		S12		S13		S14		S15		S16
	GS9 = 3		GS10 = 3		GS11 = 1		GS12 = 1		GS13 = 3		GS14 = 1		GS15 = 1		GS16 = 1
	FSAi9	RSAi9	FSAi10	RSAi10	FSAi11	RSAi11	FSAi12	RSAi12	FSAi13	RSAi13	FSAi14	RSAi14	FSAi15	RSAi15	FSAi16	RSAi16
Amoxicillin_Clavulanic	0	2	0	2	0	1	0	1	0	2	0	1	0	1	0	1
Piperacillin_Tazoboctam	0	2	0	2	0	1	0	1	0	2	0	1	0	1	0	1
Moxifloxacin	1	3	1	3	1	2	1	2	1	3	1	2	1	2	1	2

	S17		S18		S19		S20		S21		S22
	GS17 = 3		GS18 = 3		GS19 = 3		GS20 = 1		GS21 = 3		GS22 = 3
	FSAi17	RSAi17	FSAi18	RSAi18	FSAi19	RSAi19	FSAi20	RSAi20	FSAi21	RSAi21	FSAi22	RSAi22
Amoxicillin_Clavulanic	0	2	0	2	1	3	0	1	1	3	0	2
Piperacillin_Tazoboctam	0	2	0	2	1	3	1	2	1	3	0	2
Moxifloxacin	1	3	1	3	0	2	0	1	0	2	2	3

The severity evaluation of the side effects GS_j; the frequency FAS_ij and the risk RSA_ij (where i = 1,.., 3 is the set of antibiotics and j = 1,.., 22 is the set of side effects) are all presented in Table 7. The case description follows:

1	Diarrhea	2	Fulminant Hepatitis	3	Gastro Intestinal Troubles	4	Hepatic Troubles
5	Rash	6	Hematologic Troubles	7	Nausea and vomiting	8	Heart Failure
9	SNC Troubles	10	Neuropathy	11	Dizziness	12	Photosensitization
13	Ototoxicity	14	Benign Intracranial Hypertension	15	Cholestatic jaundice	16	Insomnia
17	Lyell syndrome	18	Steven Johnson Syndrome	19	Allergic Reaction	20	Cutaneous eruptions
		21	Anaphylactic shock	22	Prolongation of QT interval

Alex is a 29 years old patient, who is suffering from “Abscess Perirectal”. He is otherwise in good health without any history of medical incidents. His laboratory tests reveal that he does not have allergies and his creatinine level is 80ml/min.

For this case of infection, a surgical drainage is essential with antibiotic prescribing. Table 8 indicates the sensitivities SP_j of Alex to the side effects S_j, j = 1 22.

Table 8:

Sensitivities of Alex

SP1	SP2	SP3	SP4	SP5	SP6	SP7	SP8	SP9	SP10	SP11
0	0	1	0	0	0	1	0	0	0	0
SP12	SP13	SP14	SP15	SP16	SP17	SP18	SP19	SP20	SP21	SP22
0	0	0	0	0	0	0	0	0	0	0

We sort this list by suitability for Alex. The output of our system for this case is the following:

• P: Amoxicillin-Clavulanic, CSA = 1, Veto[R]

• P: Piperacillin-Tazoboctam, CSA = 2, Veto[R]

• P: Moxifloxacin, CSA = 1, Veto[R]

If Alex has a “Major allergy to penicillin”, the sensitivities SPi₉ and SP₂₁ change: SPi₉ = 2 and SP₂₁ = 2. The output of the system becomes:

• P: Moxifloxacin, CSA = 1, Veto[R]

• TBA: Amoxicillin-Clavulanic, CSA = 3, Veto[P]

• TBA: Piperacillin-Tazoboctam, CSA = 4, Veto[P]

The recommendations of our system match the ones in the guidelines (as seen in Fig. 4). Even in the case of “Major allergy to penicillin”, Moxifloxacin remains the only option.

Conclusion

We have developed a novel approach to help physicians with antibiotic prescriptions. Our solution combines semantic technologies with our own adaptation of the MR Sort method with Veto.

Our method sorts antibiotics in three categories: R (“recommended”), P (“possible”) and TBA (“to be avoided”) based on a small number of general rules. It is able to take into account a patient’s specific clinical criteria as well as generalize to new cases when – for example – a new antibiotic is added to the knowledge base. Using input from practitioners in the EpiCURA Hospital Center⁶ we have tuned the sensitivities and thresholds of our model and were able to validate our approach through examples that score prescription recommendations according to the risk of side effect toxicity.

In terms of future work we would like to apply our model to other types of drugs to determine if our method is applicable beyond antibiotics We plan to expand our model to other dimensions of the prescription problem including: costs, drug-drug interaction and drug-disease interaction among others.