A Medical Decision System to Prescribe Antibiotics
Adverse drug reactions lead to deaths in the healthcare system, especially the ones linked to antibiotic prescriptions. In order to reduce and prevent these, researchers at BFH have programmed a semantic decision-making system that supports doctors in their decisions alongside the usual guidelines. Our author reports here on the tool and compares it with the pure use of guidelines.
Adverse drug events (ADEs) and reactions (ADRs) are some of the most important causes of mortality in the healthcare context. In Europe, almost one in ten hospitalizations results in an adverse drug reaction (Hadi et al, 2017). Moreover, adverse drug events claim each year between 700,000 and 1.5 million casualties in the United States (Laura, 2009) alone. From these, antibiotics are the second most common cause of drug related adverse events (Gandhi, 2003), and one of the most common classes of drugs associated with medical malpractice claims (Rothschild, 2002).
Up until now, physicians have relied on guidelines, especially in the context of hospital prescriptions. Unfortunately such guidelines do not offer sufficient support to solve the problem of adverse events. Guidelines have three main disadvantages:
- The first is Maintainability, given the fact that the guidelines are static and only contain very explicit rules. As an example, when a germ becomes resistant to an antibiotic (meaning that the antibiotic is not effective anymore on that infection) a big part of the guidelines needs to be rewritten.
- The second is Straightforwardness. In complex cases, the physician has to reason about, cross-check and manually combine several different sections of the textual guidelines.
- Finally, the third is Specificity, since in order to be effective, physicians need recommendations that consider specific additional clinical characteristics of individual patients.
Therefore, we proposed a general architecture for recommendation systems adapted for this kind of context and we develop a specific system for antibiotic prescription, which we call PARS (Ben Souissi et al., 2019). The type of context that our architecture covers is characterised by highly risky decisions or decisions with high stakes. Therefore, it is essential to provide the decision maker with concrete explanations of the proposed decisions. Particularly for the medical domain, a decision that is not explained and documented cannot be trusted by the decision-maker.
Our system PARS implements, integrates and automates existing knowledge and rules of good practice. The proposed solution is intended to be used by a decision maker who must adapt his/her decision both to each subject’s specific needs and characteristics, as well as to different types of evolution.
Our approach is based on the combination of semantic technologies with MCDA (Multi-Criteria Decision Aids).
The core system of PARS is comprised of three Knowledge Sources and two Reasoning Engines (seen in Figure 1). Our three knowledge sources describe the issues (infections), the alternatives (antibiotics) and the subject (patient). These are structured into Ontologies: which are knowledge schemes of semantic web technologies that Gruber defined as « explicit specifications of a conceptualization » (Gruber, 1993; Gruber, 1995).
Figure 1 : PARS Architecture (Ben Souissi et al., 2019).
Regarding our Reasoning Engines: The first one concerns the selection of potential alternatives according to a presented issue: In our case this means selecting the potential antibiotics according to a presented infection, through ontological reasoning and matching of the infection and antibiotic ontologies. The second reasoning engine further refines these results working with the antibiotic and patient ontologies, to assess the adequacy of potential alternatives (potential antibiotics) to the characteristics of the subject (patient) (seen in Figure 2). The set of the potential antibiotics will then be sorted into three categories Recommended, Possible and To Be Avoided. For this we structure the recommendation rules into a suitable sorting method: the MR-Sort with Veto.
MR–Sort with Veto (Leroy et al., 2011) is an ELECTRE TRI simplification (Figueira et al., 2005), which in turn is part of the ELECTRE family of models (Figueira et al., 2005). ELECTRE models are Multiple Criteria Decision Aiding (MCDA for short) methods. MCDA encompasses several methods and algorithms that are designed to provide useful recommendations in a diversity of domains (Bouyssou et al., 2006). It requires both the integration of quantitative data and qualitative considerations (Park et al.,2012).
The main motivations behind our choice are: (1) MR–Sort with Veto is a NCS (“non-compensatory sorting”) method. This means that good assessments do not compensate for bad evaluations, which fits our domain better than the alternatives. (2) The Notion of Veto: Where certain criteria performances are explicit eliminators in the prescription domain. This Veto can correctly model the assignment of an alternative (antibiotic) in a worst category for such cases. Finally, (3) The method operates under simple and synthetic rules with small set of parameters, a fact that facilitates maintenance and evolution of the overall system.
Figure 2: Using MR Sort method with Veto in the Semantic Model to link and to assess antibiotics for a patient bytoxicity risk (Ben Souissi et al., 2017)
Through this structure, our solution can link and match the heterogeneous knowledge sources expressed by experts. In collaboration with a Hospital Center, we have applied this approach in the medical domain and more specifically in the prescription of antibiotics (as detailed in our already published results: (Ben Souissi et al., 2016; Ben Souissi et al., 2017; Ben Souissi, 2017; Ben Souissi et al., 2019).
We compared the recommendations proposed by the PARS system with those of the guidelines in use at a Hospital Center.
PARS gives better results since it considers each critical criterion of the patient independently, while the guidelines consider a group of critical criteria of the patient in an equivalent manner. PARS deals with additional cases or reasons for additional sensitivities (patient criteria and side effects) when compared with the guidelines, because it operates with more detailed, dynamic and structured knowledge (Ben Souissi et al., 2016; Ben Souissi, 2017; Ben Souissi et al., 2019). Finally, PARS gives recommendations for complicated scenarios, while the use of guidelines cannot be as straight-forward as our system in these cases. Indeed, when making decisions only with the guidelines the physician needs to cross-check in several places to have any amount of confidence about recommended, possible and to-be-avoided antibiotics, for each specific case. This fact raises real-world issues, during antibiotic prescription, which our system can alleviate.
- Ben Souissi, Souhir (2017). “Vers une nouvelle ggénération d’outils d’aide à la décision s’appliquant à la prévention des risques lors de la prescriptiondes antibiotiques: combinaison des technologies Web ssémantique et de l’aide multicritère à la décision”. PhD thesis. Université de Valenciennes et du Hainaut-Cambresis; Université de Mons.
- Ben Souissi, Souhir et al. (2016). “Categorizing the suitability of an alternativefor a subject”. In:IEEE Symposium Series on Computational Intelligence(SSCI). IEEE, pp. 1–8.
- Ben Souissi, Souhir et al. (2017). “Reducing the toxicity risk in antibiotic pre-scriptions by combining ontologies with a multiple criteria decision model”.In:AMIA Annual Symposium Proceedings. Vol. 2017. American Medical In-formatics Association, p. 1625.
- Ben Souissi, Souhir et al. (2019). “PARS, a system combining semantic tech-nologies with multiple criteria decision aiding for supporting antibiotic prescriptions”. In:Journal of biomedical informatics 99, p. 103304.
- Bouyssou, Denis et al. (2006).Evaluation and decision models with multiplecriteria: Stepping stones for the analyst. Vol. 86. Springer Science & BusinessMedia.
- Figueira, José, Vincent Mousseau, and Bernard Roy (2005). “ELECTRE methods”. In:Multiple criteria decision analysis: State of the art surveys. Springer,pp. 133–153.
- Gandhi, Tejal K et al. (2003). “Adverse drug events in ambulatory care”. In:New England Journal of Medicine348.16, pp. 1556–1564.
- Gruber, T (1993). “A Translational Approach to Portable Ontologies”. In:Knowledge Acquisition5.2, pp. 199–229.
- Gruber, Thomas R (1995). “Toward principles for the design of ontologies usedfor knowledge sharing?” In:International journal of human-computer studies43.5-6, pp. 907–928.
- Hadi, Muhammad Abdul, et al. (2017). “Pharmacovigilance: pharmacists’ perspective on spontaneous adverse drug reaction reporting.” Integrated pharmacy research & practice 6: 91.
- Landro, Laura (2009). “Incentives push more doctors to e-prescribe”. In:TheWall Street Journal January21, p. 2009.
- Leroy, Agnes, Vincent Mousseau, and Marc Pirlot (2011). “Learning the param-eters of a multiple criteria sorting method”. In:International Conference onAlgorithmic DecisionTheory. Springer, pp. 219–233.
- Park, Seon Gyu et al. (2012). “Shared decision support system on dental restora-tion”. In:Expert Systems with Applications39.14, pp. 11775–11781.
- Rothschild, Jeffrey M et al. (2002). “Analysis of medication-related malpracticeclaims: causes, preventability, and costs”. In:Archives of Internal Medicine162.21, pp. 2414–2420.