Ontology Outreach Advisory

General Description
Clinical trials use entry criteria to select patients for the study. The choice of these criteria is an important step in clinical trial design. To be able to compare the results of the trial with those of other trials and to assess the generalizability of the results to daily clinical practice, the entry criteria have to be compatible with definitions used in comparable trials and the agreed standard definitions of disease. Description Logics, the logical foundations for ontology modeling are obvious candidates to model these entry criterias, to allow for declarative descriptions of classes of patients with particular symptoms. This is obviously complicated when no crisp disease definition exists. For this purpose we model clinical trials in an extension of Dls, so called Rough Description Logics, and use the semantics of these languages to study 9 different clinical trials about the sepsis condition.

The problem
Our current research was motivated by a recent study of the definitions for sepsis used in clinical trials. Before a medical treatment can be used in daily clinical practice, its effect and impact on the patient have to be investigated in a clinical trial. When several trials have been performed it is interesting to compare the results of those trials. Unfortunately, the nine different trials that were investigated in [Peelen et al. , 2005] showed too much variation in their definitions of severe sepsis patients to enable a fair comparison of trial results. Illustration 1 shows the mortality rate for these nine trials. Obviously, a proper comparison of the outcomes of different trials is almost impossible if the patient population is incomparable.
One of the reasons for the problems of defining valid trials in this domain is that there is no accepted general definition of the sepsis condition. Medicine is a typical domain where concepts cannot always be described in a crisp manner. E.g., the definition of a disease is not always clear-cut, especially if a single marker is lacking that distinguishes a patient with a disease from a patient without the disease. This is common in psychiatry and in diseases in which the underlying pathology of the disease is unclear. An example of the latter is sepsis.

Sepsis is a disease in which the immune system of the patient overreacts to an infection. Due to this reaction the patient becomes severely ill, which easily results in organ failure and eventually death. The cause and underlying cellular pathways of this disease are unclear, which hinders the precise characterization of the sepsis patient. Therefore, a consensus definition of sepsis was established in 1992 to define several stages of sepsis [Bone, R.C., 1992]. This definition does not provide a precise definition of sepsis, but gives the criteria for which there was a consensus that they should at least hold for a patient with severe sepsis. In this paper we focus on the patients with severe sepsis , but for brevity we will refer to these patients as septic . The consensus statement defines patients with severe sepsis as ‘patients having a confirmed infection with at least two out of four Systemic Inflammatory Response Syndrome (SIRS) criteria:
•temperature >38◦C OR temperature <36◦C
•respiratory rate >20 breaths/min OR PaCO2<32 mmHg
•heart rate >90 beats/minute
•leucocyte count <4,000 mm3 OR >12,000 mm3
and organ dysfunction, hypoperfusion, or hypotension. From now on we refer to these criteria as the Bone criteria. Patients who have this combination of symptoms may have sepsis, however, this is not necessarily the case. We refer to these patients as being possibly septic . On the other hand, we can define a group of patients that are septic for sure, namely those who fulfill the Bone criteria and have severe multiple organ failure. We will refer to these patients as the definitely septic patients and define them as fulfilling the strict criteria: the Bone criteria plus at least three of the following symptoms of organ failure:
•pH < 7.30
•thrombocyte count < 80,000 mm3
•urine output < 0.5 ml/kg body weight/hour (provided the patient not on chronic dialysis)
•PaO2/FiO2 <• 250, and
•systolic blood pressure <90 mmHg OR vaso-active medication.
In order to be able to compare clinical trials about sepsis, we need to formalise this information in an ontology, for which we extended the standard Description Logics by rough operators, i.e. the possibility to define approximations of concepts.

The solution
Rough Description Logics (Rough DL) provides us with the possibility to describe diseases for which a crisp definition is lacking by defining lower and upper approxiamtions. In the spirit of Rough Set theory, two concepts approximate an underspecified, vague, concept as particular sub- and super-concepts, describing which elements are definitely , respectively possibly , elements of the concept.
In order to use Rough DL for patient selection we first translated the definition for each trial into a DL formula. We did the same for the Bone definition and the Strict definition of sepsis, thus building a TBox with 11 definitions for septic patients. In addition we have translated a dataset from the Dutch National Intensive Care Evaluation (NICE) registry containing information on 71,929 patients into an ABox, using the terminology fromthe TBox. With the selection criteria for the different trials and the translated data, we used a DL-reasoner to select the patients that would be eligible for the different trials (thereby mimicking the patient selection process).

We now model the strict and Bone criteria mentioned above as lower and upper approximation of sepsis. We use Rough DL to formalise and compare sepsis definitions used in different trials. Describing sepsis through approximations enforces powerful semantic consequences. Rough DL turns out to be an appropriate logical representation language to model vague concepts and provide crisp answers to queries, and can thereby assist in the validation of existing and, ultimately, the construction of new trials.

Key Benefits of Using Semantic Web Technology
In our evaluation of medical trials about sepsis patients we have shown that modeling vague knowledge can help to answer important questions in the design of clinical trials. The validation of trials
•Based on their formal definitions is already an improvement over the usual data-based validation.
•When the validation done in a declarative way using Rough DL, the logical consequences of the semantics immediately reveals inconsistencies in the trial definitions, whereas several successive queries are necessary to do the same with standard DLs.
•Finally, we claim that Rough DL can be very useful when building new trials with vaguely defined medical conditions, as they enforce better models for the selection of patients.

References
[Bone, R.C., 1992] Bone, R.C. Definitions for sepsis and organ failure and guidelines for the use of innovative therapies in sepsis. Crit Care Med, 20(6):864 - 874, 1992.
[Peelen et al., 2005] L. Peelen, N.F. De Keizer, N. Peek, E. De Jonge, R.J Bosman, and G.J. Scheffer. Influence of entry criteria on mortality risk and number of eligible patients in recent studies on severe sepsis. Crit Care Med, 33(10), 2178 - 2183, 2005.

See Other Use Cases