Knowledge Acquisition Results
Generic Model of Pharmacokinetics We constructed a preliminary global model of pharmacokinetics from the results of the domain study using UML formalism (Fig. 2
). The model distinguishes three entities: the administration protocol, the real pharmacokinetic process description and the mathematical model building.
 | Figure 2. Preliminary global model of the pharmacokinetics domain according to UML formalism. |
From an administration protocol in which the route and dosage regimen are specified (dose, administration frequency, number of intakes): the real pharmacokinetics process can be described as such:
- The parent compound or its metabolites undergo a series of reactions including absorption, distribution, metabolism and elimination processes.
- It is studied by assays for the drug in samples (of, for example, blood or urine) collected at determined times.
A mathematical pharmacokinetics model can be elaborated using the experimental data and its parameters (for example half-life, clearance, distribution volume).
Lexico-semantic Analysis The pharmacokinetics text file contained about 300,000 words and was composed of texts of 5 to 1,043 words.
The Lexter lexico-semantic extraction based on the 1,950 pharmacokinetics texts provided a lexicon of 206,500 entries (17,520 different CTs) and 20,761 textual units. Of the 17,520 different CTs, 3,132 occurred more than 3 times and only 222 occurred more than 100 times. Nouns and nominal syntagma were the most frequent syntactic units.
We selected CTs representing direct instances of major pharmacokinetics concepts described in the generic model, such as CTs related to an administration protocol, a process, an observational datum and a parameter. There were 592 such specific pharmacokinetics CTs. We also selected CTs representing useful concepts such as quantification, variation, type of disease, or a population. These nonspecific terms represented 535 CTs. Table 1
shows an example of how the more frequent CTs were grouped according to their similar meanings in pharmacokinetics-specific or nonspecific concepts.
 | Table 1 Main Semantic Groupings for the Most Frequent Candidate Terms |
Study of the lexical environment of CTs with similar meanings yielded the set of common CTs that frequently co-occur. For example, fixation was most frequently described with CTs related to a binding site, a numerical value, a substance name, a precision, an ordinal value, and a study type. These CTs also co-occurred with “binding.” We could, therefore, deduce that the concepts underlying “binding” and “fixation” are common (Table 2
.
 | Table 2Results of the Analysis of the 302 Contexts of Occurrence Linked to the Candidate Term “Binding” and of the 127 Contexts of Occurrence Linked to the Candidate Term “Fixing”: Regrouping of the Candidate Terms Found Most (more ...) |
Model Building We compared the co-occurring concepts for CTs with near meanings to identify class attributes, class relationships, and new classes. For example, instances of a reaction co-occurred most frequently with other CTs that described where, when, why, and how the reaction occurred, on what the reaction acted, and how the reaction was assessed. To describe a reaction class, its characteristics need to be identified. The reaction class should be linked to classes describing (1) properties, (2) mechanisms, and (3) substances; a measurement class is needed to illustrate it.
The new classes, relationships, and attributes were then used to refine the generic model. The initial reaction class was then linked to a property class and a mechanism class. Parent compound and metabolite were subsumed by a substance class that was linked to an origin class and a state class. Process classes and reaction classes were subsumed by a more abstract class called “explanatory data.” Observational data was replaced by a measurement class, which had links with temporal-co-ordinate class, value class and evolution class. The administration protocol was divided into four classes (administered product, administration mode, dosing schedule, and treatment schedule). The administration protocol class was aggregated with two new classes—“measurement protocol” and “population features”—to form a new abstract class called “experimental protocol.”
Description of the Final Model of Pharmacokinetics The pharmacokinetics model describes the reality of the processes obtained under experimental conditions, the compartmental pharmacokinetics model and its parameters, and the variations obtained under particular conditions. The whole detailed model is described in Figure 3
.
 | Figure 3. Object-oriented model of pharmacokinetics information found in SPCs according to UML formalism. |
• Information about the real pharmacokinetics process knowledge
Information about the real pharmacokinetics process can be separated into that which explains the behavior of the drug in the body through time, and illustrative information, which includes the results of the measurements made on samples collected at given times.
Explanatory data describe the drug-body interaction. The changes of a drug in the body can be described as a succession of four processes: absorption, distribution, metabolism, and elimination. Each can be broken up into one or more reactions with one or more mechanisms. The reactions involve a substance, which changes states because of alterations in its state of binding, activity, localization, and biotransformation. The whole process—reactions, substances, and mechanisms can have certain properties (e.g., intestinal absorption is slow) and be ordered (e.g., urinary elimination occurs mainly by glomerular filtration).
The real pharmacokinetic process can be investigated by measurement with results expressed quantitatively or qualitatively and associated with a temporal co-ordinate. These measurements can change in time with one or more phases.
• Experimental protocol
The experimental protocol describes the conditions for obtaining data on the real process. It consists of an administration protocol, a measurement protocol and a population that has certain features.
The administration protocol is made up of information about the administered drug (e.g., tablet of drug X), the mode of administration (e.g., administration by the oral route on an empty stomach), the dosing schedule (e.g: repeated administration every 8 hours of 300 mg of X), and the treatment schedule (e.g., an 8-day course).
The measurement protocol describes the conditions of measurements (e.g., by gas chromatography).
The population features concern the population sample used for experimentation. These features consist of general features (e.g., man or dog) and particular features that give details about the physiologic type (e.g., a child less than 12 years old ), the genetic type (e.g., slow acetylator), or the pathologic type (e.g., renal insufficiency with a creatinine clearance lower than 30 ml/min).
• Mathematical model
The information in the mathematical model describes the selected compartmental kinetic model applicable to the real processes observed. This information concerns either the model structure (e.g: monocompartmental model) or the pharmacokinetics model parameters (e.g., the half life is 3 hours).
• Information about the influence of factors causing variation
This describes the changes in measurements, parameters, or processes related to the changes in experimental conditions (e.g., the passage through the hematoencephalic barrier is increased when there is meningeal inflammation; the drug half-life is higher for children; the unbound fraction is higher in patients with hypoalbuminemia).
Model Evaluation Results
We did not observe any significant difference between the corpus of texts (number of sentences) evaluated by each evaluator (p > 0.657) or between evaluator responses (p > 0.591 for completeness and p > 0.456 for semantic accuracy).
Evaluation indicated that the model gives a good representation of pharmacokinetics information: it was able to represent pharmacokinetics information completely in 89% of the cases (CI95 [83%–95%]), and in 11% of the cases in an “almost complete” way (CI95 [5%–17%]). For example, this sentence was judged as almost completely represented: “According to its main biliary excretion and to its important presystemic metabolism, an accumulation of fluvastatine is shown in patients having hepatic insufficiency.” It was translated into the model into classes describing biliary excretion, metabolism and accumulation but the causal link between the various components could not be represented. There were no examples of significantly or entirely defective representation.
The meaning of the information was classified “not distorted” in 98% of the cases (CI95 [95%–100%]) and “almost not distorted” in 2% of the cases (CI95 [0%–5%]). The following sentence was judged “almost not distorted”: “Metabolic transformation by hepatic microsome enzymes (inducible)” was translated into the model as: “Metabolism is a reaction, has hepatic area, has the generic site microsome, has the specific site enzyme, has the property inducible.” The deformation of meaning is on the inducibility that is linked to metabolism instead of enzyme. There were no cases of significant or entire distortion.
The evaluators agreed on 93 texts and disagreed on only seven texts. The use of Delphi method to solve these cases was effective. A consensus was quickly obtained for all seven texts after a new evaluation. Each evaluator saw 25 texts and judged 21–25 times that the information was completely represented, 0–4 times that the information was almost completely represented, 23–25 times that the information was not distorted, and 0–2 times that the information was almost not distorted.