Medical diagnosis as a linguistic game

Medicine involves doctors and patients discussing risk and reaching complex decisions together about treatment options, amongst other things. Yet it is known that diagnostic errors occur, with a frequency (depending on how terms are defined) ranging between 2.5% and 20% [1, 2]. Most errors have only moderate negative consequences, but some do not. Modern computing technology is claimed to decrease the number of erroneous diagnosis [3‐5]. Nowadays, numerous systems of collecting patient data (electronic history taking systems) exist. Only generalist publications are cited ([5‐16], for review [5, 16]). Our approach has as starting point the fact that the collection of patient data can be considered as a patient data vector. Assignment of the symptoms or signs of a patient to a certain diagnosis is possible by a bundle of different not exclusive methods: (1) by the knowledge of the physician, (2) by comparison of the patient data with clinical pathways and guidelines, (3) by boards of specialized physicians, (4) a second opinion by an experienced medical doctor, and (5) by computer-assisted comparison with a data base of medical knowledge. The approach proposed in (5) implies, that we use formalized systems allowing sampling of medical knowledge (prototypic diseases) in a formalized manner. We describe thereafter a database of prototypes of diseases, which are oriented on the concept of vectors (patient vectors and disease vectors). For this approach three constituents are necessary: (1) patient vector, (2) a pool of disease vectors (described in this paper) (3) an algorithm allowing an interaction between both vectors. For the approach mentioned in (2) and (3) clinical decision support systems (CDSS and inference engine technology) are available [17‐19]. Most approaches concerning computer-assisted data bases of medical knowledge are based on classical object-oriented structures and an entity-relationship-model. Substantial improvement of CDSS may be provided by IT-systems improving knowledge acquisition such as MetaMap [20] and cTake [21].

Our system shows that it is possible to use original atomistic language element and an intuitive, flexible semantic network to get medical context and structures which allow complex diagnostic understanding. The diagnosis process appears as a system driven dialogue which tries to match system semantics with patient semantics. This procedure can be understood as a “linguistic game”, which provides a “third opinion” for the patient’s symptoms.

Nowadays, it is common sense that both computational linguistics and computer technology will change the practice of medicine fundamentally [39]. However, the evidence, that this change the quality of medicine will improve is sparse [18].

Here, we describe a software system consisting of a pool of pathosoms (disease entities, syndromes, symptoms or sets of disease entities). The goal of this publication is to demonstrate that with some prerequisites a computer-assisted assignment of symptoms (in our preliminary system a 5n tuple) to certain disease entities is possible. The prerequisites are (1) linguistic knowledge is delivered atomistic (2) each pathosom (prototypic disease entity) is considered as a vector (3), and the results of the analysis should be a linguistic game assisting classification of a patient vector to different pathosoms without the use of defined mathematical methods or a distance measure. After the search function of Memem7 a list of possible pathosoms explaining the patient’s complaint is given, so far, with focus on linguistic and semantic network heuristics.

For a coming assistance system of decision making in medicine three different tools are mandatory: (1) an electronic history taking which can be transformed in a patient data vector [3‐5] or extracts of electronic health care systems (2) a set of disease vectors (in our system called pathosoms) each reflecting a prototypic disease vector or an algorithms for a disease entity. (3) an algorithm comparing the patient vector with all pathosom vectors giving a ranking of possible pathosoms fitting to a given patient vector (inference engine).

Our system termed Memem7 allows today: (1) to construct a prototypic disease vector (2) to have a restricted dialogue with a 5n-tuple of a patient vector. Restricted dialogue means that, so far, only 5 tuples can be analyzed by Memem7. After this dialogue, the user is provided with n pathosoms, which fit to the 5-tuple patient vector. The goal of our approach was to demonstrate the feasibility of a knowledge collection in a formalized way.

The mathematical methods used in big data analysis will be of further interest for improving the tools allowing machine assisted medical treatment decision as described by Sun and coworkers, Slack and coworkers and Huang and coworkers [40‐42]. Our system covers approximately 4600 medical diagnoses (increasing constantly). Many trials of electronic history taking, even in very sophisticated technical settings as described by Morrison and coworkers [43] are published (70 up to 2010) ([3, 4], for review [5, 16]). However, so far, as the authors survey the literature only few open source application of an electronic history taking system are available such as OpenEMR and OpenEHR [44, 45]. In our linguistic game these histories taking systems were replaced by a 5n-tupel. The term linguistic game has some fundamental implications (1) error is part of any game (2) decision can be assisted, but not replaced by a linguistic game (3) a player can improve his results by improving his knowledge (in our system the disease vector of the pathosoms). Therefore, our system may be considered as a third opinion to a given patient vector with so far 5 tuples. Combining memem7 with tools recognizing medical texts like MetaMap [20] or cTake [21] are currently in work.

Can we give any data for accuracy of memem7? In a first testing of the system with up to 5 tuple of search terms we identified the proposed diagnosis of 190 artificial cases for 46.7%. The data available for assessment of the sensitivity for clinical decision support system (CDSS) are sparse and difficult to compare. In one recent publication of Müller and coworkers [46] a sensitivity of 96% was claimed, when case reports of the New England Journal of Medicine were used for the evaluation of the systems (DXplain, Isabel Healthcare, Diagnosis Pro, PEPID) [47‐50]. Our much less impressive data will improve with each new test as the mistakes found in the test are corrected after completing the test. Therefore, the tests are part of a continuous improvement quality circle which provides algorithms to enhance both sematic network and pathophemes (disease symptoms).

What are the differences of Memem7 to other existing medical decisions systems like e.g. HELP [18]. There are different approaches for use of computer-technology in CDSS such as probabilistic, Bayesian approaches or machine learning systems. We add a system based on atomized linguistic terms. It may be possible to combine probabilistic and linguistic approaches. The advantage of a linguistic system is that atomized terms (pathophems) can be easily added to a disease entity. The comparison between a linguistic and a probabilistic approach or its combined use is planned on basis of the next software version.

There are already many systems and tools presented for biomedical text mining and biomedical concept recognition [51, 52]. The systems mainly focus on natural language processing and concept identification through dictionary matching and various machine learning mechanisms. In contrast to these approaches Memem7 is not focused on identifying concepts in natural languages but rather on using such concepts for identification of other complex objects like health disorders. In future, the system might also be able to handle the pre-process of concept identification of written data e.g. medical patient reports. Technically Memem7 is dealing with the problem of matching ambiguous fact concepts (like symptoms, medical events, lab data) with complex concept trees (like anatomy, medical body functions, diseases). Because of the plurivalent nature of the concepts and the heuristic of the algorithms the matching processing is more like a game than a deterministic process.

In conclusion, we present an electronic system of disease entities, which can enter in a formalized dialog with a patient data vector as proof of concept of a medical navigation system based on a linguistic game.