Cognition-Based Natural Language Processing and Search for Medicine

George J. ShannonEngineering Management and Systems Engineering

Missouri University of Science and TechnologyRolla, Missouri, USA

gjscnc@mst.edu

Steven M. CornsEngineering Management and Systems Engineering

Missouri University of Science and TechnologyRolla, Missouri, USA

cornss@mst.edu

Donald C. Wunsch IIElectrical and Computer Engineering

Missouri University of Science and TechnologyRolla, Missouri, USA

dwunsch@mst.edu

Abstract —. This paper introduces a simple yet effective measure for quantifying the cognitive relevance of search results as a distance measure used to improve search precision. It also introduces a new measure for improved precision of automated concept recognition. Both are based upon recent neurodynamics research we use to propose that a bijective relationship exists between the topology of the mental notions stored in the cerebral cortex and the concepts and relations in the ontology. This enables identifying cognitive neighborhoods using an ontology-based topological covering space. These neighborhoods define a cognition relevance measure not available with existing approaches. Furthermore, the cognitive-based search approach enables natural language search criteria. Such a natural language interface allows far more complicated criteria while at the same time remaining intuitive to the casual user. Preliminary testing by a physician indicates the potential for a significant improvement in search precision when using the complicated and large search criteria typical of medicine. Further test data is being gathered to confirm these results.

Index terms — ontology, search, search relevancy, semantic search, concept-based search, automated concept recognition, healthcare informatics

I. INTRODUCTIONSearch precision for medical text is known to be poor. For

example, tests have indicated that the precision of Google Scholar is only about 8% for medical text [1].

A simple cognitive relevance measure for improved search precision in medicine is presented here along a new measure for improved precision of automated concept recognition in text. To the authors’ knowledge, neither of these approaches developed by this research exists in the literature.

The use of a cognitive approach to search was chosen for improved precision due to its efficacy (see Moskovitch, et al., [2]). In this paper we discuss the theories and approaches used in our research to implement the mechanics of cognitive-based

search. This included theories of human cognition [3-6], cognitive linguistics [7], and ontology.

Using the ontology covering space for cognitive neighborhoods leverages the relationships in the ontology. This includes both subsumptive (“is-a”) and non-subsumptive relationships. This defines a neighborhood for the conceptual space between shared by the search criteria and documents in the corpus being searched.

II. COGNITION AND ONTOLOGY FOR SEARCHOntology, cognition theory, concept/semantic search, and

cognitive grammar are reviewed briefly to provide the rationale for use of the ontology structure to compute relevancy.

A. OntologyIn our research we used the SNOMED medical ontology.

Examples presented in this paper are from this ontology. The SNOMED ontology is a subset of the Unified Medical Language System (UMLS) [8] available from the National Library of Medicine (NLM).

The ontology structure is quite simple and is stored in two database tables as shown in Figure 1:

Concepts: unique identifier, name, and whether or not the concept is a relationship type

Relationships: from concept, to concept, and type of relationship. A relationship type is itself a concept.

Figure 1: SIMPLE ONTOLOGY PERSISTENCE SCHEMA - the ontology is stored in a simple relational database structure.

No predicate logic was required. Instead, the algorithm for computing relevancy leveraged ontology relationships alone,

and was limited to relationship types already existing in the ontology. For example, in Kumar, et al. [9], logic concerning relationships is used to extract an ontology that merges medical and biological information. This approach imputed new relationship types necessary for the merger. Our approach avoids these complexities to focus on measures that improve precision.

For our purposes, all relationship types define a cognitive neighborhood.

1) SNOMED Ontology Example

Take, for example, a relatively simple phrase in the healthcare domain (Figure 2).

Figure 2: SNOMED CONCEPT EXAMPLE – Example of a complex, multi-word concept that encompasses multiple conceptual categories.

The concept “dorsolumbar spinal fusion with Harrington rod” is one concept in the SNOMED ontology. This concept is included in the anatomy, procedure, and device categories (see Figure 3 below).

Figure 3: SIMPLIFIED ONTOLOGY SNIPPET - Small subset of cognitive neighborhood for the concept "dorsolumbar spinal fusion with Harrington rod" (provided by NLM’s Terminology Services https://uts.nlm.nih.gov/home.html).

Note that the number of concepts in the cognitive neighborhood for this example concept is far more than that shown in Figure 3. This example concept is a leaf of the ontology graph, and is related to over 100 related concepts at progressively higher, more abstract cognitive levels.

B. Theory of CognitionThe ontology structure essentially serves as a proxy for the

knowledge base stored in the cerebral cortex.

Confabulation theory [3-5] (see Figure 4) explains cognition as a process that accesses the neural codes and relationships in the cerebral cortex. This process is instantiated via thalamocortical links between the thalamus and cerebral cortex.

Confabulation theory is based upon evidence that the human cognitive process exists via cooperation between approximately 4,000 paired zones in the thalamus and the cerebral cortex (summarized in Figure 4). Zones of neurons in the thalamus and cerebral cortex reflect attributes of a conceptual notion, where an attribute is stored as a set of neurons in a cortical patch (typically ~ 60 neurons). Each set of neurons defines the neural code for a particular attribute. For example, a set of neurons in the patch for color attributes store the neural code for individual colors, e.g., blue. Excitation of such a set of neurons in turn fires signals that cascade to other neural groups via knowledge links.

The feed-forward firing continues until the most plausible ending neural code is reached (i.e., winner takes all). The final neural code in this chain will identify an action or conclusion.

Cogent confabulation [3] defines cogency as a conditional probability, where, for a set of assumed facts λ={α , β , γ , δ }, the most plausible conclusion ε is the one maximizing the probability:

ϵ=argmax ( p (αβγδ∨ε ))

If confabulation is applied to language cognition then αβγδ is a set of prior words in a phrase or sentence, and ϵ is any word likely found to occur after them in the word sequence.

Hecht-Nielsen, et al. [5, 10] reported results for sentence completion experiments that apply cogent confabulation via maximization of a proxy measure considered to be “approximate proportional” to cogency as follows:

p (αβγδ∨ε )∝ p (α∨ε ) p ( β∨ε ) p (γ∨ε ) p (δ∨ε )

ε=argmax ( p (α∨ε ) p ( β∨ε ) p (γ∨ε ) p (δ∨ε ) )

Figure 4: CONFABULATION THEORY - Confabulation theory predicts the outcomes of the fast, greedy, feed-forward neural network architecture composed of the cerebral cortex, thalamus, and knowledge links which find the most plausible. Adapted from [3, 5].

Notable about these experiments was the identification of plausibly logical, linguistically correct words that can be used to complete a sentence without the need for either linguistic rules or dictionaries (e.g., grammars, lexicons, or part-of-speech tags). Furthermore, these experiments demonstrated similar results when the set of assumed facts was extended to include prior sentences. Hecht-Nielsen concluded that

grammar and syntax “exist only as emergent properties of confabulation” [10].

C. Semantic MappingResearch by Huth, et al. [6] indicates a consistency in

semantic-word storage locations in the cerebral cortex. An example of the semantic map for the brain is shown in Figure5.

Figure 5: EXAMPLE OF CEREBRAL CORTEX SEMANTIC WORD MAP – This is an example mapping between words and storage locations in the cerebral cortex. Screenshot is from YouTube https://youtu.be/k61nJkx5aDQ.

This research also demonstrated that semantically similar words are co-located in the cerebral cortex. And furthermore, that when a word has different meanings, a storage location exists for each different meaning. That is, if a word has two possible meanings, then two storage locations exist in the

cerebral cortex for that same word, one for each different conceptual meaning.

One interpretation of these results is that storage location imputes a relationship between co-located concepts. And taking this interpretation further, the ontology could be viewed

as a proxy for the cerebral cortex semantic mappings and relationships.

D. Concept/Semantic Search and Concept RecognitionCognitive-based search involves searching for conceptual

notions in lieu of keywords. Conceptual notions represented by keywords can be ambiguous, for example, keywords can be arbitrarily grouped into multiple combinations that infer fundamentally different cognitive notions.

In our research we used the MetaMap tool [11-16] from the NLM to perform automated concept recognition for both the search criteria and corpus being searched.

E. Inverse Ontology CogencyThe precision of search is obviously impacted by the

precision of concept recognition. Analysis of concept recognition precision by MetaMap suggested that improvements may be possible.

To achieve this, the cogent confabulation approach was investigated and an inverse variant of cogency developed. The cogency calculation in Equation 4 is simply inverted:

inversecogency= 1p (α∨ε ) p ( β∨ε ) p ( γ∨ε ) p ( δ∨ε )

The logarithm form was chosen as follows:

IOC (concept x|wp )={−∑i=1

n−1

ln [ p (wi|wp ) ] for n>1

0 for n=1

whereIOC (concept x|wp ) is the inverse ontology cogency

for concept x using predicate word w p

n=|N concept x|, where N concept x is the ordered set of words for the concept’s name

w i∈N concept x is the assumed fact word where i≠ n,

w p∈N concept x is the predicate word, and

p (w i|wp ) is the conditional probability of assumed fact word w i and predicate word w p occurring in the same concept name.

We applied a 6:10:2 MLP with the input layer consisting of the following:

1. Fraction of words in concept name mapped to words in text.

2. Fraction of words in text mapped to words in concept.

3. Fraction of concept maximum possible inverse cogency that is mapped.

4. Fraction of text maximum possible inverse cogency that is mapped.

5. Whether or not the text consists of a single word (true = 1, false = 0).

6. Whether or not the concept consists of a single word (true = 1, false = 0).

This MLP is the function approximator used to rank all candidate concepts during the recognition process. Training and testing consisted of 10-fold reinforcement learning with back-propagation using hand-annotated data from the NLM, as shown in Table 1.Table 1: COUNT OF NLM HAND-ANNOTATED TEXT

Number of abstracts 592

Number of annotations 3,985

F. Cognitive LinguisticsWhen analyzing MetaMap output during testing it became

apparent that limiting the concept recognition process to the noun phrase level loses fidelity, which in turn can impact precision. That is, inter and intra-phrase relationships are not taken into account in the concept recognition process. Cognitive linguistics was explored to address this.

Cognitive linguistics [17] is based upon the premise that “language is governed by general cognitive principles, rather than by a special-purpose language module.” It proposes three major hypotheses, shown in Table 2, which appear consistent with the cogent confabulation theory per Hecht-Nielsen.Table 2: COGNITIVE LINGUISTICS AND COGNITION THEORY

Cognitive Linguistics

Cognition Theory (Confabulation)

Language is not an autonomous facility

Knowledge base stored in the greedy feedforward networks in cerebral cortex is used for all cognition, including language

Grammar is conceptualization

Neural patches contain the neural codes for attributes and conceptual notions (neural code exists for words, and via the feedforward networks, link to other patches representing cognitive notions in the knowledge base)

Language emerges from language use

Language experiments using sentence completion, based solely upon conditional probabilities computed from prior language use, i.e., frequency of use. This produced rational and linguistically correct sentences without the use of lexicon, grammar, or linguistic rules

Cognitive grammar, a topic within cognitive linguistics, involves conceptualization from words in a grammatical unit (i.e., a sentence). In general, conceptualization in cognitive grammar is aimed at understanding two things [7]:

1. Things – cognitive notions , usually nouns

2. Relations – cognitive notions that are usually verbs and adjectives

These two goals are functionally equivalent to the two components of ontologies: 1) entities as concepts, and 2) relationships between concepts (Figure 1).

Unfortunately, the field of cognitive grammar does not yet possess the computational approaches necessary for automation, and hence it was necessary to use MetaMap. Concept recognition using MetaMap is limited to nouns. Hence, research discussed in the remainder of this paper used MetaMap for concept recognition in noun phrases. More sophisticated approaches using cognitive grammar is a topic for future research.

III. TOPOLOGY COVERING SPACE AND COGNITIVE RELEVANCE

The first objective of this section is to define a topology covering space for the ontology such that cognitive neighborhoods can be identified. The second objective is to define a simple measure that quantifies how much two cognitive neighborhoods are related.

A. Topology Space and Ontology Cognitive NeighborhoodThe term neighborhood used in this paper refers to a

topological covering space of related concepts.

Definition 0: Per Willard [18], a topology on a set X is a collection τ of subsets of X , called the open set, satisfying the following:

1. Any union of elements of τ belong to τ ,

2. Any finite intersection of elements of τ belong to τ,

3. Ø and X belong to τDefinition 1: An ontology can be represented as a directed,

acyclic graph that consists of a set of concepts as vertices, C , and set of relationships, R, where each relationship is a directed connection between two concepts c¿ and c¿.

∃ f :Ontology→ DAG(C , R) (7)V= {c1 , c2 , …cn } (8)R={c¿ , c¿|c¿∈V ,c¿∈V } (9)

Definition 3: the distance d between two concepts in the ontology is the length of the shortest path P in the directed graph regardless of relationship type.

P (c¿ , c¿)is ordered set {c¿ , c2 ,…c¿} (10)d (c¿ , c¿)=argminP (c¿ ,c ¿) (|P (c¿ , c¿)|) (11)Definition 4 the neighborhood N for a concept c¿ consists

of itself plus any concept cn where d (c¿ , cn )>0.

N (c¿)= {c¿ , c1 , c2 ,…cn|d ( c¿ ,cn )>0} (12)

From these definitions the following holds:

Nont (c1 , c2 , …cn )=¿ i=1¿n Nont (c i ) (13)N shared (N ont1

, Nont2, … Nontm)=¿ i=1¿m Nont i(14)A topology τ ont exists for the ontology such that conceptual neighborhood Nont exists for each concept c∈C, and shared neighborhoods N shared exists for combinations of Nont . The union and intersection of neighborhoods, defined in Equations 13 and 14, along with existence of null and full sets, ∅ and C respectively, meet Definition 0 of a topology over the ontology when the ontology is represented as DAG (C , R ).

A simple cognitive relevance measure is defined as the relative size of the shared cognitive space for the criteria and text, compared to the criteria.

r=|N criteria∩ N text|

|N criteria| (15)Figure 6 provides an example of the two neighborhoods and calculation of cognitive relevance.

Figure 6: EXAMPLE OF SEARCH AND COGNITIVE NEIGHBORHOODS – The search criteria consist of one concept, dorsolumbar spinal fusion with Harrington rod (blue). Suppose a document includes two related concepts. The green area is the covering space for the document. Cognitive relevance is the 23 concepts in the shared neighborhood divided by the 127 concepts in the criteria neighborhood.

III. TESTPrecision calculations are as follows:

precison= tptp+ fp

wheretp=true positives, andfp=false positives

A. Cognitive Relevance Real-world patient profiles were used for preliminary

testing (final testing is underway in a similar manner). The patient profiles came from the clinical practice of a cardio-thoracic surgeon. All cases were modified to remove privacy information. A total of four cases were used, ranging in length from 3 to 5 paragraphs.

The patient profiles are the search criteria. Each profile was analyzed using MetaMap, resulting in a list of concepts used to identify the patient’s cognitive neighborhood N criteria.

Search keywords were identified by the physician for each patient. This list was used to perform a PubMed search as the baseline for comparison. All abstracts returned by this search are used as the corpus for the cognitive-based search. For preliminary testing the size of the corpus was limited to approximately 30,000. Concepts were identified for these abstracts using MetaMap.

Testing consisted of the physician reviewing 30 abstracts for each patient, and indicating whether or not the abstract was relevant or not. This was done for both the keyword search results and cognitive-based search results. Results are shown in Figure 7.

Figure 7: PRECISION OF SEARCH APPROACHES – Cognitive-based approaches indicate significant precision improvement potential.

The lower precision of keyword results is consistent with other findings [1], where the precision reported for PubMed was only 6%, far less than our results. As noted earlier, the precision results for Google Scholar was reported at 8%, about the same as that reported for PubMed.

B. Inverse Cogency for Concept Recognition Results from training and testing the MLP, using hand-

annotated text provided by the NLM, are shown in Figure 8.

Figure 8: PRECISION OF CONCEPT RECOGNITION MEASURE – Improved precision evident from using MLP with six inputs, two of which are IOC-base

Use of the six input values, including the inverse cogency measure, in both the MLP and a random forest approach produced similar results. The MLP approach produced a 5% improvement in precision over best available found in the literature for MetaMap. A local instance of MetaMap was used to perform independent tests. This resulted in a precision of only 50%, far lower than reported in the literature.

IV. CONCLUSIONSThis research demonstrated the use of simple yet effective

measures for improving the precision of search and natural language processes using theories of cognition. Further work is warranted to develop more sophisticated approaches necessary to address more demanding problems. Cognitive-based methods offer an effective alternative to legacy natural language and search approaches.

References

[1] Z. S. Shariff, A. D. S. Bejaimal, M. J. Sontrop, V. A. Iansavichus, B. R. Haynes, A. M. Weir, et al., "Retrieving Clinical Evidence: A Comparison of PubMed and Google Scholar for Quick Clinical Searches," J Med Internet Res, vol. 15, p. e164, 08/15 2013.

[2] R. Moskovitch, S. B. Martins, E. Behiri, A. Weiss, and Y. Shahar, "A Comparative Evaluation of Full-text, Concept-based, and Context-sensitive Search," Journal of the American Medical Informatics Association : JAMIA, vol. 14, pp. 164-174, Mar-Apr

[3] R. Hecht-Nielsen, "Cogent confabulation," Neural Networks, vol. 18, pp. 111-115, 3// 2005.

[4] R. Hecht-Nielsen, Confabulation Theory The Mechanism of Thought. LaJolla, California: Springer, 2007.

[5] S. Solari, A. Smith, R. Minnett, and R. Hecht-Nielsen, "Confabulation Theory," Physics of Life Reviews, vol. 5, pp. 106-120, 6// 2008.

[6] A. G. Huth, W. A. de Heer, T. L. Griffiths, F. E. Theunissen, and J. L. Gallant, "Natural speech reveals the semantic maps that tile human cerebral

cortex," Nature, vol. 532, pp. 453-458, 04/28/print 2016.

[7] G. Radden and R. Dirven, Cognitive English Grammar. Amsterdam, The Netherlands: John Benjamins Publishing Company, 2007.

[8] NLM. (2013). Unified Medical Language System (UMLS). Available: http://www.nlm.nih.gov/research/umls/quickstart.html

[9] A. Kumar, Y. L. Yip, B. Smith, and P. Grenon, "Bridging the gap between medical and bioinformatics: An ontological case study in colon carcinoma," Computers in Biology and Medicine, vol. 36, pp. 694-711, 7/2006.

[10] R. Hecht-Nielsen, "The Mechanism of Thought," in Neural Networks, 2006. IJCNN '06. International Joint Conference on, 2006, pp. 419-426.

[11] A. R. Aronson, "The effect of textual variation on concept based information retrieval," Proceedings : a conference of the American Medical Informatics Association / ... AMIA Annual Fall Symposium. AMIA Fall Symposium, pp. 373-377, // 1996.

[12] A. R. Aronson, "Effective mapping of biomedical text to the UMLS Metathesaurus: the MetaMap

program," Proceedings / AMIA . Annual Symposium. AMIA Symposium, pp. 17-21, // 2001.

[13] A. R. Aronson, "MetaMap: Mapping Text to the UMLS Metathesaurus," UMLS White Paper, 2006.

[14] A. Aronson, "The Current State of MetaMap and MMTX," 2009.

[15] N. H. Shah, N. Bhatia, C. Jonquet, D. Rubin, A. P. Chiang, and M. A. Musen, "Comparison of concept recognizers for building the Open Biomedical Annotator," BMC Bioinformatics, vol. 10, p. S14, 09/17/2009.

[16] A. R. Aronson and F. M. Lang, "An overview of MetaMap: Historical perspective and recent advances," Journal of the American Medical Informatics Association, vol. 17, pp. 229-236, 2010.

[17] W. Croft and D. A. Cruse, Cognitive Linguistics: Cambridge University Press, 2004.

[18] S. Willard, General Topology. Mineola, New York: Dover Publications, Inc., 1970.