Health Articles

 

Hlt.utdallas.edu

Clinical Data-Driven Probabilistic Graph Processing Travis Goodwin and Sanda Harabagiu Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688, USA Electronic Medical Records (EMRs) encode an extraordinary amount of medical knowledge. Collecting and interpreting this knowledge,however, belies a significant level of clinical understanding. Automatically capturing the clinical information is crucial for performingcomparative effectiveness research. In this paper, we present a data-driven approach to model semantic dependencies between medicalconcepts, qualified by the beliefs of physicians. The dependencies, captured in a patient cohort graph of clinical pictures and therapies isfurther refined into a probabilistic graphical model which enables efficient inference of patient-centered treatment or test recommendations (based on probabilities). To perform inference on the graphical model, we describe a technique of smoothing the conditional likelihood of medical concepts by their semantically-similar belief values. The experimental results, as compared against clinical guidelines are verypromising.
Keywords: Information Retrieval, Bioinformatics, Patient Cohort the assertions formulated by physicians when discussing anyof the medical concepts.
An increasing abundance of clinical data is available through The 2010 i2b2/VA challenge evaluated the task of automati- massive warehouses of Electronic Medical Records (EMRs).
cally inferring six types of assertions, or belief states, used Both within the United States and across the world, hospitals to qualify medical problems in EMRs generate millions of EMRs each year. These EMRs include However, those assertions correspond to clinical information rich clinical information, consisting of detailed notes on found in only one type of EMR: discharge summaries. Be- patients' medical history, physical exam findings, lab re- cause we consider more types of EMRs, we have extended ports, radiology reports, operative reports, and discharge the problem of classifying medical assertions by consider- summaries. Clinical information contains multiple men- ing additional types of assertions. The new assertion values tions of medical problems, including observations resulting were selected based on discussions with practicing clini- from a physical exam (known as signs), features that the cians, and by following the guidelines outlined in patient observed first-hand (known as symptoms), historical and present medical problems (known as co-morbidities), in Medical concepts and their assertions were cast as nodes addition to diagnostic information. We have used the onto- in a graph which encodes a patient's clinical picture and logical definitions of medical concepts related to diseases therapy along with the potential dependencies between outlined in to capture the seman- them. We called this graph the clinical graph (CG). As tics of clinical information. Hence, we have considered the in the clinical picture is defined fact that EMRs also document the medical interventions per- as the clinical which contains the clinical findings formed during the patient's hospital stay, including medical (e.g. medical problems, signs, symptoms and tests). Like- tests and their results, as well as all the medical treatments wise, we use Scheuermann's definition of therapy as all the performed as part of the patient's therapy. These forms of treatments, cures, and preventions included within the man- clinical information are crucial for performing comparative agement plan for an individual patient. Figure illustrates effectiveness research. As shown in our representation of the CG for a patient. Given the pa- capturing the clinical information from EMRs enables the tient's hospital visit, we automatically discover the medical discovery of alternative methods to prevent, diagnose, treat, problems along with the tests and treatments documented or monitor a medical problem.
during the patient's hospital course. Medical problems, tests, It has been shown that clinical information – medical con- and treatments are qualified by their assertions and con- cepts (e.g. problems, tests and treatments) – can be automat- nected by their dependencies (e.g. when cellulitis was a ically identified from clinical texts, as described in present diagnostic, a blood culture test was conducted).
However, because medical science centers Moreover, as reported in the around asking hypotheses, experimenting with new methods clinical picture may vary widely between patients with the of care, and evaluating medical evidence, medical concepts same disease and even for the same patient during the course are associated with different degrees of belief, or assertions.
of his or her diseases. Therefore, in order to capture the vari- As such, clinical writing entails a large number of specula- ation in the corresponding clinical graphs (CGs), we have tive statements indicating the physician's belief at the time,rather than strictly quantifying a fact. In order to take into 1While the clinical phenotype refers to the set of observations account the physicians' beliefs when automatically process- related to a medical condition, the clinical phenome is the set of ing the clinical information from EMRs, we also recognized observations pertaining to a single patient.
Hospital Visit Medical Problems Diagnostic & Co-Morbidities
Treatments Atrial Fibrillation Figure 1: The Clinical picture & therapy Graph (CG).
Hospital Visits Clinical Picture & Therapy Medical Problems Treatments Clinical Picture & Therapy Medical Problems Treatments Clinical Picture & Therapy Medical Problems Treatments Clinical Picture & Therapy Medical Problems Treatments Clinical Picture & Therapy Medical Problems Treatments Figure 2: The combined Cohort Clinical Graph (CCG).
considered a patient cohort which we obtained by using the k-partite graph (where k = 4) because there are four types of system reported in Patient nodes (V, M, E and R), as illustrated in Figure It is to be cohort retrieval results in an ordered set of hospital visits noted that the edges from the CCG originate from the CGs which correspond to a cohort of patients sharing the same of patients from the cohort. We also noticed that, crucially, diagnosis (e.g. patients with As illustrated in Fig- the CCG can also be viewed as a factorization of a Markov ure this enabled us to access all the clinical pictures and network. In this way, we were able to transform the CCPT therapies from all the clinical graphs (CGs) of all patients into a probabilistic graphical model. Probabilistic graphical within a cohort. This clinical information regarding a patient models are known to be a cohort constitutes the set of all hospital visits (V), the set of state-of-the-art representation for producing probabilistic all medical problems (M), the set of all medical tests (E), inference, which we used for finding recommendations for and the set of all treatments (R), across the CGs of all the the most adequate tests or treatments for a patient, given patients belonging to the cohort. We refer to the graph that inference on the CCG.
combines all CGs as the Cohort Clinical Graph (CCG).
The remainder of this paper is organized as follows. In Sec- Given a patient cohort, the corresponding CGG was cast as a tion 2, we describe the clinical language processing requiredfor generating the CGs. Section 3 describes the construction 2Abscess is an infectious disease of the skin and soft tissue.
of the CCG, as well as how it can be transformed into a prob- abilistic graphical model. Section 4 presents the inference well as our own set of 2,349 EMR annotations. As illus- mechanisms we considered and how they may be used for trated in Figure we incorporated knowledge from many clinical test and treatment recommendation. Section 5 dis- lexico-semantic resources. In this research, we used the cusses the experimental results, and Section 6 summarizes feature set reported in Addi- the conclusions.
tionally, we have normalized the detected medical conceptsby (1) converting the surface string to lowercase, (2) filtering Medical Language Processing words belonging to words, and (3) ignoringword order.
Open-source software, such as MetaMap or,more recently, cTakes can parse EMRs Medical Assertion Classification to determine concept unique identifiers (CUIs) which corre- In order to encode the medical knowledge from EMRs with spond to entries in the Unified Medical Language System the clinical graph (CG) of each patient, we needed to au- (UMLS) However, UMLS includes tomatically qualify each medical concept with one of the many concepts that were authored according to ontological assertions given in Table We performed this automatic principles and, thus, it is too fine-grained for our purpose classification using an SVM classifier which considers in- of data-driven probabilistic processing of EMRs. In select- formation from: (a) the medical concept to be classified, ing a conceptual representation, we also evaluated the more (b) the section header where the assertion is implied, (c) general frameworks developed by the i2b2/VA challenge in features available from UMLS (extracted by MetaMap), (d) 2010 This framework was designed features reflective of negated statements, disclosed through to detect medical concepts within clinical text and assign the NegEx negation detection package, and (e) belief values one of several distinct assertions indicating the state of the are available from the Harvard General Inquirer's category author's belief for each concept. This i2b2 challenge helped information Additional details of the popularize the notion that recognizing medical concepts automatic assertion identification techniques are provided alone is not sufficient for clinical reasoning, because, when medical concepts are used in clinical texts, physicians alsoexpress their belief state about such concepts, e.g. that a Generating the Graphical Model medical problem is present or absent, that a treatment is con- For clinical decision support, it is critical to analyse the ditional on a test. The i2b2 challenge, however, considered relationships between medical problems, medical tests, and assertions only for medical problems. In our aim to build the associated treatments across patients' hospital visits. As CCG, we have extended the problem of assertion classifica- such, we must move beyond merely identifying the textual tion in two ways: (1) we have produced assertions (or belief mentions of medical concepts and their associated belief values) for all medical concepts (including treatments and values. To this end, we present a framework for modelling tests) that we have automatically identified; and (2) we have the data-driven interactions between problems, treatments, introduced 6 additional values which are defined in Table and tests. We first create a CG in which connections betweenmedical concepts are not only inferred, but their strength Medical Concept Recognition is also quantified by a weight. Because of the economy To recognize the nodes of the CCG, we have partitioned of language, relations between medical concepts are rarely medical concepts within three categories: (1) medical prob- explicitly stated, but they are rather implied. To capture lems (e.g. ATRIAL FIBRILLATION – an irregular heart beat); these implications, we postulate that co-occurrence statistics (2) medical treatments (e.g. ABLATION – the removal of can inform these relations, and further that they can also undesired tissue); and (3) medical tests (e.g. ECG – an elec- inform the strength of these relations.
trocardiogram). We detect these medical concepts using the After we create complete CGs, we can then transform the methods reported in Further, combined CGs for a cohort of patients (the CCG) into a we distinguish three sub-classes of medical problems: (a) probabilistic graphical model.
signs (observations from a physical exam), (b) symptoms (observations by the patient), (c) co-morbidities (diseases or Inferring Edges in the Cohort Clinical disorders), and (d) the diagnostic. Our method recognizes medical concepts in three steps: The nodes of the CCG are automatically discovered by the Step 1: Identification of the boundaries within text that language processing techniques described in Section 2. In refers to a medical concept; addition, we needed to infer the edges of the CCG and the Step 2: Classification of the medical concept into (1) medi- weights of the edges indicating semantics used in the clinical cal problems, (2) medical treatments, or (3) medical picture and therapy ontological definition. The observations from the clinical picture of a patient connected hospital Step 3: Classification of medical problems into (a) signs, visit (or nodes from V) to the observed medical problem (or (b) symptoms, (c) co-morbidities, or (d) diagnos- nodes from M) generating edges of type TVM. In the clin- ical picture of patients, connections between the observed Medical concepts were recognized both within the narrative (i.e. report text) and structured sections (e.g. CHIEF COM- 3In linguistics, a closed-class of words is a class of words for PLAINT) of EMRs. To do this, we used two conditional which new words are rarely introduced, for example pronouns, random fields (CRFs), trained on the i2b2 annotations as determiners, prepositions, etc.
the patient's past medical history is signif- occurred during a previous hospital visit icant for CONGESTIVE HEART FAILURE readmit him for REHAB once the WOUND occurs only during certain conditions she was given ROCEPHIN and ZITHRO- has been assigned and will occur the patient denies any CHEST PAIN at this was recommended that he be on ALLOP- has been advised, but cannot be assumed to occur there is a moderate PERICARDIAL EFFU- is currently happening may occur in the future she is to return for any WORSENING PAIN we will do a PULMONARY FUNCTION has been scheduled and will occur in the future not associated with the patient father died of LUNG CANCER I believe that this may represent worsen- may occur, but there is uncertainty ing for PULMONARY HYPERTENSION currently exists and can be assumed to persist continue DIALYSIS has been performed and completed UNASYN 3 GRAMS IV was given Table 1: Assertion values for medical concepts (typeset in SMALLCAPS) in each excerpt; "moment" refers to the specific instant when the medical concept was mentioned. Newly defined assertions are marked with an ‘*'.
Medical Concept Type Classifier Prose Concept Pattern-based Entity
Boundary Detector Non-prose Concept Problem Test Treatment
Boundary Detector Section Header Extractor External Resources for Concept Classification
Medical Assertion Lemmas, Part-Of-Speech Tags, Phrase Chunks, PropBank –Based External Resources for Assertion Classification
Figure 3: Language processing used for constructing the CGs and CCG.
medical problems (i.e. nodes from M) and results of tests in the CCG between medical problems (i.e. nodes from M) (i.e. nodes from E) exist as well, giving rise to edges of and treatments (nodes from R), generating edges of type type TME in the CCG. In addition, connection between both TMR. Similarly, we have edges between tests (i.e. nodes types of nodes (medical problems and tests) in the clinical from E) and treatments (nodes from R), generating edges of picture and therapies exist. Thus, we shall also have edges type TER. The weight of edges of each type is computed as the partition function, as given in Equation • The weight of an edge of type TVM between a visit v ∈ V and a medical problem m ∈ M is computed as the number of EMRs associated with v which also 1(v, m)Φ2(m, e)Φ3(m, r)Φ4(e, r) • The weight of an edge of type TME between a medical Probabilistic Inference problem m ∈ M and test einE is computed by the By modelling the CCG as a probabilistic graphical model, number of EMRs in which both m and e co-occur we have gained access to an incredible breadth of proba- (regardless of the patient).
bilistic information through the power of probabilistic infer- • The weight of an edge of type TMR between a medical ence. We can use this probabilistic information to construct problem m ∈ M and treatment rinR is computed by a recommendation engine enumerating the most probable the number of EMRs in which both m and r co-occur treatments for a given patient given their medical problems (regardless of the patient).
and/or their medical tests.
• The weight of an edge of type TER between a test We can use this joint distribution to calculate posterior prob- e ∈ E and treatment rinR is computed by the number ability of conducting a medical test during a particular pa- of EMRs in which both e and r co-occur (regardless of tient's hospital visit (i.e. P (E = e V = v)) as shown in the patient).
The Probabilistic Graphical Model Φ2(e, m)Φ1(v, m) In Section 3.1 we presented a co-occurrence-based method of building a cohort clinical graph (CCG). The observation Likewise, we can infer the posterior distribution of med- that this graph is in fact a k-partite graph (where k = 4) ical treatments for a given set of N medical problems, enables us to build the factorized Markov network illustrated m0, m1, . . , mN ∈ M , as the conjunction of each prob- in Figure which we call the Clinical Markov Network lem's posterior distribution, as shown in Equation Φ3(mi, r)Φ2(mi, r) Although this straightforward approach yields precise re- sults, it suffers from significant sparsity problems induced by our decision to qualify all medical concepts by the physi-cian's belief state. Rather than restricting ourselves to theinteractions between concepts exactly matching the speci- fied belief states (e.g. the likelihood that a test is conductedgiven than a problem is present), we also consider the inter-action between the same concepts with semantically similar Figure 4: The factorized Clinical Markov Network (CMN).
belief states (e.g. suggested, ordered, prescribed, condi-tional). For example, consider that assertions ONGOING andCONDUCTED both imply a strong degree of certainty that In the CMN, we assume that each vertex class (V, M, the medical concept occurred and are likely to have simi- E, or R) represents a distinct random variable in the in- lar semantic relationships despite having different temporal duced Markov network. Similarly, each of the four types of groundings. Thus, they are semantically coherent. Based weighted edges (TVM, TME, TMR, TER) have associated four on this observation, we introduce an assertion smoothing different factors to indicated the strength of the edge in the factor, S, that encodes the degree to which two assertions are semantically coherent, as given in Equation • Φ1(v, m) = weight of edge {v, m} ∈ TVM• Φ2(m, e) = weight of edge {m, e} ∈ TME• 3(m, r) = weight of edge {m, r} ∈ TMR 4(e, r) = weight of edge {e, r} ∈ TER This factorization allows us to perform efficient probabilis- tic inference by defining the joint probability as the Gibbs distribution given in Equation This smoothing factor, S(a1, a2), captures the degree bywhich occurrences for a certain medical concept labeled with the assertion a 2 may be relevant to probabilistic queries targeting the same medical concept with assertion a1. We Φ3(m, r)Φ4(e, r) estimate this value as the number of two-step paths in theCMN from any concept with assertion a1 to any concept Note that Z is the typical normalization constant equal to with assertion a2.
This assertion smoothing factor allows us to make recom- mendations for a query concept given an evidence concept (e.g. P (qc, qa)  (ec, ea)
), by considering information across all belief values weighted by their semantic similarityto the given belief values. We accomplish this by smooth- ing the co-occurrence probability as a mixture model ofthree components as shown in Equation (1) the directprobability, P , that the exact concepts co-occurred; (2) thetotal probability that the exact query concept co-occurred with the evidence concept qualified by any possible asser- tion (i.e. P P (q  (ec, ai)
, scaled by the smooth- ing factor between the encountered evidence assertion andthe desired evidence assertion, i.e. S(qa, ai); and (3) the Figure 5: Distribution of edges in the CCG.
total probability that the query concept qualified by anyassertion co-occurred with the exact evidence concept (i.e.
Cellulitis & Abscess  (ec, ea)
, scaled by the smoothing factor between the encountered query assertion and the desired query assertion, i.e. S(ai, ea).
Table 2: Precision and accuracy for the top 15 treatments P ((c, a) (d, b); δ) = for each cohort.
 λ0P (c, a)(d, b)
We annotated these EMRs with the medical concepts and P (c, a)(d, β); δ − 1
S(b, β) assertions described in Section 2.
By automatically processing the medical language in this P (c, α)(d, b); δ − 1
S(α, a) subset of EMRs, we were able to generate the Clinical Markov Network (CMN) described in Section 4, which  P (c, a)(d, b)
corresponds to a cohort of patients with cellulitis or abscess.
The distribution of edge classes in the CMN for these cohorts In order to limit the length of transitive paths considered, we is not uniform, as illustrated in Figure introduce a limiting parameter, δ, which limits the recursive Figure plots the distribution of edges in the CCG by type.
depth by which medical concepts will be smoothed (if δ = 0, Note that the distribution of edges in the CCG corresponds no smoothing will occur). This smoothing allows us to to the un-normalized probability mass of each factor in the predict the likelihood of a certain medical test or treatment CMN. It is clear from this distribution, that the majority of for a given patient by considering the dependencies encoded edges involve medical problems, with a nearly equal num- in the EMRs across all assertion values without disregarding ber of inferred dependencies between medical problems and the semantics of each assertion.
tests. In Figure the number of edges between medicalproblems and tests, T Experimental Results ME (denoted as M ↔ E), and between medical problems and treatments, TMR, denoted as M ↔ R, To produce the data-driven Clinical Markov Network are nearly equal. As such, the number of edges between med- (CMN), we used the same EMRs that enabled us to build a ical tests and treatments, TER, denoted as E ↔ R, makes patient cohort retrieval system for the medical records track up a smaller portion, indicating that there are an abundance (TRECMed) of the Text REtrieval Conference (TREC) in of medical problems listed in each EMR. This reinforces to the fact that physicians typically document all the historical, This dataset includes 95,703 de-identified possible, and related or even unrelated medical problems EMRs which were generated from multiple hospitals during observed during a patient's physical or other examinations.
2007. The EMRs were grouped into hospital visits con- In order to evaluate the validity of the inference that the sisting of one or more medical reports from each patient's CMN enables, we asked two inferential questions: (1) "what hospital stay. Thus, the EMRs were organized into 17,199 are the most probable medical treatments for a certain pa- different patient hospital visits. Each visit had the patient's tient cohort?" and (2) "which tests are most likely to be admission diagnoses, discharge diagnoses, and related ICD- conducted on patients with the given medical problem(s)?".
9 codes. We also used the 826 discharge summaries used We answered the first question by computing the conditional during the 2010 i2b2/VA challenge which contained 72,896 probability distribution for all treatments conditioned on medical concepts and their assertions.
the medical problems associated with the cohort retrieved As illustrated in Figure in addition to the hospital visits for Q1, Q2, and Q3. These probability distributions are and associated EMRs, we have also used annotations which computed according to Equation we produced on the EMRs resulting for three patient co- The second question was answered by calculating the condi- horts targeted by the queries (Q1) "patients who presented tional probability distribution over all tests conditioned on with cellulitis," (Q2) "patients diagnosed with abscess," and the hospital visits associated with each cohort, as computed (Q2) "patients suffering from both cellulitis and abscess." Cellulitis & Abscess
Treatments
Treatments
Treatments
vancomycin/ONGOING vancomycin/ONGOING vancomycin/ONGOING emergency department/ONGOING 12.61% linezolid/ONGOING procedure/CONDUCTED emergency department/ONGOING procedure/CONDUCTED linezolid/ONGOING eradication protocol/ONGOING emergency department/ONGOING eradication protocol /ONGOING procedure/CONDUCTED drainage/CONDUCTED drainage/CONDUCTED antibiotics/ONGOING iv dilaudid/ONGOING lisinopril/ONGOING antibiotics/ONGOING pain control/ONGOING vanco/HISTORICAL protonix/ONGOING protocol/ONGOING ibuprofen/ONGOING prednisone/ONGOING drainage/ONGOING ⋮ (12 rows omitted) pressure blood/CONDUCTED blood pressure/CONDUCTED physical examination/CONDUCTED 11.39%
vital signs/CONDUCTED vital signs/CONDUCTED pressure blood/CONDUCTED temperature/CONDUCTED temperature/CONDUCTED systems review/CONDUCTED systems review/CONDUCTED physical examination/CONDUCTED 6.20%
vital signs/CONDUCTED systems review/CONDUCTED palpation/CONDUCTED hemoglobin/CONDUCTED temperature/CONDUCTED respirations/CONDUCTED palpation/CONDUCTED auscultation/CONDUCTED creatinine/CONDUCTED creatinine/CONDUCTED ⋮ (3 rows omitted) auscultation/CONDUCTED physical exam/CONDUCTED
Figure 6: Treatment and test recommendations for present medical problems "cellulitis", "abscess", and both "cellulitis &abscess." The distributions of the 15 most-likely treatments and 10 The most common treatment across all patient cohorts is most-likely tests for each cohort are illustrated in Figure Vancomycin which is the most recommended treatment for We have evaluated the recommendations, as shown in Ta- methicillin-resistant Staphylococcus aureus (MRSA), the ble based on (1) the Infectious Diseases Society of Amer- most common cause of cellulitis and abscess. However, ican (IDSA)'s Practice Guidelines for the Diagnosis and after Vancomycin, the treatment distributions begin to dif- Management of Skin and Soft-Tissue Infectious fer. We have highlighted the treatment Zosyn (a mixture (2) Howe and Jones Guidelines for the Man- of Piperacillin and Tazobactam) which is an antibiotic ap- agement of Periorbital Cellulitis/Abscess proved to treat for infections such as cellulitis and abscess.
(3) Uzcategui et. al's Clinical Practice Guidelines Despite being commonly given to patients with cellulitis for the Management of Orbital Cellulitis (4.46%, the second highest-ranked treatment), it is ranked and (4) the National Library of Medicine's MED- twentieth for treating abscess, at only 0.49%. This corre- LINEplus Web Service sponds to the most typical treatment for abscessing concern-ing draining the cyst, corresponding to entries four and six.
According to these sources, we achievement a precision Additionally, more general antibiotics, such as Linezolid and within the first 15 treatments of 50% for cellulitis, 71% for Ciprofloxacin are more commonly given for abscess, as they cellulitis & abscess, and 64% for abscess. In this measure- treat a variety of underlying infections.
ment, we considered a treatment as relevant if it should bedirectly associated with the patient cohort. Note: we do not However, for the cohort of patients suffering from both consider treatments for associated symptoms (e.g. pain) as conditions, Zosyn rises to position 7 at 1.83% reflecting relevant. Additionally, because precision does not take into the fact that it is able to effectively treat both conditions.
the probability associated with each item, we have also cal- This shows the ability of the CMN to capture the interaction culated the accuracy of each distribution as the proportion between treatments for combinations of medical problems.
of probability mass assigned to relevant treatments. Using As our dataset is represented by primarily hospitalized pa- this definition, we achieve an accuracy of 58.2% for celluli- tients (rather than outpatient procedures), many of the rec- tis, 98.1% for cellulitis & abscess, and 83.6% for abscess.
ommended treatments are general purpose medications per- Before discussing specific treatments, we list the following scriped during the patients hospital stay, such as pain reliev- abridged definitions from MEDLINEplus: ers (e.g. aspirin, ibuprofen, pain control), stool softeners abscess a pocket of white blood cells, germs, and dead (e.g. colace), diaretics (e.g. lasix) and blood thinners (e.g.
tissues on the skin resulting from an infection.
cellulitis an infection of the skin and underlying tissues We have also evaluated the top 10 tests most likely to be caused by bacteria (typically streptococcal).
conducted for patients in each cohort, as illustrated in Fig- ure We observed that the likelihood of conducting a Howe, L. and Jones, N. (2004). Guidelines for the manage- physical examination has a distribution rank which varies ment of periorbital cellulitis/abscess. Clinical Otolaryn- across all cohorts. Although it is ranked second for cel- gology & Allied Sciences, 29(6):725–728.
lulitis (at 11.39% likelihood), it is ranked much lower for Koller, D. and Friedman, N. (2009). Probabilistic graphical abscess at position 12 (at 2.12% likelihood). This reflects models: principles and techniques. MIT press.
the recommendation in the guidelines for cellulitis: because Miller, N., Lacroix, E.-M., and Backus, J. E. (2000). Med- cellulitis leaves a patient vulnerable to secondary conditions, lineplus: building and maintaining the national library of a thorough physical examination should be performed. As medicine's consumer health web service. Bulletin of the such, for patients suffering from both cellulitis & abscess, Medical Library Association, 88(1):11.
the likelihood of conducting a physical examination moves Ratner, R., Eden, J., Wolman, D., Greenfield, S., and Sox, up to rank 5 (6.20%), reflecting the interaction between the H. (2009). Initial national priorities for comparative two conditions in EMRs.
effectiveness research. National Academies Press.
We also observed that the first three most-commonly con- Roberts, K. and Harabagiu, S. (2011). A flexible framework ducted tests (i.e. blood pressure, pulse, and vital signs) for deriving assertions from electronic medical records.
constitute the majority of the probability mass. This reflects Journal of the American Medical Informatics Association, a critical observation on the utility of medical test annota- tions: that the mere mention of a medical test is not sufficient Savova, G. K., Masanz, J. J., Ogren, P. V., Zheng, J., Sohn, for statistical reasoning. EMRs document a wide battery of S., Kipper-Schuler, K. C., and Chute, C. G. (2010). Mayo tests and their results for each patient allowing physicians to clinical text analysis and knowledge extraction system ascess not only their primary medical problem, but also any (ctakes): architecture, component evaluation and appli- secondary conditions or co-morbidities. In order to improve cations. Journal of the American Medical Informatics the capability of clinical reasoning enabled by the CMN, the value of tests should be considered and associated with the Scheuermann, R. H., Ceusters, W., and Smith, B. (2009).
identification of the mention of each test.
Toward an ontological treatment of disease and diagnosis.
Proceedings of the 2009 AMIA Summit on Translational In this paper, we show how medical language processing Stevens, D. L., Bisno, A. L., Chambers, H. F., Everett, enables the automatic derivation of clinical pictures and E. D., Dellinger, P., Goldstein, E. J., Gorbach, S. L., therapies for entire patient cohorts. We explain how this Hirschmann, J. V., Kaplan, E. L., Montoya, J. G., et al.
knowledge can inform a data-driven probabilistic graphical (2005). Practice guidelines for the diagnosis and manage- model on which inference can be performed in a rigorous ment of skin and soft-tissue infections. Clinical Infectious way for determining the most probable treatments for a given set of medical conditions. Further, we observe that Stone, P. J., Dunphy, D. C., and Smith, M. S. (1966). The the utility offered by medical test mentions is limited for general inquirer: A computer approach to content analy- probabilistic reasoning. Despite this, we evaluated the most likely treatments against (1) the Infectious Diseases Society Uzcategui, N., Warman, R., Smith, A., and Howard, C.
of American (IDSA)'s Practice Guidelines for the Diagnosis (1997). Clinical practice guidelines for the management and Management of Skin and Soft-Tissue Infectious of orbital cellulitis. Journal of pediatric ophthalmology (2) Howe and Jones Guidelines for the Man- and strabismus, 35(2):73–9.
agement of Periorbital Cellulitis/Abscess O., South, B. R., Shen, S., and DuVall, S. L. (2011).
(3) Uzcategui et. al's Clinical Practice Guidelines 2010 i2b2/va challenge on concepts, assertions, and re- for the Management of Orbital Cellulitis lations in clinical text. Journal of the American Medical and (4) the National Library of Medicine's MED- Informatics Association, 18(5):552–556.
LINEplus Web Service and confirmed Voorhees, E. and Hersh, W. (2012). Overview of the trec the validity the probabilistic information encoded by our 2012 medical records track. In The Twenty-First Text RE- trieval Conference Proceedings (TREC 2012), Gaithers-burg, MD. National Institute for Standards and Technol- ogy. Unpublished. Draft available at http://trec.nist.gov/.
Aronson, A. R. (2001). Effective mapping of biomedical Voorhees, E. and Tong, R. (2011). Overview of the trec text to the umls metathesaurus: the metamap program. In 2011 medical records track. In The Twentieth Text RE- Proceedings of the AMIA Symposium, page 17. American trieval Conference Proceedings (TREC 2011), Gaithers- Medical Informatics Association.
burg, MD. National Institute for Standards and Technol- Bodenreider, O. (2004). The unified medical language sys- tem (umls): integrating biomedical terminology. Nucleicacids research, 32(suppl 1):D267.
Goodwin, T. and Harabagiu, S. M. (2013). The impact of belief values on the identification of patient cohorts.
In Information Access Evaluation. Multilinguality, Mul-timodality, and Visualization, pages 155–166. SpringerBerlin Heidelberg.

Source: http://www.hlt.utdallas.edu/~travis/papers/lrec_2014.pdf