Research Article Association Patterns of Ontological Features Signify Electronic Health Records in Liver Cancer

Transcription

1 Hindawi Journal of Healthcare Engineering Volue 2017, Article ID , 9 pages Research Article Association Patterns of Ontological Features Signify Electronic Health Records in Liver Cancer Lawrence W. C. Chan, 1 S. C. Cesar Wong, 1 Choo Chiap Chiau, 2 Tak-Ming Chan, 2 Liang Tao, 2 Jinghan Feng, 2 and Keith W. H. Chiu 3 1 Departent of Health Technology and Inforatics, Hong Kong Polytechnic University, Hung Ho, Hong Kong 2 Philips Research China, Shanghai, China 3 Departent of Diagnostic Radiology, University of Hong Kong, Pok Fu La, Hong Kong Correspondence should be addressed to Lawrence W. C. Chan; wing.chi.chan@polyu.edu.hk Received 7 April 2017; Accepted 21 May 2017; Published 6 August 2017 Acadeic Editor: Zhe He Copyright 2017 Lawrence W. C. Chan et al. This is an open access article distributed under the Creative Coons Attribution License, which perits unrestricted use, distribution, and reproduction in any ediu, provided the original work is properly cited. Electronic Health Record (EHR) syste enables clinical decision support. In this study, a set of 112 abdoinal coputed toography iaging exaination reports, consisting of 59 cases of hepatocellular carcinoa (HCC) or liver etastases (so-called HCC group for siplicity) and 53 cases with no abnorality detected (NAD group), were collected fro four hospitals in Hong Kong. We extracted ters related to liver cancer fro the reports and apped the to ontological features using Systeatized Noenclature of Medicine (SNOMED) Clinical Ters (CT). The priary predictor panel was fored by these ontological features. Association levels between every two features in the HCC and NAD groups were quantified using Pearson s correlation coefficient. The HCC group reveals a distinct association pattern that signifies liver cancer and provides clinical decision support for suspected cases, otivating the inclusion of new features to for the augented predictor panel. Logistic regression analysis with stepwise forward procedure was applied to the priary and augented predictor sets, respectively. The obtained odel with the new features attained 84.7% sensitivity and 88.4% overall accuracy in distinguishing HCC fro NAD cases, which were significantly iproved when copared with that without the new features. 1. Introduction Sheer aount of clinical data hosted by the electronic health record (EHR) syste facilitates the exploration of disease signatures and potentiates the relevant clinical decision support functions [1, 2]. As a real-tie, digital patient-centered record, EHR contains a large aount of patient inforation and laboratory and test results. It provides opportunities to enhance patient care, to ebed perforance easures in clinical practice, and to ake inforation available instantly and securely to the authorized users [3]. These voluinous coplex data contain abundant input for precision edicine and big data analytics, which can extract huge knowledge to iprove the quality of healthcare [4]. Integrated exploitation of ultiple heterogeneous sources also serves for ultidisciplinary renovation like bioedical engineering. In this article, extrapolating EHRs huan lexical judgents fro coputational odels of seantics is one of the approaches that can iniize huan intervention and save huan efforts significantly. The rapid developent of EHR provides good opportunity to utilize the data for risk odeling and clinical decisions. Besides the well-structured deographics and laboratory inforation, clinical reports in EHR provide great potential for achine learning and data ining to exploit the detailed clinical inforation to iprove risk odeling and prediction. For exaple, achine learning approaches could

2 2 Journal of Healthcare Engineering New case Deographic, clinical signs, syptos, exaination report, lab results, signals, iages, and so on. Algorith EHR database Key features Fast prediction & risk odeling Siilar cases Diagnostic & therapeutic recoendations Other applications Figure 1: Clinical decision support application of EHR siilarity algorith. be developed based on adission notes and progress notes to iprove prediction of ajor adverse cardiac events (MACE) of acute coronary syndroe (ACS) [5, 6]. Extraction of key inforation fro reports is a foundation step to enable these data ining applications. As a siplifying representation in natural language processing and inforation retrieval, the bag-of-words odel has long been applied in the text clustering tasks, in which docuents are represented by independently treated single ters [7]. Without a reference terinology, a bag of words can be extracted fro a docuent to for an array of unique features whose weights are deterined by the ter frequencies and for the feature vector. However, the length of feature vector increases onotonically with the nuber of docuents in the dataset of interest, jeopardizing the practicality of the bag-of-words odel. Recently, soe researchers focused on the application of ontology for extracting the conceptual features fro docuents. Based on reference ontology, the feature vectors consist of coon fixed eleents, which have already been defined before the feature extraction. Such ontological feature vector odel could iprove the perforance of text retrieval and classification [8, 9]. In soe studies, feature vector odel has been developed for converting the clinical texts and iage patterns of an EHR into an array of nuerical values [10 13]. The support of a edical ontology is required to ap textual inforation, such as iage findings in a diagnostic report, to a feature vector [12, 13]. Systeatized Noenclature of Medicine (SNOMED) Clinical Ters (CT) is an ontological standard of clinical ters, which are organized as concepts and linked with is-a or inverse is-a relationships [14 17]. In such hierarchical structure, concepts at a particular level could be chosen as the feature concepts. Soe studies have copared SNOMED-CT with other standards, such as International Classification of Diseases (ICD) and MEDCIN [18, 19]. As a trigger to order laboratory tests, clinical conditions were extracted fro laboratory guidelines and apped to ICD10 and SNOMED-CT. It was found that ICD10 could cover 43.1% of clinical conditions only, whereas 80.1% of these conditions were apped by SNOMED-CT. For representing trauatic brain injury (TBI) concepts, SNOMED-CT yielded a sensitivity of 90%, outperforing MEDCIN whose sensitivity was 49%. Thus, SNOMED-CT was selected as the reference ontology in this study. The seantic distance between a clinical ter in EHR and a feature concept can be quantified by counting the edges along the path connecting the in the is-a hierarchy [10 12, 20, 21]. Aggregating all the seantic distances to the feature concepts generates an ontological feature vector that characterizes an EHR with its disease context. A study has perfored the evaluation and coparison between inforation content and edge counting approaches proposed by various published works against bencharks [11]. It was found that features built with edge-counting outperfored ost of the inforation content approaches. Therefore, the edge counting is necessary for weighting the features. We hypothesize that the feature association patterns derived fro the EHRs can uniquely distinguish a disease group fro the nondisease group. If such distinguishable association patterns exist, new features could be derived fro the patterns and incorporated into the existing ontological feature vector to strengthen the ontological characterization of EHRs and thus the classification perforance using siilarity algorith, as illustrated in Figure 1. The identified ontological patterns can be used to develop a clinical decision support functions. For the new cases, siilar cases retrieved fro EHR database using the patterns provide clinicians with evidence of the feasible diagnostic and therapeutic options. The siilarity search algorith based on the ontological vector odel has been successfully applied to siilar radiological iage report retrieval and siilar radiotherapy treatent plan retrieval [22 24]. In addition to the clinical evidence, the association between concepts in the patterns can be used to reind a clinician of checking the inclusion of a concept when its associated concept has already been entioned in an EHR.

3 Journal of Healthcare Engineering 3 Level 4 (feature concepts) Liver finding Abdoinal organ finding Fatty liver Level 5 Disease of liver b Disorder of spleen c Level 6 a Liver regeneration Hepatic fibrosis Splenoegaly Level 7 Cirrhosis Figure 2: Edge counting based on level 4 concepts: liver finding, abdoinal organ finding, and fatty liver. (a) Cirrhosis at level 7 is the descendant of liver finding, edge count is 3. (b) Edge count between hepatic fibrosis and liver finding is 2. (c) Splenoegaly is the descendant of abdoinal organ finding but not liver finding. Thus, edge count of splenoegaly with abdoinal organ finding is 2 and that with liver finding is infinity. Fatty liver is a feature concept, and thus, the edge count with itself is 0. Diagra was extracted fro [22]. 2. Methods 2.1. Data Collection. We collected retrospectively 112 iage reports of abdoinal coputed toography exainations fro the radiology departents of four local hospitals in Hong Kong. HCC or liver etastases were found in 59 cases (called HCC group for siplicity) and the other 53 cases had no abnorality detected (NAD group). These 112 cases were randoly selected fro the pool of iage reports where HCC or liver etastases were reported in the diagnoses of HCC cases and not reported in the diagnoses of NAD cases. Before the data collection, third party clinical personnel have reoved the patient nae, identity card nuber, telephone nuber, and address fro the reports and assigned a randoly generated unique ID to each case. We have obtained Huan Subject Ethics Approval fro the Hong Kong Polytechnic University (HSEARS ) Ontological Feature Extraction. The HCC-related clinical ters were extracted anually fro the iage reports according to SNOMED-CT curated in the Unified Medical Language Syste (UMLS; license code: NLM ). During the extraction process, the whole iage reports were read and interpreted. The negation and uncertainty of a disease, disorder, or iage finding was regarded as not detected and the corresponding ter was not considered in the ontological feature apping. Modifiers for clinical ters were not found in the iage reports. To facilitate the future studies on a bigger dataset, the extraction can be autoatic if the ters in iage reports have been already tagged by SNOMED-CT or extracted autoatically by text-ining ethods. UMLS organizes clinical ters in concepts, and SNOMED-CT defines the relationship between concepts using the is-a hierarchical tree. The extracted ters were projected to the feature concepts at a particular level to ensure consistent coparison between reports. In our previous study, a set of EHRs were collected fro 47 subjects of type II diabetic patients in Hong Kong [21]. Levels 1 4 of the SNOMED-CT hierarchy were considered as the individual candidate sets of the feature concepts. For each level, ontological feature vectors were generated using the alignent with SNOMED-CT hierarchy and the siilarity score between every possible pair of EHRs was calculated. Using SNOMOD-CT level 4, the accuracy was highest for ranking the agreeent of carotid plaque identification in EHR pairs. It is iportant to note that level 4 has already had 6964 feature ters, providing sufficient granularity for characterizing EHRs. The use of level 5 is indeed infeasible due to the treendously large nuber of features. Due to the optial classification granularity, level 4 concepts were considered as feature concepts in this work. Edge-counting approach is illustrated in Figure 2. For each report, the ontological features, a 1, a 2,, a, were generated using edge-counting approach based on the following forula: a i = p i 1 + in j=1,,n s ij, 1 where p i is the conditional probability of the ith feature concept given the occurrence of liver cancer and s ij represents the edge count between the ith feature concept and the jth clinical ter extracted fro a report. A saller edge count eans that the feature concept is conceptually closer to the clinical ter. Therefore, the iniu of the edge counts should be taken to deterine the degree of activation of a feature concept. PubMed docuent clustering has been successfully deonstrated using the edgecounting ethod [25]. With the value between 0 and 1, a i indicates the relevance between the ith feature concept and a clinical ter in a report. Such relevance can be odulated by the conditional probability, p i, which is estiated by the specific terweighting approach [22]. Indeed, a siilarity easure derived fro direction cosine represents the su of the product of ontological features. Each product of corresponding features eliinates the square root, and the value p i becoes

4 4 Journal of Healthcare Engineering the weight associated with the product of the degree of feature concept activation between two EHRs. It is obvious that the values of a i follow a nonnoral distribution in the HCC and NAD populations, which violates the assuption of statistical analysis using Pearson correlation coefficient. Rank-based inverse noralization is a popular approach that converts the feature values to those norally distributed across individuals [26]. Those features with zero values do not cause any effect on the characterization of iage reports and the association patterns between features. Thus, those zero-valued features were excluded in inverse noralization process and reain unchanged. For each feature concept, the nonzero values of a i were ranked by R i 1, N N 0 aong reports of the group where N and N 0 are, respectively, the total nuber of reports and the nuber of zero-valued features in the group. The activation value of the ith feature concept is given by z i = Φ 1 R i ξ N N 0 2ξ +1, 2 where Φ 1 represents the standard noral quantile function and ξ denotes a constant, whose value is given by zero as suggested by van der Waerden [27]. The activation values of a feature concept across a group for the following vector: u i = z i 1, z i 2,, z i N T 3 F d F n D = ax F d C F n C C C = prob C d C C = prob C n C D α = γα 4 1, where α is the significance level, that is, 0.05, γ 0 05 =31, and k =30 in this study. The critical value of D is , which has been proved by exhaustive coputer siulations [28] New Features Derived fro Association Patterns. It is interesting to explore soe new features, which signify the iage reports of HCC cases, based on the above-entioned ontological association patterns. The first new feature, z 1 k, is the square of the su of activation values characterizing the iage report of the kth case in a group. z 1 k = z i k The expected value of this new feature can be estiated by its average over the group Note that nonzero z i k follows noral distribution, N 0, 1, after inverse noralization Ontological Association Patterns. The association level between two feature concepts was denoted by C d i, j for the HCC group and C n i, j for the NAD group, as given by the following forulas: C d i, j = r u di, u dj = 1 N d z N di k z dj k d C n i, j = r u ni, u nj = 1 N n z N ni k z nj k, n 4 N 1 N z 1 k = 1 N N = 1 N N z i k i=,j= = i j,,j=1 i=,j= i j,,j=1 2 i=,j= z i k 2 +2 z i k z j k i j,,j=1 N 1 N z i k z j k Ci, j 7 where u di and u dj represent the vectors weighting the ith and jth feature concepts across the HCC group; u ni and u nj represent the vector weighting the ith and jth feature concepts across the NAD group; and r u i, u j is Pearson correlation coefficient between two arrays. Two sets of correlation coefficients, C d and C n, in the HCC and NAD groups fored two cuulative distributions, F d and F n, which were copared using two-saple Kologorov-Sirnov (KS). To test the significant difference, the axiu deviation between two cuulative distributions, D value, was copared with its critical value, D α, which is derived based on our developed ethod [28] and given by following equations. A correlation threshold, at which F d and F n were extreely deviated, can be identified and used to characterize the perturbed ontological association pattern. It is clearly shown that the expected value of this new feature fors the lower bound of the su of association levels over all possible pairs of features in the group. The second new feature, z 2 k, is the square of the su of the absolute values of activation values characterizing the iage report of the kth case in a group. z 2 k = z i k The expected value of the second new feature, again, can be estiated by its average over the group. 2 8

5 Journal of Healthcare Engineering 5 N 1 N z 2 k = 1 N N = 1 N N z i k i=,j= = i j,,j=1 i=,j= i j,,j=1 2 i=,j= z i k 2 +2 z i k z j k i j,,j=1 N 1 N z i k z j k Ci, j The above forula clearly shows that the expected value of the second new feature defines the upper bound of the su of association levels over all possible pairs of features in the group. When the KS test indicates that the ontological association patterns of two groups are significantly different, we expect that the su of association levels of a group is distinguishable fro that of the other group. Therefore, the new features could signify the difference between two groups Logistic Regression. The statistical analysis was perfored by SPSS (IBM SPSS Statistics 22; Aronk, NY). Binary logistic regression selects and estiates the optial subset of independent variables for predicting categorical outcoe Y coded by 1 or 0, which represents HCC and NAD in this work. Stepwise forward procedure was used to obtain the logistic regression odel where the potential predictors were prioritized and entered into the odel one by one until the predictive power was optiized. The procedure results in the following odel with M predictors, logit = ln P 1 P = β 0 + β 1 X 1 + β 2 X β M X M, 10 where logit is the estiated log odds of Y =1, P is the estiated probability of Y =1, X i is the ith predictor entered into the odels, and i is the coefficient associated with the ith predictor for i =1,, k. The statistical significance of the association between the outcoe and each predictor is indicated by p <005. For a well-balanced saple, we assue 50% of the cases will be classified as Y =1and the cut-off of logit is set at 0. Saple is ibalanced when the nuber of cases with an outcoe category is about 2 5 ties that with the other category. For an ibalanced saple, the constant 0 is corrected by deducting the log odds of Y =1 observed in the saple. Onibus test of odel coefficients indicate the overall perforance of an identified odel. Two sets of candidate predictors, priary set and augented set, are considered for identifying the logistic regression odels. The priary set consists of the activation values of feature concepts: z 1 k, z 2 k,, z k. The augented set is coposed of the activation values of feature concepts and three new features derived fro the association patterns: z 1 k, z 2 k,, z k, z 1 k, z 2 k Experiental Settings. Figure 3 illustrates the flow chart of the experiental steps perfored in the study Perforance Evaluation. Sensitivity, specificity, and overall accuracy were used to evaluate the perforance of two logistic regression odels based on the priary and augented predictor sets. To exaine the agreeent between priary predictor odel (PPM) and augented predictor odel (APM), 2 2 contingency tables for HCC, NAD, and all cases are constructed. The McNear test is used to copare sensitivities, specificities, and overall accuracy of two odels. The difference in perforance is considered significant if the P value is less than Results Edge counting Ter weighting Inverse noralization Priary predictor set Logistic regression HCC versus NAD Priary predictor odel Association patterns HCC versus NAD Augented predictor set Logistic regression HCC versus NAD Augented predictor odel Figure 3: Flow chart of the perfored experiental steps Extracted Features. Fro 59 and 53 iage reports of respective HCC and NAD groups, 38 clinical ters were extracted and apped to 38 unique concepts in UMLS. Based on the approach illustrated in Figure 2, these ters were then projected to 30 feature concepts at level 4 of SNOMED-CT is-a hierarchy (Table 1). After counting the edges and estiating the conditional probabilities of these concepts, their weightings were calculated and fored and atrices for HCC and NAD groups Ontological Association Patterns. The association level between every two feature concepts was calculated. We generated 435 association levels for each of HCC and NAD groups. Figure 4 shows the cuulative distributions of association levels for the two groups and their difference. The axiu deviation, D =0333, was found at C =003 and greater than its critical value. Therefore, the two ontological association patterns are significantly different Priary Predictor Model. The stepwise forward procedure stops at step 2 where the prediction accuracy is optial, yielding the following regression:

6 6 Journal of Healthcare Engineering Table 1: Feature concepts and feature vectors of representative NAD and HCC cases. Class NAD HCC Abdoinal organ finding Blood vessel finding 0 0 Disorder of body cavity Disorder of body syste Disorder of cardiovascular syste Disorder of digestive syste Disorder of soft tissue Disorder of trunk Finding of trunk structure Liver finding Radiologic finding Cyst of abdoen 0 0 Mass of body region Mass of digestive structure Neoplastic disease 0 0 Growth alteration 0 0 Iaging result abnoral 0 0 Mechanical abnorality Finding of biliary tract Heorrhage into peritoneal cavity 0 0 Disorder of connective tissue 0 0 Degenerative abnorality 0 0 Trauatic and/or nontrauatic injury of anatoical site 0 0 Abnoral radiologic density, diffuse Iaging result abnoral Abnoral radiologic density, irregular 0 0 Abnoral radiologic density, nodular 0 0 Abnoral radiologic density, sall area 0 0 Multiple lesions Finding of nuber of lesions logit = 1 28 z z , 11 where z 11 represents the activation value of radiologic finding and z 27, abnoral radiologic density, nodular. The predictor z 11, radiologic finding, is significantly associated with the log-odds of HCC (p =0016). The constant has been adjusted to copensate the ibalanced NAD and HCC cases. Onibus test shows that the variance of log-odds explained by the odel is significantly greater than the unexplained variance (χ 2 = , df = 2, p =0002). For logit 0, the case is NAD ore likely than HCC. For logit > 0, the case is HCC ore likely than NAD. Classifier based on this odel is illustrated in Figure 5(a). The y-axis represents the linear cobination of z 11 and z 27 in the above equation. The horizontal dotted line indicates the threshold level in the equation, 0.114, above which a lesion is classified as HCC and otherwise, NAD. Cuulative fraction of feature pairs 1 NAD HCC Difference Ontological association level (C) Figure 4: Two distinct ontological association patterns. Cuulative distributions of ontological association levels across NAD and HCC groups are indicated by dash-dotted and dash lines, respectively. Solid line represents the difference between these two cuulative distributions Augented Predictor Model. The stepwise forward procedure stops at step 5 where the prediction accuracy is optial, yielding the following regression: logit = 0 586z z z z z , where z 1 and z 2 are the squares of su of the ontological features and their absolute values, which were incorporated into the odel in the first two steps; z 11, z 13, and z 25 represent the activation values of radiologic finding, ass of body region, and iaging result abnoral, respectively, which were included in steps 3 5. The augented predictors, z 1 and z 2, and the priary predictors, z 11, radiologic finding, z 25, and iaging result abnoral are significantly associated with the log-odds of HCC (p =0014, 0.006, 0.003, and 0.04). The constant has been also adjusted to copensate the ibalanced NAD and HCC cases. Onibus test shows that the variance of log-odds explained by the odel is significantly greater than the unexplained variance (χ 2 = , df = 5, p <0001). For logit 0, the case is NAD ore likely than HCC. For logit > 0, the case is HCC ore likely than NAD. Classifier based on this odel is illustrated in Figure 5(b). In step 5, the linear cobination of the augented predictors, z 1 and z 2, fors the y-axis and that of the priary predictors, z 11, z 13, and z 25, the x-axis. The classifier is represented by the dotted line Perforance Coparison of Models. Using the PPM, 98.1%, that is, 52 out of 53 NAD cases, and 57.6%, that is, 34 out of 59 HCC cases, are correctly classified. The overall accuracy is 76.8%. Using the APM, the correctly classified HCC cases increase significantly to 84.7% (p <00001), which consist of 50 out of 59 HCC cases. Although the correctly classified NAD cases are reduced slightly to 92.5% (p =0250), the APM raises the overall accuracy to 12

7 Journal of Healthcare Engineering z z Priary predictor odel z 11 : radiologic finding z 27 : abnoral radiologic density, nodular No iproveent in accuracy by introducing any new predictor Step 2 Step NAD HCC (a) Augented predictor odel z z z 1 : squared su of ontological features z 2 : squared su of absolute values of ontological features NAD Step 2 Step 5 HCC z z z 11 : radiologic finding z 13 : ass of body region z 25 : iaging result abnoral 3.72 z z z (b) Figure 5: Logistic regression classification results using priary and augented predictor odels. (a) Priary predictor odel (PPM): the identification procedure incorporated z 11, radiologic finding, z 27, and abnoral radiologic density, nodular in the first two steps. The process stopped at step 2 as new predictor cannot ake any iproveent in classification. (b) Augented predictor odel (APM): two new features, z 1 and z 2, were included in the odel in the first two steps of the procedure. The predictors, z 1 and z 2, represent the squared sus of ontological features and their absolute values, respectively. The identification proceeds to step 5 that extends the feature space into three predictor diensions, z 11, z 13, and z 25, representing radiologic finding, ass of body region, and iaging result abnoral. 88.4% significantly (p <00001). The coparison of perforances is suarized in Table Discussion This study illustrated an approach for characterizing textual iage reports by nuerical values that weight the alignent of report contents with the ontological standard. Such approach has been deonstrated in our previous study where all the level 4 feature concepts of SNOMED-CT were considered to characterize the sae set of iage reports used in this study [22]. Using the specific ter weighting, the highest overall accuracy, 74.3%, was attained for apping report pairs based on the siilarity easure of odified direction cosine. In this study, the features were further converted to standardized values, following N 0, 1, by inverse noralization [26]. Such conversion could help reduce the noise or outlier that was induced to the features through the edge-counting approach. The converted features were considered as priary predictors. Binary logistic regression odel, identified using the priary predictors as the candidates in stepwise forward procedure, was used to classify the reports. The overall accuracy was increased to 76.8%. It was shown that the interfeature association levels in HCC and NAD groups exhibited significantly different distributions where the feature concepts have particularly strong association in HCC [29]. This observation led to the derivation of two new features, which are squared sus of the existing features and their absolute values. We proved that the expected values of these two new features, which are estiated by their averages, represent the lower and upper liits of the su of association levels over the group. The new features were cobined with the existing features to provide the augented predictor set for the stepwise forward procedure. It was found that the overall accuracy was significantly increased to 88.4% (p <00001). The sensitivity, an iportant diagnostic perforance indicator, was also significantly increased fro 57.6% to 84.7% (p <00001).

8 8 Journal of Healthcare Engineering Table 2: Perforances of priary predictor odel (PPM) and augented predictor odel (APM) in classifying NAD and HCC cases. HCC and NAD predictions are indicated by + and, respectively. Model Classification rate NAD (53 cases) HCC (59 cases) Overall accuracy (112 cases) PPM 98.1% (52 cases) 57.6% (34 cases) 76.8% (86 cases) APM 92.5% (49 cases) 84.7% (50 cases) 88.4% (99 cases) Contingency table PPM APM PPM APM PPM APM McNear test p =0250 p <00001 p <00001 Besides the first two new features, we identified the feature concepts: radiologic finding, ass of body region, and iaging result abnoral. For new suspected cases, this panel of predictors representing a disease signature can be used to assist the clinical decision when associations of those pairs are observed. In future work, the discovered signature should be validated with independent data before its clinical applications. The detailed underlying eanings of the signature in patient anageent should be further explored using big data analytics. An alternative application of the identified association patterns is the detection of inaccurate edical coding. When a disease is diagnosed, the coactivated feature concepts can be obtained and checked against the pairs in the diseasespecific patterns. Potential inaccurate coding can be detected and the clinicians will be alerted. On a public health level, systeatic failure in appropriate edical coding ay result in under- or overadjustent to case-ix easureents when assessing quality of care [30]. In soe healthcare odels, this will also affect billing, reiburseent, and insurance clais [31]. Soe observed iage patterns entioned in the iage reports cannot be apped to concepts in SNOMED-CT. For exaple, intravenous contrast injection induces changes of pixel optical density in different phases of CT scan. Contrast enhanceent in particular phases is critically iportant for HCC diagnosis. However, SNOMED-CT has not defined the concepts, which could represent closely contrast enhanceent, arterial enhanceent, and hyperdensity in arterial phase. This is a liitation of this study that hindered the precision of the proposed predictor odel. 5. Conclusions This study deonstrated the extraction of ontological features fro iage report contents based on the ontological standard. Cobining new features, derived fro the differential association patterns, with the ontological features fors a panel of augented predictors that signifies the HCC iage reports. Conflicts of Interest The authors declare that they have no conflicts of interest. Acknowledgents This research was supported financially by Philips Research China. References [1] B. J. Peter, J. J. Lars, and B. Søren, Mining electronic health records: towards better research applications and clinical care, Nature Reviews Genetics, vol. 13, no. 6, pp , [2] W. Ceuster and B. Sith, Strategies for referent tracking in electronic health records, Journal of Bioedical Inforatics, vol. 39, pp , [3] P. Wu, C. Cheng, C. D. Kaddi, J. Venugopalan, R. Hoffan, and M. D. Wang, Oic and electronic health record big data analytics for precision edicine, IEEE Transactions on Bioedical Engineering, vol. 64, no. 2, pp , [4] M. R. Cowie, J. I. Bloster, L. H. Curtis et al., Electronic health records to facilitate clinical research, Clinical Research in Cardiology, vol. 106, no. 1, pp. 1 9, [5] D. Hu, Z. Huang, T. M. Chan, W. Dong, X. Lu, and H. Duan, Utilizing Chinese adission records for MACE prediction of acute coronary syndroe, International Journal of Environental Research and Public Health, vol. 13, no. 9, [6] Z. Huang, T. M. Chan, and W. Dong, MACE prediction of acute coronary syndroe via boosted resapling classification using electronic edical records, Journal of Bioedical Inforatics, vol. 66, pp , [7] G. Salton and M. J. McGill, Introduction to Modern Inforation Retrieval, McGraw-Hill Inc., [8] A. Hotho, S. Staab, and G. Stue, Wordnet iproves text docuent clustering, in Proceedings of the Seantic Web Workshop at SIGIR-2003, 26th Annual International ACM SIGIR Conference, [9] S. Bloehdorn and A. Hotho, Text classification by boosting weak learners based on ters and concepts, in Proceedings of the 4th IEEE International Conference on Data Mining, pp , Brighton, UK, Noveber [10] L. W. Chan, I. F. Benzie, Y. Liu, and C. R. Shyu, Is the interpatient coincidence of a subclinical disorder related to EHR siilarity?, in 2011 IEEE 13th International Conference on e-health Networking, Applications and Services, pp , Colubia, MO, June [11] D. Sánchez, M. Batet, D. Isern, and A. Valls, Ontologybased seantic siilarity: a new feature-based approach,

9 Journal of Healthcare Engineering 9 Expert Systes with Applications, vol. 39, no. 9, pp , [12] M. Batet, D. Sánchez, and V. Aida, An ontology-based easure to copute seantic siilarity in bioedicine, Journal of Bioedical Inforatics, vol. 44, pp , [13] R. L. Richesson, J. E. Andrew, and J. P. Krischer, Use of SNOMD CT to represent clinical research data: a seantic characterization of data ites on case report fors in vasculitis research, Journal of the Aerican Medical Inforatics Association, vol. 13, no. 5, pp , [14] G. B. Melton, S. Parsons, F. P. Morrison, A. S. Rothschild, M. Markatou, and G. Hripcsak, Inter-patient distance etrics using SNOMED CT defining relationships, Journal of Bioedical Inforatics, vol. 39, no. 6, pp , [15] T. Pedersen, S. V. Pakhoov, S. Patwardhan, and C. G. Chute, Measures of seantic siilarity and relatedness in the bioedical doain, Journal of Bioedical Inforatics, vol. 40, no. 3, pp , [16] H. Wasseran and J. Wang, An applied evaluation of SNOMED CT as a clinical vocabulary for the coputerized diagnosis and proble list, in AMIA Syposiu, pp , Aerican Medical Inforatics Association, [17] M. I. Lieberan, T. N. Ricciardi, and F. E. Masarie, The use of SNOMED CT siplifies querying of a clinical data warehouse, in AMIA Syposiu, p. 910, Aerican Medical Inforatics Association, [18] M. Yasini, V. Ebrahiinia, C. Duclos, A. Venot, and J. B. Lay, Coparing the use of SNOMED CT and ICD10 for coding clinical conditions to ipleent laboratory guidelines, in Volue 186: Data and Knowledge for Medical Decision Support, B. Blobel, Ed., pp , IOS Press, [19] D. Montella, S. H. Brown, P. L. Elkin et al., Coparison of SNOMED CT versus Medcin terinology concept coverage for ild trauatic brain injury, in AMIA Syposiu, pp , Aerican Medical Inforatics Association, [20] P. W. Lord, R. D. Stevens, A. Brass, and C. A. Goble, Investigating seantic siilarity easures across the gene ontology: the relationship between sequence and annotation, Bioinforatics, vol. 19, no. 10, pp , [21] L. W. Chan, Y. Liu, C. R. Shyu, and I. F. Benzie, A SNOMED supported ontological vector odel for subclinical disorder detection using EHR siilarity, Engineering Applications of Artificial Intelligence, vol. 24, pp , [22] L. W. Chan, Y. Liu, T. Chan et al., PubMed-supported clinical ter weighting approach for iproving inter-patient siilarity easure in diagnosis prediction, BMC Medical Inforatics and Decision Making, vol. 15, p. 43, [23] E. S. Liu, V. W. Wu, B. Harris, M. Foote, M. Lehan, and C. LWC, Vector-odel-supported optiization in voluetric-odulated arc stereotactic radiotherapy planning for brain etastasis, Medical Dosietry, vol. 42, no. 2, pp , [24] E. S. Liu, V. W. Wu, B. Harris, M. Lehan, D. Pryor, and L. W. C. Chan, Vector-odel-supported approach in prostate plan optiization, Medical Dosietry, vol. 42, no. 2, pp , [25] X. Zhang, L. Jing, X. Hu, M. Ng, and X. Zhou, A coparative study of ontology based ter siilarity easures on PubMed docuent clustering, Lecture Notes in Coputer Science, vol. 4443, pp , [26] T. M. Beasley, S. Erickson, and D. B. Allison, Rank-based inverse noral transforations are increasingly used, but are they erited? Behavior Genetics, vol. 39, no. 5, pp , [27] B. L. van der Waerden, Order tests for the two-saple proble and their power, Indagationes Matheaticae, vol. 14, pp , [28] L. W. Chan, X. Lin, G. Yung et al., Novel structural coexpression analysis linking the NPM1-associated ribosoal biogenesis network to chronic yelogenous leukeia, Scientific Reports, vol. 5, article 10973, [29] L. W. Chan, S. C. Wong, and K. W. Chiu, Ontological features of electronic health records reveal distinct association patterns in liver cancer, in 2016 IEEE International Conference on Bioinforatics and Bioedicine (BIBM), pp , Shenzhen, Deceber [30] S. F. Jencks, Case-ix easureent and assessing quality of hospital care, Health Care Financing Review, vol. 1987, Suppleent, pp , [31] M. Slight, S. E. Belley, W. Shrader, and M. Endicott, ICD-10: targeting reporting and reiburseent, Health Manageent Technology, vol. 37, no. 3, pp. 8 11, 2016.

10 Volue Volue Volue Volue Volue Volue 2014 International Journal of Rotating Machinery Journal of Volue 201 The Scientific World Journal Volue 2014 Journal of Sensors Volue 2014 International Journal of Distributed Sensor Networks Journal of Control Science and Engineering Advances in Civil Engineering Subit your anuscripts at Journal of Robotics Journal of Electrical and Coputer Engineering Volue 2014 Advances in OptoElectronics Volue 2014 VLSI Design International Journal of Navigation and Observation Volue 2014 Modelling & Siulation in Engineering Volue 2014 International Journal of International Journal of Antennas and Cheical Engineering Propagation Active and Passive Electronic Coponents Shock and Vibration Advances in Acoustics and Vibration Volue Volue Volue Volue Volue 2014