A Fuzzy-Based Clinical Decision Support System for Coeliac Disease

Coeliac disease (CD) is a permanent inflammatory disease of the small intestine characterized by the destruction of the mucous membrane of this intestinal tract. Coeliac disease represents the most frequent food intolerance and affects about 1% of the population, but it is severely underdiagnosed. Currently available guidelines require CD-specific serology and atrophic histology in duodenal biopsy samples to diagnose CD in adults. In paediatric CD, but recently in adults also, non-invasive diagnostic strategies have become increasingly popular. In order to increase the rates of correct diagnosis of the disease without the use of biopsy, researchers have recently been using approaches based on artificial intelligence techniques. In this work, we present a Clinical Decision Support System (CDSS)system for supporting CD diagnosis, developed in the context of the Italy-Malta cross-border project ITAMA. The implemented CDSS has been based on a neural-network-based fuzzy classifier. The system was developed and tested using a Virtual Database and a Real Database acquired during the ITAMA project. Analysis on 10,000 virtual patients shows that the system achieved an accuracy of 99% and a sensitivity of 99%. On 19,415 real patients, of which 109 with a confirmed diagnosis of coeliac disease, the system achieved 99.6% accuracy, 85.7% sensitivity, 99.6% specificity and 96% precision. Such results show that the developed system can be used effectively to support the diagnosis of the CD by reducing the appeal to invasive techniques such as biopsy.


I. INTRODUCTION
Coeliac disease (CD) is a rapidly expanding disorder both 19 in terms of prevalence in the world and in terms of a more 20 significant number of diagnosed patients; it is an autoim- 21 mune disease that can occur at all stages of life. Advances 22 in understanding the pathogenetic and genetic factors that 23 The associate editor coordinating the review of this manuscript and approving it for publication was Yizhang Jiang .
influence risk have led to the development and refinement 24 of diagnostic tools. It is a chronic disease of the small 25 intestine characterized by an abnormal immune response; 26 the latter is due to exposure to gluten present in the diet 27 in genetically predisposed subjects. The ''environmental'' 28 factor triggering coeliac disease is represented by gluten, 29 a protein complex contained in some cereals (wheat, barley, 30 rye) [1]. Coeliac disease is an autoimmune disorder induced 31 by dietary gluten in genetically predisposed subjects. CD has 32 VOLUME 10, 2022 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ a prevalence of ∼ 1% in many populations around the world, of anti tTG_IgA kits. 78 In order to satisfy the heterogeneity of data and their com- (Artificial Intelligence), interoperability (multidisciplinarity) 89 and data sources (Cloud, open data,. . . ). Thus, decision sup-90 port systems become a relevant part of the tools that use 91 artificial intelligence (AI). They have the task of solving open 92 questions to deepen the understanding of the correlations 93 between the data representing the events. The DSS, supported 94 by multidisciplinary approaches, revisits how data are treated 95 and analysed and generates virtually and globally cognitive 96 pathways that highlight unconventional solutions and consid-97 erations. This approach improves how data are analysed and 98 understood, their knowledge and correlation. 99 knowledge-based, data-driven, or lacking a priori knowl-100 edge. The strategy of the former is based on rules that are 101 not necessarily deterministic; it recovers data from informa-102 tion systems (i.e.: databases) or in real-time from Biometric 103 systems and evaluates the rules involved. Finally, it produces 104 an output event (Alarm, screening, diagnostic pathway,. . . ). 105 Non-knowledge-based DSSs are data-driven, and the output 106 events result from modelling applications on machine learn-107 ing with no specific medical knowledge needs to take into 108 account to set up the model. Such a model without knowl-109 edge, adopted for the creation of the DSS, is currently being 110 studied in the scientific community; they are incredibly com-111 plex to implement, and they leave no room for understanding 112 the results, whether they are correct or incorrect, even when 113 they have a high degree of sensitivity and specificity. 114 The inherent imprecision of medical data, as well as the 115 fact that a patient enters the diagnostic pathways from differ-116 ent medical sources (as an outpatient, an inpatient, referred by 117 a physician, after blood tests and other unrelated diagnostic 118 tools have been administered) makes standard classification 119 methods less easy to adapt to the CDSS backend. While 120 well-known classifiers such as SVM and NN produce precise 121 results on binary classification problems, the diagnosis of 122 a coeliac patient requires a number of steps, and a CDSS 123 should offer prioritisation advice on each of these steps, 124 regardless the completion of the whole diagnostic pathway. 125 Fuzzy classifiers allow for taking into account this inherent 126 dynamicity and imprecision. 127 In this work, we present a fuzzy-based Clinical Diagnostic 128 Support System developed within the ITAMA project (hence-129 forth ITAMACDSS). ITAMA (ICT Tools for the diagnosis 130 of Autoimmune diseases in the Mediterranean Area, [8]) 131 is a cross-border project between Italy and Malta funded 132 by the European Regional Development Fund within the 133 INTERREG V-A Italia -Malta Cooperation Programme, 134 in which the common territorial challenge is to improve the 135 quality of life and well-being of the population affected by 136 autoimmune diseases, containing the costs of health systems 137 through a strategic commissioning demand towards the world 138 of research. In Sicily and Malta, autoimmune diseases present 139 a high incidence, probably due to the high consumption of 140 starchy foods. In ITAMA, a mass screening was carried out on 141 more than 20,000 Maltese children. The screening was based 142 on a Medical History Questionnaire (MHQ) and a Point-143 of-Care Test (PoCT). Children tested positive based on the 144 result of the MHQ, of the PoCT, or both, were invited for 145 further investigation, and in particular for the anti-Actin IgA 146 to verify the possibility of avoiding biopsy in a large number 147 of patients.

148
The implemented CDSS is based on a fuzzy classifier using 149 neural networks. The system was developed and tested using   if applied to training with large data sets; but large data sets 193 are not always available in the medical domain [9].

197
A dataset with 150 medical records was used for training 198 and testing, and the system was trained using a resilient 199 backpropagation algorithm with a ten-fold cross-validation 200 scheme to assess the generalization of the system thus, the 201 results show an accuracy of the system is 81%.

202
The introduction of DSSs for diagnosing CD could 203 improve diagnostic work-up, allowing cost, time and labour 204 savings and improving the procedure's safety, avoiding 205 biopsy sampling and prolonged sedation associated with the 206 multiple biopsy protocol. In particular, DSS based on Fuzzy 207 Logic are enjoying growing research interest in solving clas-208 sification problems in a wide range of application fields [11], 209 especially in medicine, where the possibility of presenting 210 classification results together with a measurement of the 211 association is very tempting [12].

212
The interest of the scientific community in the develop-213 ment of DSS systems, also thanks to new performing machine 214 learning techniques, is certainly growing [13], [14]. However, 215 the problem of developing CD diagnosis support systems 216 is still poorly explored, perhaps due to the difficulty of the 217 problem but undoubtedly also due to the lack of public 218 databases. For example, in a recent review work [15], after 219 proper research, the authors have identified only 41 publi-220 cations consisting of original work describing techniques for 221 computer-aided CD diagnosis. 222 Gadermayr et al. [16] summarize recent trends in 223 computer-aided coeliac disease diagnosis based on upper 224 endoscopy and proposed pipelines for fully-automated 225 patient-wise diagnosis and for integrating expert knowledge 226 into the automated decision process.

227
In [17], the authors presented a feature descriptor for 228 the classification of video capsule endoscopy images. 229 In addition, they introduced a system for small intestine 230 motility characterization based on deep CNN for individ-231 ual motility events. Experimental results showed a mean 232 classification accuracy of 96% for six intestinal motility 233 events.

234
In the last few years, deep learning methods have also been 235 used to classify endoscopic images. In this context, the best 236 known convolutional neural networks, i.e. AlexNet [18], [19], 237 GoogLeNet [20], VGGf net [21], and VGG16 net [22], have 238 been used for this purpose. 239 Wang et al. in [23] propose a deep-learning-based method-240 ology to recalibrate the module to identify images with 241 regions significant for celiac disease from healthy ones. The 242 developed module determines the most salient feature on the 243 features' map and is hooked to a Support Vector Machine 244 and a k-nearest neighbour module to perform a linear dis-245 criminant analysis. Their method reports 95.94%, 97.20% 246 and 95.63% for accuracy, sensitivity and specificity with the 247 10-time 10-fold cross-validation strategy. 248 Amirkhani et al. [24] developed a method based on a fuzzy 249 cognitive map (FCM) and a possibilistic fuzzy C-means clus-250 tering algorithm (PFCM) for the categorization of CD. The 251 research goal was to develop an expert system for classifying 252 patients with CD into three grades A, B1, and B2, which is the 253 latest grading method available. Three experts have extracted 254 seven key defining features of CDs that were considered FCM 255 VOLUME 10, 2022 concepts. For the three analysis classes, the authors obtained 256 88%, 90% and 91% accuracy, respectively.

257
The authors of [25] proposed a fuzzy logic-based method 258 to predict coeliac disease by entering sharp values of various 259 symptoms. The analysis was conducted on 700 individuals. 260 The system, which was based on the Mamdani model, shows     The Point-of-Care Test (PoCT) has negative or inconclu-314 sive outcome (highly unlikely positive PoCT), maintaining 315 the distribution of inconclusive (1:600), and considering a 316 number of defective tests equal to 1:1200, which is consistent 317 with literature. In the case of negative PoCT and negative 318 Questionnaire, the Blood Test has missing values. Otherwise 319 (positive PoCT or positive MHQ), the logic detailed below 320 is followed. First, a value for the total IgA is generated, 321 following the PoCT result: if PoCT is negative, IgA generated 322 value is higher than the threshold with mean 7 and variance 323 2 stdev; if PoCT is inconclusive, IgA has lower random values 324 with random distribution between 0 and 0.25. In the case 325 of a deficit of the total IgA, a value is generated for the 326 tTG_IgG with mean 2 and variance 2 stdev and the value of 327 the tTG_IgA is missing. In the other cases (i.e. if the total 328 IgAs are sufficient) a value for the tTG_IgA is generated from 329 a Gaussian distribution with mean 4.5 and variance 2 stdev, 330 and the value for the tTG_IgG will be missing.

331
If the blood test is positive, the Biopsy will obviously have 332 a negative result (class 1 or 2 -the patient being generated is 333 CD-negative), otherwise it has a missing value.  The PoCT is positive (599:600) or inconclusive (1:600). 341 Blood Tests follow a logic similar to that for negative cases 342 but will always be positive: first a value for the total IgA 343 is generated. In the case of inconclusive PoCT, with mean 344 0.125 and variance 1 stdev, otherwise mean 8 and variance 345 2 stdev. In the case of inconclusive PoCT, a value is generated 346 for the tTG_IgG with mean 14 and variance 2 stdev and 347 the value of the tTG_IgA will remain missing. Values in 348 the negative range are discarded. In the case of a positive 349 PoCT, a value for the tTG_IgA is generated from a Gaussian 350 distribution with mean 24 and long tail on the right, and the 351 value for the tTG_IgG will remain missing. Values in the 352 negative range are discarded 353 The Biopsy has assigned a random positive evaluation with 354 uniform 1:3 distribution among classes 3a, 3b, 3c. it is ready to be extended for further possible functions to be 363 supported. Other general guidelines and best practices were 364 followed for its design, such as data isolation and interoper-365 ability.

366
As far as data isolation is concerned, the DB schema is  As far as interoperability is concerned, we used some well-  The database also keeps track of all the collected data, 411 also those not directly usable for the project's goals but still 412 useful for side statistics (e.g.: defective PoCTs or incomplete 413 personal information).

414
The amount of stored data is 95.8 GiB, of which 31.4 MiB 415 corresponds to internal data, and the remaining corresponds 416 to indexed images of PoCTs and mucosal deposits. The fol-417 lowing figures 1-3 show the distribution of participants by 418 age, gender, and ethnicity.

419
The following tables and figures show the distribution by 420 age, gender, and ethnicity of patients who tested positive at 421 the PoCT (Table 1, Fig. 4), patients who declared five or 422 more symptoms at the MHQ (Table 2, Fig. 5), and patients 423 diagnosed as celiac (Table 3, Fig. 6).     methodologies commonly used in Artificial Intelligence. The 433 purpose of ITAMACDSS is to allow decision-makers in the 434 coeliac disease diagnostic process to evaluate better the status 435 of a subject that has entered the diagnostic pathway by priori-436 tizing the subjects for which a positive diagnosis is more plau-437 sible, at the same time reducing diagnosis costs and optimis-438 ing the use of costly and uncomfortable medical procedures. 439 ITAMACDSS conforms to standard practices and general 440 philosophy in the discipline: a CDSS supports and supple-441 ments physicians, it does not replace them. Furthermore, 442   The ITAMACDSS team has decided to base the classifier 458 portion of the system, which is described in the present 459 article, on a fuzzy paradigm. Fuzziness in input and out-460 put provides a more natural expression, which lends itself 461 better to introspection and conforms to the anthropocentric 462 principles that medicine should adhere to. A fuzzy-based 463 classifier can also deal with multiple explicantia for a single 464 explicandum, i.e. the same semantic data input is expressed 465 using different syntactic models [27], [28]. This feature is 466 extremely important with data of the kind we have dealt 467 with in the project, as attested by the different semantics that 468 are present in the database. Fuzziness offers a more natural 469 treatment for missing data, which can be considered to belong 470 to each fuzzy set with shallow confidence; as the diagnostic 471 pathways go on, ownership will increase in the correct class 472 and decrease elsewhere. The same results are more difficult 473 and less natural to obtain with classical architectures, as miss-474 ing data is usually swept under the rug of neurons, and treated 475 as an obstacle and not as a natural part of the phenomena that 476 are analysed. As well, since in coeliac disease diagnosis, the 477 ground truth is only obtained through an invasive procedure, 478 the fact that the system will inherently generate approximate, 479 imprecise results should also encourage the clinician to trust 480 the system and to gain complete control of the diagnostic 481 decisions, incorporating suggestion from the CDSS in a less 482 invasive instance.

483
The implemented CDSS is based on a fuzzy classifier 484 using neural networks, feed-forward backpropagation with 485 momentum. Development took place in Python 3.6 language 486 using the appropriate ML libraries. The system was trained 487 and tested on real data databases acquired in the ITAMA 488 Project and on the Virtual DB previously described.  The first process of data cleaning was applied to the bulk 502 data. First, each subject data consisting only of 0 or missing 503 values was excluded. Then, the optimization and tuning of the 504 system were carried out according to the scheme presented 505 in Fig. 8, based on a    • Step F:the process stops when adding other best vari-532 ables does not significantly improve classification, 533 as defined by a threshold selected by previous knowl-534 edge of the problem As the original number of vari-535 ables did not impede the classification speed, no a-priori 536 maximum number of variables was forced on the opti-537 misation process.

538
Following the variables optimisation process, an additional 539 parameters' choice was applied to the fuzzy neural network 540 by applying a simple gradient with momentum descent to the 541 number of neurons in each level and the number of levels. The 542 complete optimisation process has selected the parameters 543 reported in Table 4.

545
Once the classifier is completed, ITAMACDSS is interfaced 546 with the ITAMADB system to offer diagnostic support to 547 clinicians and health care personnel, and the system can 548 accept new patients' data as input. ITAMACDSS commu-549 nicates with ITAMADB System via a public API, using 550 a microservice architecture, once again shielded from the 551 final user. In the following, a brief description of the public 552 API is given. All the API commands are available using a 553 common endpoint. The following commands describe the 554 API. After the classifier has been trained, each time a new patient 585 data set is passed to the system, a Risk Factor (RF) and 586 a Confidence Factor (CF) are computed. RF is the clas-587 sification result, considered as the best class that fits the 588 data (MIN, LOW, MED, HIG, MAX). CF is a parameter 589 that addresses the idea that information about a patient is 590 partial and increases over time until a specific diagnosis is 591 reached. The confidence factor is computed according to the 592 following: each stage of the diagnostic pathway is associated 593 to a weight factor w i , according to clinical experience and 594 previous knowledge; in the simplest version, a linear scale 595 in the [0, 1] interval that mirrors the stage in the diagnostic 596 pathway. In the first version of the application, w 1 = 0.30, 597 w 2 = 0.55, w 3 = 0.70. Each weight is then multiplied by 598 a function of the level of belonging to the fuzzy set selected 599 by the algorithm. In the first version of the application, the 600 value is w i · 2(fs i ). The resulting value is then thresholded 601 in the [0, 1] range. Such value can be displayed next to the 602 patients' data directly or using any eidetic device (such as 603 colour, intensity, shapes).

605
As for the analysis conducted on the virtual database, training 606 of the already optimized model has been carried out using 607 10 batches of 10K virtual patients each, split 70/30 (training 608 set/test set), following the procedure detailed in the previous 609 section, that has been previously submitted to the system, 610 using custom-made Python code.

611
Due to the significant unbalance between the cardinality of 612 positive and negative virtual patients (negative positive), 613 the positive/negative ratio has been fixed in both training 614 and test sets in order to avoid having sets with only nega-615 tive patients. Table 5 shows the confusion matrix obtained 616 by averaging the results of 10 rounds of 10,000 virtual 617 patients each (repeated validation of random subsampling) 618 split between train and test sets according to the configuration 619 described in the previous section. Patients with no useful data 620 (i.e. 0 or NaN in all columns) have been omitted. 621 VOLUME 10, 2022 The sensitivity and specificity of the method can be 622 obtained from the confusion matrix in Table 5: i.e. as in equation (2): The values obtained for these two parameters are:   For the study on the real database, a total of 19,415 patients, 644 of which 109 diagnosed with coeliac disease, were analysed. 645 Also for this database, the same parameters used for the 646 virtual database and already presented in Table 4 were used. 647 While the positive/negative ratio is lower (0.006 for real data, 648 0.01 for virtual data), such values are in the same order of 649 magnitude, and allow a direct comparison of results from 650 the two models. Training for the real data model has been 651 carried out using the following parameters: split 80/20 (train-652 ing set/test set), with fixed positive/negative ratio. Patients 653 with no useful data (i.e. 0 or NaN in all columns) have 654 been omitted. Averaged and rounded results from 1000 split 655 batches using Monte Carlo repeated sub-sampling validation 656 are as in Table 6 (TP mean = 23.78, stdev = 1.41; FP mean = 657 1.00, stdev = 0.17; FN mean = 3.67, stdev = 1.36; TN 658 mean = 1399.76, stdev = 0.65). An example of the typical 659 membership function obtained through the classification pro-660 cess is given in Fig. 9.

661
By means of the confusion matrix in Table 6

670
In order to highlight the effectiveness of the proposed 671 method, an analysis was also carried out using the best known 672 and most used classifiers for medical imaging; SVM, kNN, 673 neural network. The comparison of the results obtained by the 674 various classifiers on the virtual database is shown in Table 7, 675  while the comparison of results on the real database is shown 676 in Table 8.   The fuzzy classification additionally allows to calculate risk 704 and confidence factors, that can be usefully employed in 705 evaluating priorities in the diagnostic pathway of the patients. 706 Furthermore, fuzzy approaches allow for suggestions on the 707 next steps to follow during the diagnostic pathway, instead of 708 at the end of it, as natural for other methods.

709
The data acquired and hosted in the DB highlights various 710 medical conditions associated with specific symptoms and 711 signs. The CDSS helps assessing the physical health of a 712 person by providing both a tool to support the diagnosis of 713 coeliac disease and a device capable of verifying the correct-714 ness of the progress of investigation during the identifica-715 tion of the disorders. Furthermore, the CDSS includes new 716 mathematical methodologies relating to the area of Artificial 717 Intelligence. These models are used to determine both the 718 functions of belonging to the various classes (Min to Max) 719 and the relative values to predict the onset of pathology.

720
A virtual DB was created which allowed for the tuning of 721 the proposed method. The performance of the CDSS on the 722 real and virtual database were comparable, thus confirming 723 the goodness of the implemented virtual DB. The results 724 obtained by classifying the virtual patients' data are much 725 better than what is usually obtained by classifiers applied 726 to clinical diagnosis. This can be explained by the fact that 727 since data is artificially created from the application of known 728 distributions (albeit through a random generation of parame-729 ters), the classifier had the chance to understand the overlying 730 distribution in an optimal way.

731
It can be observed that the ITAMACDSS classifier, when 732 used on real patients' data and in conjunction with a confi-733 dence factor assessment mechanism, represents a good pre-734 dictor of coeliac disease, and that by using the system as a 735 diagnostic support it is possible to support the clinician in 736 assessing the coeliac status of a patient with high accuracy 737 and precision by looking at blood tests and PoCT results, 738 reducing the number of costly and uncomfortable procedures 739 such as biopsy.

740
Further work remains to be done on a better correlation 741 between virtual and real data, in order to obtain a virtual 742 model of the coeliac parameters that can be helpful in further 743 revising clinical guidelines. Another area that can be consid-744 ered an active research topic lies in the presentation of the 745 tool's suggestions to the users of the systems, clinicians and 746 health personnel alike.

747
The evaluation process confirms the system's robustness 748 in the presence of a large amount of data (∼22K subjects) 749 and the adequacy of the results compared to international 750 statistics. However, the system can move towards more 751 advanced intelligent systems to support medical diagnostics. 752 The improved computational performance, the identification 753 of new diagnostic paths derived from data analysis and the 754 re-edition of the DB containing the data will mitigate the 755 necessary energy consumption. Furthermore, these aspects 756 will extend the field of action of the CDSS to other branches 757 of medicine in which a decision-making system based on 758 Artificial Intelligence finds huge interest and diversified 759 applications.    He is currently a Senior Data Scientist at synbrAIn. 978 He has been in academia for several years, with 979 more than three years of postdoctoral activities in 980 (University of Palermo) and (University of Hert-981 fordshire, U.K.). His research interests include the 982 field of human-computer interaction and applied 983 artificial intelligence, and he has visited and collaborated with many research 984 institute in Italy and Europe. He has authored more than 30 scientific 985 publications, and served as a conference organizer and a peer reviewer for 986 various international journals. He has also been an Adjunct Lecturer with 987 the University of Palermo. Prior to, and in parallel with his academic career, 988 he worked also as a web and mobile software developer, and a tech writer 989 and a content manager for several webzines.  [1985][1986][1987]. In this period, he participated in the 1060 experiment ALEPH, one of the four experiments 1061 at LEP, the accelerator site at CERN. In January 1987, he continued his 1062 research activity in the Italian component of ALEPH, as an Initial Researcher 1063 at the INFN, Bari Section. In 1998 and 1999, he was also a part of the 1064 CMS collaboration, one of the two experiments on particle physics that will 1065 come into operation, in 2008, at LHC (CERN), dealing with central tracking 1066 with silicon detectors. In November 1999, he was appointed as an Associate 1067 Professor in applied physics, as a winner of the national competition, at the 1068 Faculty of Engineering, University of Palermo. Since, he has been carrying 1069 out his research activity in applied physics at the Department of Physics and 1070 Chemistry, University of Palermo. Since January 2005, he has been a Full 1071 Professor in applied physics.