In this study, the clinical and radiological screening diagnosis of IPF was subjected to system development. Two hundred ninety cases were diagnosed as DPLDs or IPFs and treated from December 2002 to February 2008 at Seoul National University Hospital and Yonsei University Hospital. One hundred nine of these patients (Seoul National University Hospital, 7; Yonsei University Hospital, 102) who sufficiently underwent diagnostic examination were included in the study. Among the patients whose diseases were not definitely diagnosed histologically, those having heart or renal failure that might influence radiological or clinical findings and those having other pulmonary diseases (lung cancer, chronic obstructive pulmonary disease) were excluded from this study.

Expert knowledge on IPF diagnosis was acquired by conducting interviews with respiratory specialists, radiologists, medical domain engineers, and system developers at regular group meetings. Knowledge modeling began with collecting knowledge from various pulmonary disease guidelines. The knowledge working group collected and defined diagnostic knowledge for DPLD and IPF, based on clinical experiences and medical records, through periodic study meetings. Simultaneously, websites, journals, and textbooks were studied with the intent to acquire up-to-date data and knowledge.

To refine the collected knowledge, we added and verified respiratory physicians' experience and clinical practice knowledge and applied it to the IPF CDSS diagnosis rules. These rules were applied to confirmed cases for validation. The differences shown in the respective disease groups were statistically analyzed by the chi-square test, and the information that showed a statistical significance was given clinical significance on the basis of respiratory specialist experience and knowledge. Rules recognized for providing clinically meaningful differences, which had been verified for validity based on clinically confirmed cases, were aligned with the actual diagnosis process and expressed in IF-THEN form for each stage of the process, according to the information acquisition time and method and importance of the differentiation priority.

Since the defined knowledge was designed to focus on increasing discrimination power, it was built by including as many patients as possible with the disease in the initial stage of the system, and thereafter excluding cases showing a clear decrease in possibility. The reasoning method consisted of repeatedly narrowing the range of target diseases with each condition added to the entire range of diseases, until the final diagnosis was reached.

In this study, however, we did not include surgical lung biopsy results in the diagnosis process of the IPF CDSS. If a final decision could not be made using other major and minor criteria testing results, the IPF CDSS would recommend a surgical lung biopsy as the next test procedure and physicians would carefully make a decision based on the patient's state.

The knowledge base was reconstructed by rearranging the knowledge collected and arranged in that base. The final diagnosis was made after information for diagnosing DPLD and IPF was acquired through the patient's medical history, pulmonary examination, radiological examination, BAL, and lung biopsy. As it was difficult to extract from the examination reports, the "insidious onset of otherwise unexplained dyspnea on exertion" was excluded from the criteria. A comprehensive knowledge base was rebuilt, which included the clinical and image findings required for the IPF diagnosis standard as presented by the joint statement of the ATS/ERS [3,11] and IPF diagnosis presented by Hunninghake et al. [17].

The data warehouse we constructed functions as an integrated hospital information system that potentializes the final clinical diagnosis and HRCT diagnosis. Figure 1 presents the integration architecture of 2 CDSSs and other hospital information systems, including PACS, using the CVS and data warehouse. The data warehouse in Figure 1 consists of HRCT image data from PACS, patient clinical information, patient physical examination results, and laboratory test results. The data mart could be handled with multidimensional online analytical processing (OLAP), being linked with the database of HRCT data and knowledge base of the CDSSs for DPLD and IPF diagnosis. It processes various and complicated queries requested by CDSSs. We used Microsoft SQL Server 2000 to develop the data warehouse and Microsoft Analysis Services to analyze the clinical data from the order communication system (OCS), laboratory information system (LIS), and other hospital information systems.

The DPLD CDSS integrated with the CVS was developed in our previous project and deduced diagnosis from 12 DPLD disease classifications [13]. The IPF CDSS helps respiratory specialists to diagnose IPF by using collected patient information from the data warehouse integrating hospital information systems. The interface and inference engine for the IPF CDSS was developed using .NET with the Visual Web Developer 2005 express edition and extensible markup language (XML), web standard language.

We developed an optional function to integrate the IPF CDSS with the CVS. In clinical practice, a respiratory physician makes a diagnosis with the help of a radiologistp the CS, using the a comprehensive knowledge base was rebuilt, which included tth automatic HRCT interpretation results from the CVS. The respiratory physician can access the information from the CVS.

Expert knowledge regarding IPF diagnosis was acquired by holding interviews with respiratory specialists, radiologists, medical domain engineers, and system developers at regular group meetings.

IPF, also termed cryptogenic fibrosing aveolitis (CFA) in Europe [18], is a distinct and specific form of chronic fibrosing interstitial pneumonia with no known cause. The ATS/ERS revision of the classification of DPLD [11] established the guidelines for the classification of IIPs, distinguishing between IPF and other IIPs [1]. Since they have a clinical similarity but their clinical courses and responses to therapy are different, the diagnostic process in patients with suspected DPLD begins by making the distinction between idiopathic and nonidiopathic diseases. For example, the diagnosis of sarcoidosis needs a compatible clinical picture of a systemic disease, histological demonstration of noncaseating granulomas, and exclusion of other diseases capable of producing a similar histological or clinical picture [19].

The common clinical feature of IPF is insidious progressive shortness of breath or dyspnea for at least 3-4 months [5,11]. Cough and minimal or no sputum is a general feature of IIPs. The presence of clubbing on physical examination is helpful in suggesting a diagnosis of IPF and pulmonary fibrosis associated with rheumatoid arthritis, asbestosis, or fibrosing NSIP [18]. A thorough clinical diagnostic approach begins with a detailed medical history, physical exam, PFTs, chest x-ray, lung function studies, blood tests, and tissue analysis [5,11]. An important goal of these multidisciplinary clinical approaches is the early diagnosis of the disease, which is enhanced by diagnostic agreement [16,20].

If clinical findings fail to suggest an alternative diagnosis, HRCT image findings should be obtained [21]. HRCT scanning has greatly improved the ability to visualize and characterize abnormalities in IPF. Extensive reticulation, traction bronchiectasis, honeycombing cysts, a subpleural predominance, and a somewhat patchy distribution are strongly suggestive of HRCT findings in IPF. The appearance of GGOs on HRCT is not characteristic of IPF [5]. Therefore, HRCT allows the clinician to make the distinction between "possible IPF" and "non-IPF" [5,18]. Hunninghake et al. have suggested that relatively accurate clinical and radiologic data, interpreted by an experienced pulmonologist or radiologist, are sufficient to diagnose IPF, precluding and even obviating the need for a lung biopsy [10,18].

Transbronchial biopsy (TBBx) and BAL have limited use as diagnostic tools, but are sufficient to help exclude other causes of pulmonary interstitial infiltrates, such as sarcoidosis, hypersensitivity pneumonitis, malignant disease, and infection, when surgical biopsy is not an option [5]. A recent article in Chest [19] suggested that the TBBx may have some utility with well-defined criteria. Berbescu et al. [21] examined TBBx samples from 22 UIP patients.

Restrictive pulmonary physiology is the classic finding on pulmonary function testing in IPF [20,22]. There is impaired gas transfer reflected in a decrease in carbon monoxide diffusion in the lung (DL_CO) and alveolar oxygen partial pressure (PaO₂), which is associated with increased mortality [18,23]. An increase in oxygen desaturation on exercise is well correlated with the extent of disease on HRCT. There may be hypoxemia with an increased alveolar-arterial O₂ gradient (A-aPO₂). In addition, although the typical findings of pulmonary function tests (PFTs) in IPF are consistent with restrictive impairment, PFTs or arterial blood gases (ABGs) may be normal. Typical findings of airway restriction include a decreased forced vital capacity (FVC) and total lung capacity (TLC) and DLco [3,5,11].

The presence of typical findings on clinical and radiologic examination may allow for a confident diagnosis, precluding the need for a surgical lung biopsy. A surgical lung biopsy in suspected IPF remains the gold standard for diagnosis, but it can be limited to those patients who show an atypical pattern, such as a predominant lymphocytosis on HRCT or in BAL, or an atypical clinical presentation, such as young age or short duration of illness. The surgical lung biopsy should be carefully considered and performed on the bases of HRCT findings [5].

In this study, we did not include surgical lung biopsy results in the IPF CDSS diagnostic process since the diagnosis of the CDSS focused on differentiating IPF from other DPLDs in primary care. Instead, the IPF CDSS recommends surgical lung biopsy as the next test procedure, and that physicians make a careful decision according to the patient's state.

For knowledge redefinition, we added verified respiratory physicians' experience and clinical practice knowledge to the IPF CDSS diagnosis rules. The collected knowledge was applied to cases of confirmed diagnosis for validation. The information that showed a statistical significance was given clinical significance on the basis of the respiratory specialists' experience and knowledge.

The collected knowledge based on major and minor criteria for the clinical diagnosis of IPF suggested by the joint statement of the ATS/ERS guidelines [3,11] and HRCT image findings for IPF diagnosis suggested by Hunninghake et al. [17,24] was refined and redefined. The differences shown in each DPLD group were statistically analyzed by the chi-square test, which was applied to differentiate IPF from other DPLD groups unrelated to connective tissue disease (CTD). In this analysis, we excluded an unexplained insidious dyspnea because it is currently difficult to acquire such cases.

For the first step, to differentiate IPF from CTD unrelated to other DPLD groups [2], the 18 IPF cases confirmed by diagnosis through surgical biopsy were compared with the 10 cases of other DPLD groups confirmed clinically (Table 2). The characteristics of these 2 diseases are apparently different in a population.

The different clinical findings between IPF and other DPLD disease groups were analyzed by the chi-square test. A p-value lower than 0.05 was considered statistically significant. The results (Appendixes 1 and 2) show that, among the clinical features, dyspnea on exertion (DOE) (IPF, 77.8%; other DPLDs unrelated to CTD, 30.3%; p < 0.05), duration from the beginning of having symptoms to visiting clinics (IPF, more than 14 days; other DPLDs unrelated to CTD, less than 12 days), and the presence of Velcro rale (IPF, 72.2%; other DPLDs unrelated to CTD, 20.2%; p < 0.05) were apparently different between IPF diseases and other DPLDs unrelated to CTD. The radiologic findings analyzed by the chi-square test demonstrate that consolidation, traction bronchiectasis, GGO, lower location of honeycombing, and subpleural lesions were significantly different between IPF and other DPLDs unrelated to CTD (p < 0.05) (Appendix 3).

To differentiate IPF from other DPLDs related to CTD, however, the 18 IPF cases confirmed with the diagnosis of surgical biopsy were compared to 10 cases of other DPLDs related to CTD (Appendixes 4 and 5). The characteristics of these 2 diseases have similarities in radiological diagnosis, but are apparently distinguished by clinical features and prognosis by a secondary cause of DPLD. The different clinical findings between IPF and other DPLDs related to CTD were analyzed by the chi-square test. A p-value lower than 0.05 was considered statistically significant. The different clinical findings between IPF and other DPLDs related to CTD were analyzed by the chi-square test. The results (Appendix 6) show that among the clinical features, ex-smoker (IPF, 11.1%; DPLDs related to CTD, 0.0%; p < 0.05), cough (IPF, 88.9%; DPLDs related to CTD, 10.0%; p < 0.05), DOE (IPF, 77.8%; DPLDs related to CTD, 40.0%; p < 0.05), weight loss (IPF, 11.1%; DPLDs related to CTD, 0.0%; p < 0.05), and the presence of Velcro rales (IPF, 72.2%; DPLDs related to CTD, 30.0%; p < 0.05) were apparently different between IPF diseases and other DPLDs related to CTD. The radiologic findings analyzed by the chi-square test (p < 0.05) show that consolidation, traction bronchiectasis, GGO, lower location of honeycombing, and subpleural lesions were significantly different between IPF and other DPLDs related to CTD (Appendix 7).

When the collected knowledge from guidelines such as the joint statement of the ATS/ERS and refined knowledge by statistical analysis were applied to 18 cases confirmed as IPF, the former diagnosed 1 case (6%) and the latter diagnosed 14 cases (78%). If traction bronchiectasis, one of the HRCT image findings, was removed from the diagnosis rules, the diagnosis performance of the collected knowledge was enhanced by 6 cases (33%). In addition, when the collected knowledge from guidelines such as the joint statement of the ATS/ERS and refined knowledge by statistical analysis were applied to 11 cases under medical treatment, the former diagnosed 1 case (9%) and the latter diagnosed 7 cases (64%). If traction bronchiectasis, one of the HRCT image findings, was removed from the diagnosis rules, the diagnosis performance of the collected knowledge was enhanced by 3 cases (27%). Therefore, refined knowledge performed better than collected knowledge.

The 3 diagnosis steps for IPF were extracted from refined knowledge. Figure 2 shows the flow chart of these steps, which we refined from the respiratory physicians' experience and clinical practice knowledge.

As the order of priority of the 3 steps was not provided in IPF diagnosis research literature, the order of priority in the diagnosis process was designed based on the empirical knowledge of experts verified by clinical trials at the redefinition stage of this research, and was reflected in the diagnosis process of CDSS.

Step 1 is the step for confirming the presence of clinical features. When the program is run by calling a patient's clinical information from the data warehouse through the data interface forms or with the manual input of information, and if, based on that basic clinical information, there is a chance of IPF, the clinician proceeds to the next step. If the likelihood of IPF is small, clinical information enhancement or PFT implementation is recommended. After the PFT result is input into the system, the next process is suggested based on the result. In step 2, the step diagnosing radiological features, the HRCT image findings and reading results are automatically received as direct input from the DPLD CDSS and CVS. From the DPLD CDSS, the diagnosis results confirmed by the radiologist are received as text. For high-level users able to read the information, in addition to the text reading option, direct image examination is made available as a diagnostic reference. When the clinical information and radiological information are ready, the likelihood of IPF is analyzed and the physician either modifies the information or proceeds to the next step. Step 3, the step diagnosing the presence of CTD-related DPLDs, after selecting the target group with a high likelihood of IPF, confirms the CTD-related DPLDs in order to eliminate the possibility of other associated diseases. Verification of antinuclear antibodies (ANAs) and rheumatoid factors, the markers for CTDs, is strictly enforced, and proceeding to the next step is allowed only after the physician confirms whether there are any associated CTDs present. Step 4, the diagnosis step to eliminate the presence of other associated diseases, as in cases of CTDs, eliminates cases of important associated lung diseases that are expected to bring substantial change to the prognosis and treatment guidelines. If lymphocytosis (in terms of the BAL differential count, lymphocyte > 15%) is shown from BAL, even if it clinically falls under IPF, the user is shown other diseases so that they can also be considered. Since the likelihood of IPF decreases for people under the age of 50 years, the user is given the opportunity to consider other associated diseases. The likelihood of IPF as the final inferred diagnosis result is presented for the data, which has been confirmed by the user in a conversation form. For the CDSS inference application, the knowledge is expressed with IF-THEN rules.

Patient clinical information from the OCS, past diagnostic information, physical and laboratory test information, HRCT image information, and knowledge and experience based on inquiry are needed in order for respiratory specialists to clinically diagnose IPF. Thus, hospital information systems should be integrated so that such diagnostic processes, as well as the process of knowledge acquisition, can be implemented in the system.

The diagnosis process of the IPF CDSS developed in this study is as follows (Figure 3). When the IPF CDSS program is started, a patient case list, which already exists in the data warehouse, appears on the CDSS. On clicking a patient name or number in the list, all laboratory examinations from the data warehouse, including clinical information and the diagnosis results from the DPLD CDSS of the given patient, are read into the IPF CDSS and displayed on the screen. The input of the confirmed diagnosis result among the 12 diseases from the DPLD CDSS, based on the image interpretation from the CVS, are both automatically and directly made.

If an examination required for IPF diagnosis is missing, the program recommends the given examination, carries out the diagnosis when all the information needed for that diagnosis has been input, and presents the diagnosis confirmation for the physician to review. If making the final diagnosis is difficult based on the given result, the system gives the physician the option to recommend surgical lung biopsy, based on the patient's condition.

IPF CDSS validation was implemented. First, the IPF diagnosis algorithm validation was implemented for the filtering process to express the diagnosis knowledge as a rule.

To verify the validity of the IPF diagnosis algorithm, 290 histological confirmation cases of DPLDs were collected for validation of the diagnosis rules and algorithms for 3 years (2005. 2.-2008. 2.) from Seoul National University Hospital and Severance Hospital in Yonsei University. The cases included HRCT image data, clinical laboratory test results, patient physical test results, and lung biopsy results. Out of the 109 cases that were critically diagnosed as IPF and being treated, 81 (74.3%) were diagnosed as IPF by the IPF CDSS. Likewise, the IPF CDSS was applied to the 28 cases that were diagnosed as other DPLDs. As a result, there were no cases where other diseases had been diagnosed as IPF (Table 3).

Meanwhile, validity analysis of the CVS-integrated CDSS was implemented with 10 patient cases diagnosed as IPF. The system detected 100% honeycombing from the HRCT images of the 10 patient cases, and the CDSS using the IPF diagnosis rules showed an accuracy rate of 98%. Considering such results, the IPF CDSS developed in this study, with its accurate IPF diagnosis decision-making, is considered able to aid in the quality of diagnosis.

For the evaluation of the CDSS, the DPLD CDSS was set in a clinical environment. Five radiologists at Seoul National University Hospital and 2 respiratory physicians at Yonsei University Hospital were educated to use the CDSS program, and evaluated the system usability. The evaluation of demonstrative operation was analyzed. When ascertaining the precision of diagnosis, the automatic image detection of the CVS and radiologists' gross examination indicated a consentaneity of about 99%.

A questionnaire was conducted on user satisfaction, qualitative improvement, and proposals, among other things, to examine the outcomes after clinical experiment. The users mostly felt that the CDSS would be of help to diagnose diseases and most of them were satisfied with the results. Many users suggested that HRCT findings should be classified as to meaning rather than presented in alphabetical order. Surveys and system use evaluations will be regularly collected and reflected iteratively for system modification, enhancement, and improvement to increase the accuracy and efficiency of diagnosis.