Original article: Clinical research SARCOIDOSIS VASCULITIS AND DIFFUSE LUNG DISEASES 2016; 33; 355-371 © Mattioli 1885 High-resolution computed tomography to differentiate chronic diffuse infiltrative lung diseases with chronic multifocal consolidation patterns using logical analysis of data Constance de Margerie-Mellon1*, Geneviève Dion2*, Julien Darlay3, Imene Ridene4, Marianne Kambouchner5, Nadia Brauner3, Michel Brauner6, Dominique Valeyre7, Pierre-Yves Brillet6 1Department of Radiology, Hôpital Saint-Louis, Université Paris 7 Denis Diderot, Sorbonne Paris-Cité, Paris, France; 2 Department of Pneumology, Centre de Recherche de l’Institut Universitaire de Cardiologie et de Pneumologie de Québec, Québec, Québec, Canada; 3 Labo- ratoire G-SCOP, Université Joseph Fourier, Grenoble, France; 4 Department of Radiology, Hôpital Abderrahmane Mami, Ariana, Tunisie; 5Department of Pathology, EA2363 Laboratory, Hôpital Avicenne, Université Paris 13, Sorbonne Paris Cité, Bobigny, France; 6 Department of Radiology, EA2363 Laboratory, Hôpital Avicenne, Université Paris 13, Sorbonne Paris Cité, Bobigny, France; 7Department of Pneumology, EA2363 Laboratory, Hôpital Avicenne, Université Paris 13, Sorbonne Paris Cité, Bobigny, France Abstract. Background: Chronic lung consolidation has a limited number of differential diagnoses requiring dis- tinct managements. The aim of the study was to investigate how logical analysis of data (LAD) can support their diagnosis at HRCT (high-resolution computed tomography). Methods: One hundred twenty-four patients were retrospectively included and classified into 8 diagnosis categories: sarcoidosis (n=35), connective tissue disease (n=21), adenocarcinoma (n=17), lymphoma (n=13), cryptogenic organizing pneumonia (n=11), drug-induced lung disease (n=9), chronic eosinophilic pneumonia (n =7) and miscellaneous (n=11). First, we investigated the patterns and models (association of patterns characterizing a disease) built-up by the LAD from combina- tions of HRCT attributes (n=51). Second, data were recomputed by adding simple clinical attributes (n=14) to the analysis. Third, cluster analysis was performed to explain LAD failures. Results: HRCT models reached a sensitivity >80% and a specificity >90% for adenocarcinoma and chronic eosinophilic pneumonia. The same thresholds were obtained for sarcoidosis, connective tissue disease, and drug-induced lung diseases when clini- cal attributes were added to HRCT. LAD failed to provide a satisfactory model for lymphoma and cryptogenic organizing pneumonia, with overlap between both diseases shown on cluster analysis. Conclusion: LAD provides relevant models that can be used as a diagnosis support for the radiologist. It highlights the need to add clinical data in the analysis due to frequent overlap between diseases at HRCT. (Sarcoidosis Vasc Diffuse Lung Dis 2016; 33: 355-371) Key words: interstitial lung disease, computed tomography, medical informatics Introduction Received: 7 Jenuary 2016 Consolidation appears on high-resolution com- Accepted after revision: 14 March 2016 Correspondence: Constance de Margerie-Mellon, puted tomography (HRCT) as a homogeneous in- Department of Radiology, Hôpital Saint-Louis, crease in pulmonary parenchymal attenuation that Assistance Publique-Hôpitaux de Paris, obscures the margins of vessels and airway walls (1). Université Paris 7 Denis Diderot, Sorbonne Paris-Cité, Paris, France Chronic consolidations may be due to diagnoses re- quiring distinct managements, especially regarding *These authors contributed equally to this work 356 C. de Margerie-Mellon, G. Dion, J. Darlay, et al. malignancy. HRCT plays a key role in the diagnostic consisted of understanding the overlaps between approach along with clinical data (2, 3). Consolida- diseases by analyzing misclassified or non-classified tions can be defined by their morphological charac- cases by LAD during the agglomeration of models teristics (contours, density, and bronchogram), distri- step and by using cluster analysis. Our aim was to bution, or their association with other features. Some obtain for each disease a model with the highest sen- HRCT features are considered highly indicative of sitivity and specificity. a disease. For instance, a bulging fissure sign favors adenocarcinoma (4), a superior and peripheral dis- Patient selection tribution advocates chronic eosinophilic pneumonia, and an association with a nodular galaxy sign sug- Three hundred one cases of possible chronic dif- gests sarcoidosis (5). However, the diseases often fuse infiltrative lung diseases with multifocal consoli- share similar HRCT features, and the specificity of dation on lung HRCT that were referred at a tertiary isolated signs remains to be determined. Moreover, care hospital (HÔpital Avicenne, Bobigny, France) in the daily practice, diagnosis is usually made by from January 1988 to August 2009 were retrospec- associating features, and the determination of perti- tively reviewed. Cases with predominant or exclusive nent combinations needs to be investigated. multifocal consolidation on lung HRCT, a definitive The development of data mining provides inno- diagnosis of chronic diffuse infiltrative lung diseases vating tools to extract relevant information from a da- and clinical symptoms evolving for at least 2 months tabase. Among the various available methods, logical were included, after optimal work-up and antibiotic analysis of data (LAD) allows, from a large set of data, treatment leading to exclusion of chronic infectious the identification of informative combinations of fea- diseases. Cases with the following criteria were ex- tures (named attributes) suggestive of a specific diag- cluded: lone lung consolidation (n=12), absence of nosis (6). LAD has already been applied in a series of lung consolidation at presentation (n=15), absence of medical studies (7-10), particularly for the diagnosis a definite diagnosis (no histological proof (n=47), in- of chronic diffuse infiltrative lung diseases presenting complete files (n=48), other (n=8)), prior comorbidity with predominant ground-glass opacity (11). Results known to be confusing (i.e., cardiac failure, known provided by the LAD constitute a diagnostic decision lung neoplasia, other chronic infiltrative lung disease support for non-expert radiologists (11). or lung infection) (n=16), immunodeficiency (HIV/ Our study aimed to investigate how logical AIDS, organ transplantation, immunosuppressive analysis of data (LAD) may improve the diagnosis of drug or corticosteroids use) (n=5) and clinical symp- chronic pulmonary consolidations at HRCT. toms evolving for less than 2 months (n=28). Even- tually, 124 cases were included and analyzed for this study (Appendix A). Methods Clinical data collection The present study was retrospective and mono- centric. It was validated by a local ethics committee, Clinical and epidemiological data were collected and informed consent was waived. First, we worked retrospectively from clinical records using a stand- on the clinical and HRCT attributes. We calculated ardized data sheet by a pulmonologist with expertise the interobserver agreement (Cohen’s Kappa test in infiltrative lung diseases (GD). For each case, the using the following κ ranges: 0.21-0.40=poor, 0.41- final diagnosis was confirmed in consensus by a pul- 0.60=fair, 0.61-0.80=moderate, 0.81-1.00=good), monologist (DV), a radiologist (MB) and a patholo- sensitivity and specificity of each HRCT attribute. gist (MK) specialized and experienced (> 20 years of Second, we defined HRCT and HRCT + clinical experience) in infiltrative lung disease according to models for the main diseases by combining attrib- strict criteria (Appendix B). Subsequently, cases were utes, and we calculated their sensitivity and speci- classified into eight categories: sarcoidosis (n=35), ficity. Our endpoint was to fulfill a specificity >90% connective tissue disease (n=21), adenocarcinoma (first optimal scenario) or >80% (second scenario) (n=17), lymphoma (n=13), cryptogenic organizing with the highest sensitivity as possible. The last step pneumonia (n=11), drug-induced lung disease (n=9), Logical analysis of data in chronic lung consolidation 357 chronic eosinophilic pneumonia (n=7) and miscella- chovascular, lobar or having no predominant distri- neous (n=11, Appendix B). bution. The cranio-caudal distribution was evaluated The following clinical and epidemiological at- in three zones: the upper zone above the level of the tributes were included (n=14, Table 1): age, gen- carina, middle zone between the level of the carina der, exposure to drugs known to cause lung disease, and level of the inferior pulmonary veins, and lower smoking history, Caucasian ethnicity, cough, dysp- zone below the level of the inferior pulmonary veins. nea, weight loss, asthenia, fever, extrathoracic visceral The axial distribution was also evaluated in three involvement, rheumatologic symptoms, crackles and zones: the central zone situated in the inner half of Raynaud syndrome or myalgias. Each attribute was the lung, peripheral zone referring to the outer half binary coded (presence=1 or absence=0) for the LAD of the lung and peripheral and subpleural zones situ- analysis except for age (in years). ated under the pleura. The antero-posterior distri- bution was evaluated in three equal zones (anterior, HRCT Protocol middle and posterior). Thirty-two additional HRCT features describing other pleuro-pulmonary and HRCT was performed using a Toshiba X-Press hilo-mediastinal abnormalities