| Part I.Study of comorbidities in patients with ILDsBackground and objectives: Comorbidity not only further complicate the diagnosis of ILD,but may also,to some extent,influence the choice of treatment options and disease prognosis.The high incidence of certain comorbidity in patients with ILD suggests that there may be similar pathogenesis between several diseases,or that one disease may be a background condition for the development of other diseases.However,there is a lack of large scale studies on the comorbidity characteristics of ILD,especially in combination with VTE and malignancies.Based on this,the aim of our study was to conduct a single centre and large scale analysis of the comorbidity characteristics of different ILD subtypes,with a particular focus on the occurrence of VTE and malignancies.Methods: Information on patients with ILD clearly diagnosed by MDT at China-Japan Friendship Hospital was collected,and patients were classified into the following subtypes according to the 2013 American Thoracic Society/European Respiratory Society(ATS/ERS)guidelines[1]: IPF,NSIP/i NSIP,COP,HP,CTD-ILD,AAV-ILD,and other categories of ILD.Among them,CTD-ILD was classified in detail according to etiology into IIM-ILD,SS-ILD,RA-ILD,SLEILD,and other CTD-ILD.Basic information and comorbidity of patients were obtained through the electronic medical record.VTE includes pulmonary thromboembolism(PTE)and deep vein thrombosis(DVT).The occurrence of malignancies was obtained through electronic medical records and telephone follow-up,and the follow-up time was at least 6 months.The Mann-Whitney U test,Fisher exact test or chi-square test were used to assess differences between groups for categorical or continuous variables.Results: 1.Analysis of comorbidities characteristics of patients with ILD(1)A total of 5009 patients with ILD who were hospitalized in China-Japan Friendship Hospital from January 2016 to March 2022 were included.The incidence of comorbidities in patients with ILD was 78.3%.The incidence of comorbidities in patients with IPF,NSIP,CTD-ILD,COP,and HP were 82.2%,80.8%,76.9%,75.5%,and 73.6%,respectively,which with the highest incidence of 1-3 comorbidities of 77.1%,74.5%,75.7%,73.6%,and 70.8%,respectively.According to different etiologies of CTD-ILD,the incidence of comorbidities in patients with SLE-ILD,RA-ILD,IIM-ILD,and SS-ILD was 82.6%,80.3%,77.6%,and 76.8%,respectively.(2)The comorbidities with the three highest incidences were cardiovascular system disease,respiratory system disease,and diabetes,with incidence of 34.1%,17.2%,and 16.1%,respectively.The incidence of comorbidity increased significantly with age,with 70.0%,81.4%,and 88.1% of patients in the <60,60-69,and ≥70 years groups,respectively.2.Incidence and risk factors of VTE in patients with ILD during hospitalization A total of 5009 patients with ILD who were hospitalized in China-Japan Friendship Hospital from January 2016 to March 2022 were included.The incidence of VTE was 2.6%,of which 0.3% was combined PTE and DVT,0.7% was PTE only,and 1.6% was DVT only.The incidence of VTE in patients with AAV-ILD,HP and IPF was 7.9%,3.6%,and 3.5%,respectively.The incidence of VTE,DVT,PTE,both PTE and DVT in patients with CTDILD was 3.0%,2.3%,0.4%,and 0.3%,respectively.The incidence of VTE increased with the prolongation of time in hospital,with length of hospitalization time <7 days,7-14 days,15-21 days,and >21 days,the incidence was 2.2%,2.3%,3.3%,and 6.4%,respectively.Different ILD subtypes,advanced age,respiratory failure,and varicose veins were independent risk factors for the development of VTE in patients with ILD.3.Analysis of characteristics of malignancies in patients with ILD A total of 5350 patients with ILD hospitalized in China-Japan Friendship Hospital from January 2015 to December 2021 were included.The incidence of malignancies in ILD population was 4.6%,among which lung cancer was the most common,followed by digestive system malignancies,breast cancer,lymphoma,genitourinary system malignancies,thyroid cancer,and head and neck malignancies,with incidence of 1.9%,0.9%,0.6%,0.4%,0.4%,0.3%,and 0.1%,respectively.Patients with IIP had a higher incidence of malignancy than patients with CTD-ILD,at 5.3% and 2.5% respectively.In different ILD subtypes,the incidence of malignancies in patients with COP,IPF,AAV-ILD,i NSIP,CTD-ILD,HP,and sarcoidosis was 6.8%,5.0%,4.7%,4.3%,2.5%,2.2%,and 1.2%,respectively.Lung cancer was the most common malignancies in IPF,with an incidence of 3.9%.The incidence of malignancy in patients with RA-ILD,IIM-ILD and SS-ILD in CTD-ILD was 2.4%,2.3% and 2.3%,respectively.In patients with ILD in combination with lung cancer,adenocarcinoma was the most common,followed by squamous carcinoma and other pathological types at 53.5%,28.7%,and 17.8 respectively.Conclusions: 1.The type of comorbidity varied among different ILD subtypes.As the age increased of patients with ILD,the incidence and the number of comorbidity increased significantly.Focusing on the aggregation characteristics of comorbidity in each subtype of ILD is conducive to the development of more precise prevention and treatment plans,and provides useful information for the exploration of the mechanisms of comorbidity.2.The incidence of VTE in hospitalized patients with ILD was 2.6%,and the incidence of DVT was higher than that of PTE.The incidence of VTE increased with the length of time in hospital.Risk factors for the development of VTE in patients with ILD such as ILD category,advanced age,respiratory failure,and varicose veins should be closely monitored.The types of malignancies occurred varied considerably between ILD subtypes,and physicians should formulate personalized early screening and management strategies according to the characteristics of malignancies in different ILD subtypes.Part II.Analysis of the assessment value of artificial intelligence models for ILDBackground and objectives:F-ILD,represented by IPF has become an important public health problem.Early identification of different ILD subtypes and quantitative analysis of different lesion patterns facilitates early diagnosis of the disease,assessment of disease severity and decision making for therapy options.The purpose of this part of the study was to develop a lung graph-based machine learning model to identify f-ILD and achieve the classification of several common ILD subtypes.At the same time,quantitative analysis was performed for IPF to assess disease progression.Two main approaches were used for quantitative analysis,one was to build a deep learning model to identify and quantify different lesion patterns,especially fibrosis patterns.The second was to quantitative and visual analysis lung shrinking based on elastic registration techniques to assess disease progression.Methods: Patients with ILD clearly diagnosed by MDT at our institution from January 2017 to December 2022 who were eligible for the study design and 50 healthy physicals were included in the study.1.A lung graph machine learning model based on HRCT was proposed to realize the identification of f-ILD and the classification of common ILDs.All HRCTs were randomly split five times(80% data for training,20% data for testing),in each split,the dimensionality reduction operation was performed on the training set and the five-fold cross-validation method was used to select the best model.The optimal model was validated on the test set.The model prediction results were further compared with physician assessments in an independent validation set.The area under the receiver operating characteristic curve(AUC),sensitivity,specificity,positive predictive value,and negative predictive value were used as evaluation indicators of model performance.Weighted kappa coefficients(kw)were used to assess the interobserver agreement between the model and radiologists.2.Deep learning model was used to segment four types of lesion patterns of IPF,including honeycombing(HC),reticular pattern(RE),traction bronchiectasis(TRBR),ground glass opacity(GGO).Dice coefficient,recall,precision,and false positives were used to evaluate model segmentation performance.Two methods were used for quantitative analysis:(1)Deep learning model-based dynamic lesion quantification included 246 HRCTs of 123 IPF patients.Patients were divided into groups of disease stability and disease progression according to lung function,and the differences of the extent of lesions and pulmonary vascular parameters between the two groups were analyzed.(2)Quantitative analysis of lung shrinking in IPF based on elastic registration technique strictly included 138 HRCTs of 69 eligible IPF patients.Point-to-point elastic registration was performed on the baseline and follow-up HRCTs to obtain deformation maps of the whole lung.Jacobian determinants were calculated from the deformation fields and after logarithmic transformation,the corresponding log_jac values were expressed on the color maps to describe visual or functional progression,and the correlation between log_jac values and indicators of pulmonary vascular and pulmonary function was further analyzed.Results: 1.Lung graph-based machine learning model for the identification of f-ILD A total of 417 HRCTs of 279 patients were included(f-ILD: 156 patients,223 HRCTs;non-f-ILD: 123 patients,194 HRCTs).The Weighted Ensemble model showed the best predictive performance in cross-validation in the five randomly divided training sets.Model performance was further assessed in the independent validation set.The classification accuracy of the model was significantly higher than that of 3 radiologists at both the CT sequence level and the patient level.At the patient level,the diagnostic accuracy of the model and radiologists A,B,and C was 0.986(95% CI: 0.959 to 1.000),0.918(95% CI: 0.849 to 0.973),0.822(95% CI: 0.726 to 0.904),and 0.904(95% CI: 0.836 to 0.973).There were significant statistical differences of AUC values between the model and three physicians(P<0.05).2.Lung graph-based machine learning model for the classification of common ILDs A total of 767 HRCTs of 459 patients with IPF,NSIP,COP,and HP were included.In the independent validation set,the model and radiologists had the highest diagnostic accuracy for IPF compared to NSIP,COP,and HP.At the CT sequence level,the diagnostic accuracy of the model was lower than that of radiologist A,and reached or even exceeded that of radiologists B and C,with 0.930(95% CI: 0.880 to 0.970),0.970(95% CI: 0.930 to 1.000),0.930(95% CI: 0.870 to 0.970),and 0.880(95% CI: 0.800 to 0.940),respectively.The model also achieved good classification performance in multi-classification,with a macro AUC value of 0.924(95% CI:0.885 to 0.957)at the CT sequence level.The diagnostic accuracy of the model was better than that of radiologists B and C,but lower than that of radiologist A at 0.730(95% CI: 0.640,0.810),0.680(95% CI: 0.580 to 0.770),0.580(95% CI: 0.480 to 0.670),and 0.770(95% CI: 0.680 to 0.850),respectively.3.Image features recognition of IPF based on deep learning model A total of 327 HRCTs from 327 patients with IPF and 50 healthy individuals with physical examination without chest lesions were included.The deep learning model had a lung segmentation precision on the test set was 0.995 ± 0.009.The overall lesion segmentation precision,Dice coefficient,and the recall was 0.965 ± 0.038,0.972 ± 0.036,and 0.980 ± 0.046,respectively.The fibrosis pattern segmentation precision was 0.960±0.067.Among different lesion patterns,RE and TRBR had the highest segmentation precision,followed by HC and GGO,which was 0.946±0.072,0.927±0.145,0.777±0.338,and 0.720±0.380,respectively.The precision of overall lesion segmentation decreased with the increase of disease severity,and the precision of GAP I,GAP II,and GAP III was 0.967±0.040,0.964±0.036,and 0.959±0.037,respectively.4.Dynamic quantitative analysis of IPF based on deep learning model There were no significant differences in baseline pulmonary function indicators,total lesion volume,extent of lesions at different sites of the lung,and pulmonary vascular parameters between the disease stability and disease progression groups(P>0.05).During follow-up,patients in the disease progression group had a significantly lower median total lung volume compared to those in the disease stability group,with-278.01 ml(IQR,-901.64 to-91.54 ml)and-40.76 ml(IQR,-212.45 to 286.86 ml),respectively(P<0.001).The median change in total lesion volume and total lesion ratio was significantly higher in the disease progression group than in the disease stability group,with total lesion volume changes of 214.73 ml(IQR,68.26 to 501.46 ml)and 3.67 ml(IQR,-71.70 to 85.33 ml),respectively(P=0.001).The decrease in pulmonary vascular volume and the number of pulmonary vascular branches was more obvious in the disease progression group than in the disease stability group.The change in total lesion volume ratio were negatively correlated with change in the percentage of predicted diffusing lung capacity for carbon monoxide(DLco%)(r=-0.57,P<0.001).The changes of pulmonary vascular related parameters were positively correlated with DLco%(r value range from 0.27 to 0.53)and percentage of predicted forced vital capacity(FVC%)(r value range from 0.44 to 0.61)(P<0.001).5.Evaluation the progression of IPF based on elastic registration technology The Jacobian maps of IPF patients showed that the lung base shrinking significantly in the visual and functional progression groups.The mean log_jac values were significantly lower in patients with functional progression compared to patients with functional stability.The mean log_jac value was positively correlated with pulmonary function and pulmonary vascular related parameters(P<0.05).Conclusions: 1.The lung graph-based machine learning model achieved high performance in f-ILD identification and classification of common ILD subtypes,which reached the performance of experienced radiologists,and had great significance potential to assist doctors in the early and accurate diagnosis of ILD.2.In this study,a deep learning model was established to achieve high segmentation precision in lung regions and typical lesions of patients of IPF,which to some extent overcome the limitation of interobserver differences in visual assessment.Model provided a basis for quantitative analysis of the degree of IPF lesions.3.Compared with the baseline imaging parameters,the dynamic quantitative changes of lesion volume and pulmonary vascular related parameters may be important imaging markers for the evaluation of disease severity and the prediction of disease progression in patients with IPF.Elastic registration of follow-up and baseline HRCTs allowed quantification of IPF lung shrinking to assess disease progression,and the results could be visualized by Jacobian maps. |
PDF Full Text Request