From December 2014 to May 2019, 1,321 inpatients from the Tuberculosis Department and Respiratory Department of West China Hospital of Sichuan University were initially recruited, during which time 1,765 outpatient health examination personnel were also recruited. A TB diagnosis was made according to the WS 288-2017 Diagnostic Criteria for Tuberculosis (Ministry of Health, People’s Republic of China, 2008), and all cases were confirmed by two experienced respiratory physicians. Finally, a total of 2,759 (12 + 798 + 1,650 + 299) subjects were ultimately included in this study, as detailed in Table 1. The experimental procedures met the ethical requirements, and the study was authorized by the Ethics Committee of West China Hospital of Sichuan University [No. 2014(198)].

The inclusion criteria for patients with suspected PTB were as follows: (1) history of epidemiological exposure to TB; (2) signs and symptoms of active PTB at the time of consultation; (3) clear microbiological and radiological imaging evidence of PTB; and (4) effective response to anti-TB treatment. The exclusion criteria for patients with PTB were as follows: (1) lung diseases other than TB; (2) positive HIV, HBV or HCV serology; (3) history of chronic liver/renal/hematologic diseases or autoimmune diseases; (4) pregnancy; and (5) immune dysfunction or history of immunosuppressive therapy or booster agents.

The inclusion criteria for the 1,650 medical examination volunteers from the same period were as follows: no previous history of TB infection or negative MTB bacteriology (i.e., smear antacid staining, MTB culture, and TB-DNA) and confirmation as healthy upon physical examination.

The inclusion criteria for the patients with non-TB lung disease were as follows: (1) three consecutive negative MTB bacteriological tests (smear antacid staining, MTB culture, and TB-DNA examination); (2) clear diagnosis of a non-TB lung disease; and (3) effective antibiotic treatment or other non-TB drug regimens and confirmed by one year of follow-up observation. The 299 cases of non-TB lung diseases were dominated by pneumonia, pneumonia with other diseases, and lung cancer.

All imaging items and laboratory tests were performed at the Department of Imaging and Department of Laboratory Medicine of West China Hospital, respectively, and clinical tests and reports were completed by radiologists of West China Hospital who were unaware of the final diagnosis and grouping of patients.

RNA isolation and cDNA preparation: peripheral blood mononuclear cells (PBMCs) were isolated using DAKEWE human lymphocyte isolation tubes, and the total RNA from PBMCs was extracted by the TRIzol method. The concentration and purity of the extracted RNA were measured using a NanoDrop ND-1000 Microspectrophotometer, and RNA samples with absorbance ratios >1.8 at 260/280 nm and >2.0 at 260/230 nm were considered of acceptable quality. The integrity of the RNA samples was examined using agarose gel electrophoresis, and high-quality samples were retained for subsequent experiments. Complementary deoxyribonucleic acid (DNA) was prepared by removing genomic DNA contamination from the reaction system using the PrimeScript RT reagent kit.

An HTA2.0 microarray chip was used to detect the expression of lncRNAs in the primary screening cohort. The expression data of lncRNAs were analyzed, and Gene Ontology (GO) and pathway bioinformatics analyses were performed on the differential lncRNAs. Candidate lncRNAs were screened by the aforementioned criteria.

Total RNA from PBMCs of all subjects in the selection cohort was extracted, and the relative expression of candidate lncRNAs was detected by qRT-PCR. Relative lncRNA expression was measured in a blinded fashion, normalized to the endogenous control GAPDH, and calculated according to the 2^–ΔΔCq method (Mavridis et al., 2013), where ΔCq = Cq lncRNA − Cq GAPDH, ΔΔCq = ΔCq − Δaverage Cq (healthy subjects). The expression differences between TB patients and healthy controls (HCs) (TB vs. HC), TB patients and non-TB DC patients (TB vs. non-TB DC) were compared, and candidate lncRNAs with differences can be used as new molecular markers of tuberculosis and determined to be suitable for construction. Model target lncRNAs.

The patient population (TB patients and patients with non-TB lung disease) of the selection cohort was randomly divided into a training set (399 TB + 150 non-TB DC) and a validation set (399 TB + 149 non-TB DC). The training set was used for training the model, and the validation set was used for validating the model. Based on the lncRNA and EHR data in the training set, the R “bestglm” function was used for variable subset selection, and binary logistic regression was used to construct three models: “lncRNA + EHR,” “EHR,” and “lncRNA.” Their diagnostic performances were compared to determine the optimal differential diagnosis model for PTB.

Presentation and evaluation of the nomogram: the optimal model was visualized as a nomogram to assess its discrimination and calibration, and the nomogram was validated internally by applying a 500 bootstrap self-sampling method and externally validated with independent validation set data. The decision curve analysis (DCA) net benefit method was used to compare the clinical potentials of the nomograms and diagnostic models.

HyLown Power and Sample Size Calculators¹ were used to estimate the required sample size of the lncRNA population study (Bacchetti and Leung, 2002) with the following parameters: α = 0.05, power value = 80%. Normally distributed count data were expressed as the mean ± standard deviation (X ± SD), and non-normally distributed count data were expressed using median and interquartile spacing [M (P25–P75)]. When comparing two groups, categorical variables were compared using the Chi-square test, and continuous variables were compared using the t-test or Mann–Whitney U test. Correlations between lncRNAs and clinical data were analyzed using the Spearman rank correlation test. All tests were two-sided probability tests, and P-values < 0.05 indicated a statistically significant difference. Data were analyzed, and the results were plotted using SPSS v24.0 and GraphPad Prism v8.0. Primary screening of modeling features (variables) was performed using R version 3.6.1 software and SPSS v24.0.

There were no statistically significant differences in age or sex between TB patients and HCs in the primary screening cohort (7 males and 5 females, aged 22–59 years). The basic demographic information and laboratory indices of the selection cohort are detailed in Table 2, and 40 EHR indicators were collected for the modeling study. There were no statistically significant differences between the TB and HC groups matched for age, sex and body mass index (P > 0.05). The mean age of all patients in the disease control group (non-TB DC) was greater than that of the TB patients (P < 0.001). Significant differences were found in multiple laboratory tests, signs and symptoms, and imaging, suggesting that these indicators could be used as preliminary screening variables for modeling.

The quality control data of the microarray assays in the primary screening cohort of this study met the standards, indicating that the Affymetrix HTA 2.0 gene chip test was successful and can be used for lncRNA expression data analysis. A total of 325 differentially expressed lncRNAs were identified in the HTA 2.0 microarray between the TB and HC groups, of which 287 were upregulated and 38 were downregulated. According to the parameters fold-change >2, original signal value >2⁵, and no literature report available, the four candidate lncRNAs were three upregulated lncRNAs, namely, TCONS_00013664 (chr6:85677081–85678394), ENST00000416679 (chr7:26392–35472), and TCONS_00001838 (chr1:223963605–223971768), one downregulated lncRNA, n406498 (Chr2:113599027–113601576).

The candidate lncRNAs were further bioinformatically analyzed to predict the signal transduction pathways and biological processes that might be associated with them, and the results are shown in Figure 2. Pathway analysis showed a total of 30 enriched signaling pathways for the differentially expressed lncRNAs, of which the top 10 enriched scores were mainly associated with the MAPK signaling pathway, apoptosis, cell cycle, and the Jak-STAT signaling pathway. GO analysis showed that the differentially expressed lncRNAs may be involved in regulating numerous biological processes, such as lipopolysaccharide response, apoptosis, inflammatory response, immune response, and small molecule metabolism.

qRT-PCR was used to detect the expression of the four candidate lncRNAs in the selection cohort of TB patients and HCs. It was found that lncRNA TCONS_00013664 had extremely low expression (Cq > 35) in both groups and was excluded from further analysis (Bustin et al., 2009). The relative expression levels of the remaining three lncRNAs are shown in Figure 3 and Table 3. The relative expression levels of ENST00000416679, TCONS_00001838, and n406498 in the HC group were 0.94, 1.12, and 0.99, in the TB group were 0.62, 0.33, and 0.70, and in the non-TB DC group were 0.68, 0.54, and 0.91, respectively. The expression of the three lncRNAs was significantly different between the TB and HC groups (all P < 0.001). The area under the ROC curves (AUCs) of the three lncRNAs were all >0.50, the AUC of lncRNA TCONS_00001838 was 0.828, the sensitivity was 76.24%, and the specificity was 72.93%. The AUC of lncRNA ENST00000416679 was 0.622, the sensitivity was 38.24%, and the specificity was 81.45%. The AUC of lncRNA n406498 was 0.602, the sensitivity was 37.88%, and the specificity was 82.58%, the results are shown in Supplementary Table 1. The expression of TCONS_00001838 and n406498 was significantly different between the TB and non-TB DC groups (P < 0.001). Suggesting that TCONS_00001838 and n406498 could be used as preliminary screening variables for modeling.

To establish a diagnostic model for distinguishing tuberculosis patients from non-tuberculous lung disease patients, 12 characteristic variables were initially included in this study for model construction: the 2 lncRNA loci and 10 EHR indicators mentioned above (age, hemoglobin, white blood cell count, lymphocyte count, gamma interferon release test, hypothermia, weight loss, night sweats, imaging examinations showing polymorphic changes, and calcified foci). Three binary logistic regression models were constructed: an “LncRNA + EHR” combined model, “LncRNA” model and “EHR” model. The optimal subset of variables included in each model was evaluated by the exhaustive search method. Finally, the above two lncRNA loci and eight EHR indicators (age, hemoglobin, lymphocyte count, gamma interferon release test, weight loss, night sweats, imaging examinations showing polymorphic changes, and calcified foci) were selected as optimal subset of variables. The variance inflation factors of the characteristic variables ranged from 1.06 to 1.19, indicating that there was no covariance among the indicators included in the models.

ROC curves of the three models in the selection cohort are shown in Figure 4. In the training set, the AUCs of the three models are shown in Table 4. The AUC of the “LncRNA + EHR” model was 0.89, the highest among the three models which was significantly higher than that of the “LncRNA” and “EHR” models (P < 0.001), For the differential diagnosis of TB, the “LncRNA + EHR” model had a sensitivity of 0.86 and a specificity of 0.76, This shows that the optimal model for the differential diagnosis of PTB was the “LncRNA + EHR” model. The likelihood ratio test, McFadden R² test and Nagelkerke R² test were used to assess the goodness of fit of the “LncRNA + EHR” model with 10 variables, and the results were P < 0.0001, 0.372, and 0.5, respectively. Indicating that the model variables have good goodness of fit and that there is no overfitting.

A nomogram of the visualized “LncRNA + EHR” model is shown in Figure 5, and the distribution of the scores and weights of the variables in the nomogram showed that TCONS_00001838 (lncTP), age (AGE), and lymphocyte count (L) were all negatively correlated. In contrast, ENST00000416679 (lncPA), hemoglobin (Hb), weight loss (weight loss), night sweating (Night_sweat), imaging showing polymorphic changes (CT_polymorphic), imaging showing calcification (CT_calcification), and gamma interferon release test (TB_IGRA) were the positively correlated variables.

The accuracy of 500 “bootstrap” self-sampling validations using the training set data was 0.83 (0.80–0.86), and the calibration curve and Hosmer–Lemeshow test were not significantly different (P = 0.755), confirming the high consistency between the predicted results of the nomogram and the actual results. Using the validation set data to validate the three models, the “LncRNA + EHR” model also had the best diagnostic performance for the differential diagnosis of TB, with an accuracy of 0.79 (0.75–0.82), a sensitivity of 0.81 (0.78–0.86), a specificity of 0.73 (0.64–0.79), and an AUC of 0.86. The calibration curve and the Hosmer–Lemeshow test (P = 0.174) showed that the nomogram also showed good predictive compliance for the differential diagnosis of TB in the validation set. Using DCA to compare the potential clinical application of the combined model with the conventional “EHR” model, the results of the DCA analysis in Figure 6 show that at a threshold of 0.2–1, the nomogram (blue dashed line) is more accurate than the conventional “EHR” model (red line). This shows that the use of the nomogram may improve the prediction level of PTB, indicating that the nomogram has better predictive value for clinical application.