Nontuberculous mycobacteria (NTM) refer to a broad category of mycobacteria excluding the Mycobacterium tuberculosis complex and the Mycobacterium leprae complex. To date, approximately 200 species and 13 subspecies have been identified, most of which are opportunistic pathogens. These bacteria can invade the human body through the respiratory tract, gastrointestinal tract, skin, and other pathways, affecting multiple sites such as the lungs, skin and soft tissues, lymph nodes, and bones. Among these, the lungs are one of the most common sites of infection, leading to a condition known as nontuberculous mycobacterial lung disease (NTM-LD) (Kumar et al., 2024). In recent years, the incidence and prevalence of NTM-LD have been steadily increasing worldwide (Johansen et al., 2020). A 2024 epidemiological study confirmed that NTM constitute approximately 6.8% of sputum acid-fast bacilli smear-positive cases misdiagnosed as tuberculosis (Chen et al., 2024), underscoring significant risks of diagnostic errors and inappropriate therapeutic interventions. Moreover, the intrinsic resistance of NTM species to conventional anti-tuberculosis regimens necessitates protracted, multifaceted treatment protocols, which frequently entail severe drug-related toxicities and elevated treatment discontinuation rates. Consequently, NTM has emerged as a critical global threat to respiratory health. NTM-LD now represents not only a dominant etiology of chronic pulmonary infectious pathology but also a pivotal domain demanding intensified surveillance, advanced diagnostic methodologies, and targeted research initiatives.

NTM-LD often presents clinically with symptoms such as coughing with sputum hemoptysis, chest tightness, shortness of breath, low-grade fever, and night sweats. These symptoms are very similar to the clinical manifestations of pulmonary tuberculosis lung disease (PTB-LD) caused by infection with Mycobacterium tuberculosis. The clinical differentiation between NTM-LD and PTB-LD remains challenging due to similar manifestations. Current diagnostic workflows prioritize sputum smear microscopy, specifically Acid - Fast Bacilli (AFB) staining, as an initial screening tool for mycobacterial infections, but this method cannot distinguish NTM from Mycobacterium tuberculosis complex. Definitive diagnosis of NTM requires bacterial culture followed by species identification—a process that takes approximately 8 weeks under stringent biosafety protocols (Haworth et al., 2017). During this prolonged diagnostic period, patients with AFB-positive results are often empirically prescribed anti-tuberculosis therapy (ATT) to mitigate potential PTB transmission risks. However, this approach poses dual challenges: (1) overtreatment, exposing a subset of patients to unnecessary drug toxicity without clinical benefit, and (2) delayed targeted therapy, as most NTM species demonstrate intrinsic resistance to standard anti-tuberculosis agents, necessitating species-specific multidrug regimens distinct from PTB protocols (Conyers and Saunders, 2024). Therefore, establishing a stratified management system based on accurate differentiation between NTM-LD and PTB-LD holds significant clinical implications for achieving personalized diagnosis and treatment as well as optimal allocation of medical resources. With the advancement of analytical methods, the application of multifactorial mathematical models in the medical field has gradually expanded. Currently, numerous studies have explored the imaging-based differentiation between NTM-LD and PTB-LD, encompassing multiple modalities such as chest radiography and computed tomography (CT) (Xing et al., 2020; Park et al., 2023). However, existing research has predominantly focused on the discriminative value of individual imaging features, with limited systematic integration of clinical data (e.g., age, comorbidity profiles, and clinical symptoms) and quantitative imaging parameters for multidimensional analysis. Notably, there remains a paucity of exploration into integrating clinical-imaging multimodal indicators to construct mathematical prediction models. This study aims to evaluate the basic clinical data, symptoms and signs, quantitative CT imaging indicators, and comorbid conditions of NTM-LD and PTB-LD patients treated at our hospital since 2016. Through the development and validation of disease prediction models, we aim to provide predictive tools for early differential diagnosis of NTM-LD and PTB-LD, thereby enabling precision risk stratification with differentiated resource allocation and targeted interventions.

Differentiating NTM-LD from PTB-LD remains a significant diagnostic challenge in respiratory medicine. This study developed a high-performance predictive model by integrating multidimensional indicators, including clinical characteristics, radiological patterns, and comorbidities. Furthermore, we designed an interactive web-based calculator to visualize the model’s outputs, providing a novel tool for early clinical discrimination between these two entities.

This study revealed distinct clinical profiles between NTM-LD and PTB-LD. Specifically, patients with NTM-LD demonstrated a significantly higher prevalence of chronic cough with sputum production (OR=6.185, P<0.001) and hemoptysis (OR=6.702, P<0.001), which aligns with the airway injury patterns in NTM-LD reported by Youssefnia A.et al (Youssefnia et al., 2022). The virulence of NTM is relatively lower compared to Mycobacterium tuberculosis, yet their enhanced adhesive capacity enables persistent colonization of airway mucosa. Characterized by unique lipid-rich cell wall components that resist phagocytic clearance and impede their elimination, NTM induce chronic airway inflammation through sustained antigenic stimulation. This pathological process promotes excessive mucus secretion and consequently exacerbates clinical manifestations of cough with sputum production (Schiff et al., 2019). Furthermore, NTM-triggered inflammatory responses disrupt vascular integrity in bronchial walls, resulting in enhanced vascular permeability. Concurrently, airway smooth muscle hyperreactivity and bronchoconstriction elevate intraluminal pressure, which synergistically exacerbates microvascular damage and significantly increases the risk of hemoptysis (Zhang et al., 2019). Moreover, our study identified several distinct imaging biomarkers-bronchiectasis, centrilobular nodules, and thin-walled cavities-that collectively reflect the unique pathophysiological features of nontuberculous mycobacterial lung disease (NTM-LD). The strong association between bronchiectasis and NTM-LD (OR=28.604, P<0.001) further supports the predominantly airway-centered, chronic inflammatory nature of this condition. Bronchiectasis is both a predisposing factor and a consequence of NTM infection (Retuerto-Guerrero et al., 2024). Structural airway damage facilitates persistent bacterial colonization, while chronic inflammation from NTM infection exacerbates mucociliary dysfunction and smooth muscle hyperreactivity, leading to progressive airway dilation—a self-perpetuating cycle known as the “vicious cycle hypothesis” in chronic respiratory diseases (Dartois and Dick, 2024). Notably, our findings align with those of Chu et al (Chu et al., 2015), highlighting its utility as a distinguishing imaging feature.We also observed a significantly higher prevalence of centrilobular nodules in NTM-LD (OR=11.122, P<0.001), consistent with findings by Iakobachvili et al (Iakobachvili et al., 2022). These nodules arise from bronchial spread of infection along the airway tree and represent early inflammatory foci within terminal bronchioles, typically appearing as “tree-in-bud” opacities on high-resolution computed tomography (HRCT). This pattern is widely recognized as a hallmark of NTM-LD, reflecting pathogen colonization of the airway epithelium and localized granulomatous inflammation without extensive tissue necrosis. Another key imaging feature in NTM-LD was the presence of thin-walled cavities (OR=2.556, P=0.002), which are believed to develop secondarily to long-standing bronchiectasis and parenchymal remodeling. These cavities are typically small, multiple, and predominantly located in the upper lobes, especially in the subclavicular and middle lobe regions. This observation is strongly supported by prior HRCT studies demonstrating that thin—walled or even “bubbly”—appearing cavities are more common in NTM-LD than in PTB-LD (Evans et al., 1996). Comparative analysis of underlying comorbidities revealed distinct pathophysiological predispositions: the NTM-LD cohort exhibited heightened susceptibility to autoimmune disorders (OR=7.280, P=0.001), whereas diabetes mellitus predominated in PTB-LD patients (OR=0.310, P=0.002), a dichotomy reflecting pathogen-specific modulation of immune microenvironmental susceptibility. The indolent and chronic progression of NTM infection drives sustained low-grade activation of the host immune system, disrupting the homeostatic immune regulatory network. This persistent immunostimulation compromises immune cell-mediated recognition and tolerance mechanisms toward self-antigens, thereby triggering aberrant autoimmune responses against host tissues and elevating susceptibility to autoimmune disorders (Liu et al., 2023). The elevated prevalence of diabetes mellitus in PTB-LD patients is mechanistically attributed to the hyperglycemic microenvironment serving as a nutritional reservoir that facilitates Mycobacterium tuberculosis proliferation, survival, and dissemination. Moreover, hyperglycemia affects cell metabolism, inhibits the functions of immune cells such as macrophages, reduces the body’s phagocytic and killing ability against Mycobacterium tuberculosis, and increases the risk of infection (Kamei et al., 2023).

The predictive model constructed based on the aforementioned key differences demonstrated favorable diagnostic performance, with a training cohort AUC of 0.923 (95% CI: 0.898-0.948), validation cohort AUC of 0.892 (95% CI: 0.849-0.936), and overall accuracy of 81.5%. It significantly outperforms existing models: Liu et al’s (Liu et al., 2023). chest X-ray DenseNet model (internal/external AUC: 0.86 ± 0.006/0.64 ± 0.017), Hu et al.’s (Hu et al., 2025) radiomics model (AUC 90.2%), and CT-only models—all limited by single-modality reliance, neglect of clinical context (e.g., symptoms like cough with sputum/hemoptysis and comorbidities), inability to capture clinical heterogeneity, and smaller sample sizes. By integrating clinical variables with imaging features, this model overcomes single-modality limitations, achieving superior discriminative power (higher AUC than comparative models), robust calibration (94.7% alignment of calibration curve with ideal reference; Hosmer-Lemeshow test, P=0.176), and confirming clinical net benefit via decision curve analysis-outperforming “treat-alls” or “no-intervention” strategies across threshold probabilities. It provides a reliable tool for precise differentiation between NTM-LD and PTB-LD, aiding in reducing misdiagnoses and optimizing patient management.

Furthermore, the transparent logistic regression-based algorithm employed in our model offers better interpretability for clinical practitioners compared to complex deep learning architectures, facilitating its potential integration into routine clinical decision-making processes. The innovatively developed web-based calculator (accessible at https://dynamic-diagram.shinyapps.io/DynNomapp/) facilitates clinical translation of our model by enabling physicians to obtain real-time personalized risk assessments through the input of eight binary variables. This tool supports clinical decision-making by stratifying patients according to their probability of NTM-LD. Individuals predicted as high-risk should avoid unnecessary anti-tuberculosis therapy (ATT) and instead undergo timely NTM-specific diagnostic testing—such as sputum culture for nontuberculous mycobacteria, species identification, NTM specific PCR, or metagenomic next-generation sequencing (mNGS)—to confirm the diagnosis and guide appropriate management. In contrast, patients classified as low-risk may be considered for early initiation of ATT if clinical and radiological features suggest pulmonary tuberculosis, provided that microbiological workup is concurrently pursued to confirm or exclude PTB-LD. The current version of the application represents a proof-of-concept implementation, designed to demonstrate the feasibility of translating our predictive model into a practical clinical decision-support tool. However, we acknowledge that formal usability testing and integration into routine clinical workflows have not yet been conducted. Future studies should evaluate the tool’s acceptability, efficiency, and impact on clinical decision-making through structured user experience assessments—such as System Usability Scale (SUS) surveys or cognitive task analyses—among frontline healthcare providers. Furthermore, pilot implementation in outpatient respiratory or infectious disease clinics could provide valuable insights into its compatibility with existing electronic health record (EHR) systems and standard diagnostic pathways.

This study has several limitations. First, the single-center, retrospective design may introduce selection bias and limit the generalizability of the findings to broader populations. Second, the relatively modest sample size (n=600) constrains statistical power for comprehensive subgroup analyses, particularly for rare clinical phenotypes. Furthermore, although multiple statistical and validation techniques were employed to minimize overfitting—such as internal bootstrapping and stepwise variable selection—some degree of model overfitting may still persist. To address these limitations, we plan to conduct a prospective, multi-center cohort study to externally validate the model’s performance across diverse healthcare settings and patient populations. Independent validation in larger, geographically representative cohorts will be critical to confirm its robustness, transportability, and clinical utility.

The developed NTM-LD/PTB-LD differentiation model demonstrates significant clinical utility for optimizing diagnostic workflows. The implementation of its web-based tool (accessible at https://dynamic-diagram.shinyapps.io/DynNomapp/) could transform current diagnostic paradigms by reducing reliance on invasive procedures, while providing real-time decision-making support for precision antimicrobial therapy. Future efforts will focus on clinical validation and global dissemination of this tool to enhance tuberculosis management in resource-variable settings.