Assessing the risk of concurrent mycoplasma pneumoniae pneumonia in children with tracheobronchial tuberculosis: retrospective study

Introduction

Tuberculosis (TB), recognized as a Category B notifiable infectious disease in China, is preventable and manageable (MacNeil et al., 2020). Annually, TB affects over 2 million children globally, leading to the death of over 200 children every day (Shakoor & Mir, 2022). The non-specific manifestations of pediatric TB frequently lead to its misdiagnosis or underdiagnosis, suggesting that its actual prevalence could surpass current estimates (Marangu & Zar, 2019; van Heerden et al., 2021). Despite a reported decrease in TB incidence within China, the country’s large population presents a persistent, significant ongoing health challenge. Consequently, China remains among the nations with the highest TB burden globally (Pang et al., 2019).

Tracheobronchial tuberculosis (TBTB) explicitly targets the mucosa, submucosa, smooth muscle, cartilage, and epithelium of the trachea and bronchi, classifying it as a distinct TB subtype that affects the lower respiratory tract (Liu et al., 2022). Pediatric patients with TBTB present with non-specific symptoms, such as fever, cough, and wheezing. However, these can lead to critical complications, such as airway constriction, occlusion, sensitivity, and pulmonary atelectasis (Zhuang et al., 2023), highlighting the need for early diagnosis and prompt intervention to enhance patient prognosis and quality of life.

Conversely, lobar pneumonia, primarily caused by Streptococcus pneumoniae, is more frequently observed in young adults. This condition is characterized by the abrupt onset of high fever, chills, chest discomfort, cough, and dyspnea (Zhuang et al., 2023). The primary pathology initiates in the alveoli and rapidly spreads to affect one or several lobes. Mycoplasma pneumoniae pneumonia constitutes 10% to 40% of pediatric hospitalizations, exhibiting cyclical outbreaks every 3 to 7 years. Each outbreak can persist for a year, marking significant periodic public health challenges (Rothberg, 2022).

Distinguishing between TBTB and lobar pneumonia presents a diagnostic challenge, as both conditions frequently start with non-specific symptoms such as fever, cough, and dyspnea (Kim et al., 2020). From a radiological perspective, both diseases show signs of inflammatory lung changes, including solid masses or pulmonary infiltrates.

In light of the symptomatic overlap between TBTB and lobar pneumonia, this study introduces an innovative machine learning-based risk prediction model to enhance diagnostic precision. This model, particularly through the XGBoost algorithm, represents a significant advancement in clinical diagnostics, substantially reducing the risk of misdiagnosis. It provides a sophisticated tool for clinicians to differentiate between these conditions more accurately, ensuring targeted and effective patient treatment. Furthermore, this approach enhances diagnostic accuracy and contributes to a deeper understanding of the pathophysiology of these diseases, thereby establishing a new benchmark in their management.

Materials and Methods

Sample screening

Out of the initial cohort, 212 samples fulfilled the inclusion, exclusion, and diagnostic criteria. Of these, 170 cases identified with lobar pneumonia formed the “pneumonia group,” and 42 cases diagnosed with TBTB alongside Mycoplasma pneumoniae pneumonia constituted the “combined group.” A flowchart was constructed to detail the sample selection process (Fig. 1).

Results

Signature identification using XGBoost

The XGBoost algorithm was applied to the patient data to identify significant variable associations (Fig. 2A). Subsequently, those with <1% contribution were excluded, leaving us with a refined list of eight critical signatures: geographical area, duration of cough, history of fever, cough history, SIRI, WBC count, presence of wet rhonchi, and SII.

Evaluation of model performance

Upon concluding the study, each model’s performance was compared by examining the ROC and PR prediction curves. The AUC values were as follows: XGBoost: 0.992, decision tree: 0.943, and logistic regression: 0.851 (Figs. 3A and 3B, Table 4). The DeLong test affirmed the XGBoost model’s superior predictive capabilities over the decision tree and logistic regression models (p < 0.01, Table 5), ranking them as XGBoost > decision tree > logistic regression. Furthermore, the Akaike Information Criterion (AIC) assessment revealed the XGBoost model (AIC = 43.226) had a better fit compared to the Decision Tree (87.259) and logistic regression models (135.231) (Fig. 3C). The XGBoost model demonstrated the highest average net benefit of 0.764227, outperforming the decision tree (0.708071) and logistic regression models (0.625883). Consequently, the XGBoost model is recommended for diagnosing pediatric patients with TBTB coexisting with Mycoplasma pneumoniae pneumonia (Table S1).

Table 4:

Detailed parameters of the ROC curve.

Variables	AUC	CI	Sensitivity	Specificity	Jordon Index
XGBoost	0.996	[0.992–1.000]	97.64%	97.61%	95.26%
Decision tree	0.943	[0.905–0.982]	91.77%	95.24%	87.00%
Logistics	0.851	[0.761–0.941]	87.65%	83.33%	70.98%

DOI: 10.7717/peerj.17164/table-4

Table 5:

DeLong test comparing the effectiveness of the three models.

	z	p	Difference between AUC	Difference between SD	95% CI
					Low	High
XGBoost-decision tree	2.787	0.005	0.053	0.148	0.016	0.091
XGBoost-logistics	3.156	0.002	0.145	0.219	0.055	0.236
Decision trees-logistics	2.019	0.044	0.092	0.255	0.003	0.181

DOI: 10.7717/peerj.17164/table-5

Note:

z-score, test statistic in DeLong test; SD, standard deviation; AUC, area under the curve; CI, confidence interval.

XGBoost model validation

The original dataset was divided into validation groups 1 and 2, following a 7:3 distribution to validate the XGBoost model. Analysis of the ROC and PR curves revealed an AUC of 0.997 for both groups. The AIC values were 21.571 for the validation one group and 13.612 for the validation teo group, indicating a robust model fit. The average net benefits within the XGBoost model were 0.804 for validation 1 and 0.796 for validation two, demonstrating the model’s predictive accuracy (Figs. 4A and 4B). Furthermore, a nomogram was constructed based on eight characteristic factors. Using the profile of a hypothetical pediatric patient (city, fever history, cough history, cough duration of 15 d, absence of wet rhonchi, WBC count of 15.99, SII of 3,479.47, and SIRI of 9.01), a morbidity prediction score of (0 + 40 + 100 + 76 + 0 + 26 + 30 + 15) 287 points, corresponding to a 72% morbidity rate for this child (Fig. 4C).

Discussion

Mycoplasma pneumoniae pneumonia represents a significant respiratory infection in children, characterized by symptoms such as cough and fever, potentially leading to long-term pulmonary damage (Wang et al., 2023; Zhu et al., 2023). Although imaging diagnostics are crucial, their frequent application in pediatric care raises concerns about elevated radiation exposure risks (Kersemans et al., 2011), highlighting the importance of weighing the diagnostic advantages against potential health risks.

Emerging technological innovations have marked the beginning of a new era in diagnostics through machine learning (Nature Medicine, 2023; Nature, 2023). Machine learning empowers physicians to expedite and refine diagnostic accuracy by analyzing extensive biomedical datasets. Algorithms in machine learning pave the way for the development of automated and unbiased diagnostic models, potentially diminishing the reliance on imaging modalities and subsequently decreasing the radiation exposure risk for children (Garriga et al., 2022). Complex, high-dimensional biomedical data could be comprehensively analyzed through the proactive adoption of these algorithms. Such analysis improves diagnostic precision and offers critical insights into monitoring therapeutic efficacy (Swanson et al., 2023). As an invaluable tool to truncate diagnostic latency, machine learning provides significant potential for the early detection and intervention of diseases, potentially saving numerous lives.

The landscape of modern medical research and clinical diagnostics has shifted considerably from traditional paradigms to the integration of data analytics and machine learning. Within this evolving field, models such as XGBoost, decision tree, and logistic regression have gained prominence, showcasing their pivotal significance in clinical medicine (Kalyani et al., 2023). This study aimed to ascertain the diagnostic efficacy of these models in pediatric patients with coexisting TBTB and Mycoplasma pneumoniae pneumonia.

Based on ROC and PR curves, our assessment demonstrated that the XGBoost model, with an AUC value of 0.992, surpassed its counterparts in diagnosing this specific pediatric condition. The DeLong test further accentuated XGBoost’s diagnostic superiority. The effectiveness ranking was XGBoost > decision tree > logistic regression. The AIC metric endorsed XGBoost’s superior model fit with a score of 43.226. Furthermore, DCA indicated XGBoost’s commendable average net benefit of 0.764227, the highest among all evaluated models. These findings suggest XGBoost as the preferred choice for diagnosing children with concurrent Mycoplasma pneumoniae pneumonia and TBTB, promising superior diagnostic accuracy and reliability.

The variability in the performance of different models may be ascribed to factors such as the models’ complexity, their ability to extract relevant signatures, control overfitting, the quality and volume of data, and the tuning of parameters. XGBoost stands out for its adaptability and potent predictive capabilities when handling complex clinical datasets. Although decision tree models exhibit reasonable efficacy, they potentially lag behind XGBoost in certain aspects due to their inherent simplicity and absence of a sophisticated signature selection mechanism, making logistic regression models less suitable for complex diagnostic challenges due to their foundational simplicity and lack of a robust signature selection process.

Supporting these findings, Hu et al. (2022) demonstrated that XGBoost outperformed LASSO in internal validation, though LASSO showed a marginal advantage in external validation. Furthermore, XGBoost was superior to logistic regression models in predicting outcomes for hepatocellular and gallbladder carcinomas (Zhang et al., 2021).

Selecting the appropriate machine learning model selection is paramount in clinical diagnostics and prognostics. For instance, a study on blood pressure variation during hemodialysis analyzed various models without determining a definitive best-performing model (Huang et al., 2020), suggesting that the model’s efficacy may depend on specific dataset characteristics and task specifications. Similarly, research involving patients with COVID-19 assessed the effectiveness of multiple models without furnishing comparative outcomes, indicating that model efficacies could vary across distinct disease domains and objectives (Hong et al., 2022). Another comparative study between XGBoost and logistic regression for predicting ICU mortality in patients with rheumatic heart disease underscored the importance of comparing machine learning models to enhance clinical diagnostics (Xu et al., 2022). Therefore, selecting an appropriate model that aligns with the specific clinical challenge and dataset is crucial (Adamson & Welch, 2019).

Our investigation identified eight significant variables intrinsically linked to pediatric TBTB coexistent with Mycoplasma pneumoniae pneumonia utilizing the XGBoost model. The interplay of these variables aligns logically within biological and clinical science realms.

Factors such as limited healthcare infrastructure and inadequate sanitation in rural settings may predispose children to heightened exposure to Mycobacterium tuberculosis and other respiratory pathogens (Pinto et al., 2023). An extended duration of cough may indicate the persistent and severe nature of these respiratory infections. A prolonged cough, without a history of fever, could serve as an early indicator of TBTB, while fewer cough episodes might suggest heightened sensitivity or susceptibility to respiratory pathogens among children (Kim et al., 2022). The SIRI and leukocyte counts potentially reflect the body’s immunological response to infection, signaling acute inflammatory reactions. A decrease in the incidence of wet rales suggests obstructions in respiratory mucus due to pathogenic activity, aligning with the pathogenesis of Mycoplasma pneumoniae pneumonia (Venkatappa et al., 2023). An increased SII supports the presence of significant inflammation and infection, corroborating the severity of the concurrent conditions (VanValkenburg et al., 2022). Analyzing these variables and their association with the disease provides a deeper understanding of the pathophysiology and clinical presentation of TBTB in conjunction with Mycoplasma pneumoniae pneumonia. However, supplemental research endeavors are warranted to elucidate the complex interactions among these variables and their impact on the disease and to probe the underlying biological processes.

A validation phase was incorporated for the XGBoost model into our study for further introspection. This process involved dividing the sample into two validation groups, 1 and 2, using a 7:3 ratio. The outcomes showcased exceptionally high AUC values of 0.997 for both groups, confirming the model’s high predictive accuracy and reliability across varying datasets. The AIC scores and the calculations of average net benefit further validated the model’s efficacy. Furthermore, a nomogram constructed from eight characteristic variables demonstrated its practical application by estimating a 72% morbidity rate for a hypothetical patient case. This extension of our initial analysis emphasizes the potential of machine learning in enhancing diagnostic precision, underscoring the significance of our methodological contributions to the discipline.

The study acknowledges several limitations that merit consideration. The relatively small sample size may not have fully represented the diversity of the target population. The retrospective design of the study may have introduced selection bias. Although we evaluated three machine learning models, numerous other models could apply to clinical diagnostics. Furthermore, excluding certain risk factors and possible confounders could have influenced the study’s results and interpretations. Expanding the sample size, employing a prospective design, exploring a broader selection of models, and encompassing a comprehensive set of risk factors are recommended for enhanced precision and reliability.

Conclusions

This study demonstrates the exceptional diagnostic precision of the XGBoost model in identifying cases of pediatric TBTB with Mycoplasma pneumoniae pneumonia and identifying eight critical factors that offer insights into the disease’s pathogenesis. The observation highlights the significant potential of machine learning technologies, particularly the XGBoost model, in refining clinical diagnostic processes and their prospective incorporation into medical practice.

Supplemental Information