Prediction model for prolonged fever in patients with Mycoplasma pneumoniae pneumonia: a retrospective study of 716 pediatric patients

Mycoplasma pneumoniae (MP) infections are generally mild and self-limiting. However, patients of every age can develop a severe and progressive course during treatment with appropriate antibiotic therapy [1]. The underlying mechanisms are unclear, but a direct microbe effect, macrolide resistance, and excessive immunological response of the host are commonly suggested [2, 3]. Macrolide antibiotics have been generally preferred as the first-choice agents for MP infections because secondary antibiotics such as tetracyclines and fluoroquinolones are not recommended because of the risk of severe adverse events, especially in pediatric patients.

Macrolide resistance rates have risen throughout the world and vary across countries[4‐7]. Although macrolides could be continued in cases of mild to moderate infections irrespective of their resistance, replacement by alternative antibiotics or additional corticosteroids have been shown to improve radiological abnormalities and clinical symptoms[8, 9]. Additionally, the severity of the disease is partially related to the degree to which the host immune response reacts to infection. The concept of immune-mediated lung disease provides a basis for consideration of immunomodulatory therapy in addition to conventional antimicrobial therapies for the management of MP infections [10].

The appropriate time for alternative treatment is not clarified, but it still depends on the physician’s decision. Alternative treatments are delayed on some occasions owing to concerns regarding toxicities and adverse effects of secondary antibiotics or the possibility of blurred diagnosis caused by corticosteroids, leading to aggravation of the clinical course. Protracted courses of fever or worsening respiratory exertion despite treatment with macrolides are reported to complicate atelectasis, parapneumonic effusion, bronchiolitis obliterans, necrotizing pneumonitis, pulmonary abscess, and systemic inflammatory response syndrome [2, 11‐14]. At the initiation phase of macrolide therapy, physicians find it difficult to predict patients with a prolonged or severe clinical course. Previous studies have suggested individual cut-off values for inflammatory markers to differentiate between the patients with or without clinical and radiological progression after macrolide therapy for 7 days or longer [8, 15‐20]. Identifying patients who are expected to undergo a prolonged or severe clinical course would help in providing them with timely secondary treatment and mitigating their clinical course [8, 17, 21, 22].

This study aims to identify the predictive factors for prolonged fever in patients with MP pneumonia with readily accessible clinical, laboratory, and radiological data and to develop a predictive model for these patients in whom timely initiation of secondary treatment options should be considered.

To prevent the progression of MP pneumonia resulting in severe and prolonged clinical course, early recognition and timely treatment is important for patients who display clinical and radiological aggravation during macrolide therapy [8, 16, 23, 24]. To our knowledge, this is the first study to demonstrate a prediction model for refractory MP pneumonia based on readily accessible sequential fever data in addition to clinical, laboratory, and radiologic variables at admission. For prediction of RMPP-3, a conventional logistic model using only radiologic grouping showed increased sensitivity (63.3%) than the model using laboratory values, including CRP and LDH. Adding laboratory values in the prediction model using radiologic grouping did not meaningfully contribute to an increase in sensitivity (64.6%). For the prediction of RMPP-5, laboratory values and radiologic grouping showed lower sensitivities ranging from 12.9 to 16.1%. However, prediction models using the predefined fever profiles showed significantly increased sensitivity for predicting RMPP-5, and neural network models using 12 sequential fever data showed a greatly increased sensitivity of 64.5%. Predicting high-risk patients for refractory MP pneumonia would enable physicians to calibrate their expectations of progression in these patients and to provide earlier alternative treatment.

Several studies have tried to identify predictors for refractory MP pneumonia and have suggested individual cut-off values of inflammatory markers, namely CRP, LDH, and ferritin or cytokines, such as IL-6, IL-8, IL-10, IL-18, and interferon-gamma [15‐17, 19, 23]. However, the application of these findings in clinical practice is limited by lower prediction power or accessibility of the tests. Although it is plausible that increased inflammatory cytokines are related to the severity of MP pneumonia, serum cytokine assays are mostly limited for research purposes and are not routinely measured. Bronchoscopy and bronchoalveolar lavage studies are useful tools not only for identifying the causative organism but also for the removal of mucosal plugs in severe pneumonia, but are generally performed for a small proportion of MP pneumonia cases. The requirement of sedation, the necessity of special equipment, and the need for an experienced bronchoscopist limit their accessibility.

A study using CRP value of 16.5 mg/L as the cut-off value showed a sensitivity of 74.7% and a specificity of 77.2% for predicting refractory MP pneumonia [18]. However, our prediction model created using the CRP level of the training cohort showed a sensitivity of 32.9% for the prediction of RMPP-3 in the test cohort even when it was combined with the LDH level (Table 5). For the prediction of RMPP-5, our prediction model using the CRP level showed a lower sensitivity of 12.9% even when used in combination with ALC and LDH levels (Table 6). Previous prediction models, created without validation, are inevitably vulnerable to model overfitting, resulting from institutional selection bias, which limits their clinical use. Therefore, a reasonable prediction model should undergo internal validation by a separate test cohort or external validation using data from another institution. Thus, it is understandable that previously identified laboratory markers such as CRP and LDH showed lower sensitivities (below 30%) for predicting RMPP-3 and RMPP-5 in our cohorts (Tables 5, 6). Such low sensitivities limit their clinical application for the timely detection of refractory MP pneumonia. To overcome such bias, our 716 enrolled patients were divided into training and test datasets for internal validation, which prevented overfitting and created a reasonable prediction model.

For prediction of RMPP-3, according to a previous study, initial radiologic grouping was the most prominent predictor [25]. While the underlying mechanisms are still unclear, the pattern of pulmonary lesions in MP infection is reported to be influenced by the characteristics of host cell-mediated immunity [26, 27]. Thus, radiological evidence of lung involvement is consistent with the strong host immune response in RMPP.

Both initial laboratory values and radiologic grouping showed limited prediction power for the prediction of RMPP-5. However, we tried to predict RMPP using initially available data and focused on the fever data during the initial 48-h period. Inflammatory cytokines involved in the immunopathogenesis of MP infection are reported to be increased in RMPP [3, 20, 28]. Since these cytokines act as endogenous pyrogens that play a pivotal role in inducing fever response, their levels are associated with core body temperature [29]. Although initial single time-point data were limited for predicting RMPP-5, the prediction model using predefined fever profiles showed a two-fold increase in sensitivity (16.1% to 32.3%), and the DNN model using all 12 sequential fever data within 48 h showed a four-fold increase in sensitivity (64.5%) for predicting RMPP-5. Theoretically, DNN is a black-box approach, and the causes of superior prediction power of the DNN model cannot be identified. However, the greatly increased sensitivity for predicting RMPP-5 with the DNN model using only the initial 48-h fever data is noteworthy.

The major purpose of our grouping that included RMPP-3 and RMPP-5 was to evaluate the prediction power of the statistical model at two separate time points, to compare their prediction power, and to infer the causes for the difference. The prediction power of our statistical models for the later event (RMPP-5) was considerably lower than that for the early event (RMPP-3). Evaluating the model prediction power at separate time points enabled us to trace the changing trends in the variables of the prediction models at different time points. We identified fever profiles and radiologic grading as the most effective predictors that have superior prediction power for the ‘later event’ (RMPP-5).

The main limitation of our study is its retrospective design based on a limited number of inpatients from a single center, which might have introduced a selection bias. However, our prediction models underwent internal validation. Prediction models were created only from the data in the training cohort, and their prediction power was estimated in the test cohort, which was not used for model development. Nevertheless, external validation of our model in a prospective, large-scale cohort is needed for validating our results. Second, a possibility of under-diagnosis and over-diagnosis in MPP exists because of false negative IgM antibodies in the early stage or persistent IgM antibodies in convalescent patients with recent infection. We attempted to minimize these misdiagnoses through our strict exclusion criteria. Third, prediction models were not developed using tests, namely cytokines or FOB, which were reported to be significant. We especially focused on the accessibility of the tests, and those tests were not considered useful in usual clinical practice. Lastly, data on macrolide resistance were not included. Although febrile days during macrolide administration were reported to be greater in macrolide-resistant patients (3.5–4.0 days vs. 1.0–1.5 days) [9, 30], prolonged fever in RMPP patients may not imply macrolide resistance because fever might have resolved spontaneously in some macrolide-resistant patients. The clinical efficacy of macrolide for treating MP infection may not only reflect its direct antimicrobial activity but also reflect its anti-inflammatory effects [31].

Development of tests based on data obtained from routine examination of vital signs and its integration into the clinical workflow can be more effective than utilizing new tests that are less verified and less accessible. Further studies utilizing such potential data are needed for improving the prediction power.

In summary, our study showed that for prediction of RMPP-3, a conventional logistic model using only radiologic grouping showed a favorable predictive power than the model using initial laboratory values. In contrast, RMPP-5 could not be effectively predicted using the initial laboratory and radiologic data, which were previously reported to be significantly predictive. However, the prediction models using predefined fever profiles showed a two-fold increase in sensitivity (16.1–32.3%), and the DNN model using all 12 sequential fever data within 48 h showed a four-fold increase in sensitivity (64.5%). Further studies using more advanced mathematical models based on easily accessible large-sized clinical data are anticipated to be helpful for predicting RMPP.