Flail chest is the most severe type of chest trauma, is found in approximately 4% of patients with rib fractures and is defined as at least 3 or 4 consecutive rib fractures in at least two places causing paradoxical movement of the chest wall (1, 2). It has been described in 30% of patients with significant chest trauma requiring intensive care, and flail chest has an adverse impact on their respiratory function and causes several associated complications, such as pneumonia and acute respiratory distress syndrome (ARDS) (3).

Pneumonia is a common, severe, preventable complication in patients with flail chest, with an estimated incidence rate ranging from 21% to 33.5% (4–6). Moreover, pneumonia morbidity is an important indicator of care quality for patients with rib fractures, including risks of prolonged mechanical ventilation, increased mortality, and poor long-term outcomes (5). There are some challenges in managing pneumonia (7). Overestimating the likelihood of pneumonia can cause inappropriate use of antibiotics, leading to the emergence of multidrug-resistant and invasive fungal infections. Conversely, underestimating the likelihood of pneumonia leads to undertreatment and increased mortality from severe nosocomial infections (8, 9). Thus, to avoid possible adverse disease progression and unfavourable outcomes, it is essential that physicians accurately identify patients at high risk of pneumonia at an early stage and tailor individualised preventive treatment (e.g., enhanced monitoring, lung hygiene, pain management) (10).

The Clinical Pulmonary Infection Score (CPIS) has been proposed and used clinically for decades and has shown some shortcomings in guiding the management of pneumonia (8). Consequently, many studies have been conducted to design and construct more accurate and stable risk scores, such as the Scoring System for Pneumonia Risk in Pulmonary Contusion Patients (11) and “RibScore”, a novel radiographic score based on fracture patterns, have been performed (12). These prediction models have been developed based on a generalised linear approach; in addition, the “RibScore” only considers anatomical factors (13). However, many risk factors may exhibit a nonlinear relationship with the outcome and may not fully apply to the traditional linear regression prediction model (14).

Machine learning (ML), a branch of artificial intelligence methods, can develop models from medical data to make clinical decisions and assist doctors in their routine work (15). In previous studies, several classical machine learning algorithms have been tested to predict the risk of pneumonia, including the prediction of stroke-associated pneumonia (16), ventilator-associated pneumonia (17), and postoperative pneumonia (18). However, studies in which researchers use machine learning to predict pneumonia risk in flail chest patients are rare. Even though these algorithms have not yet become widely accepted and used in clinical decision-making, they have tremendous potential in medicine (19).

In this study, we aimed to develop an advanced machine learning model, assess how well it predicts the risk of pneumonia in flail chest patients, and provide new approaches for individualized analyses of pneumonia risk factors in hospitalized patients with flail chest.

As a severe chest trauma, flail chest not only affects the stability of the chest wall but also poses a threat to circulation and breathing, which can directly affect the clinical course and outcome of the patient (30). Furthermore, multiple rib fractures and pulmonary contusions are significant trauma factors contributing to the development of pneumonia (31). Therefore, early detection of pneumonia is critical for timely interventions to improve clinical outcomes and reduce costs (32). In our study, we evaluated the performance of seven models for the occurrence of pneumonia in flail chest using clinical data from the first 48 h. The XGBoost model with 32 clinical features showed the best prediction ability, and we had some plans to avoid overfitting, such as feature selection by Fisher score and 3-fold cross-validation in the training set.

Predictive modelling to aid clinical decision-making is not a new concept. The Scoring System for Pneumonia Risk in Pulmonary Contusion Patients, the “RibScore”, a novel radiographic score based on fracture patterns, and the Clinical Pulmonary Infection Score (CPIS) are all examples of current operational scoring tools (8, 11, 12). The CPIS tool has shown moderate reliability and accuracy in diagnosing nosocomial pneumonia (33, 34). The CPIS, however, is based on radiographic and laboratory values following symptom onset, as with other tools (22). Thus, CPIS may assist clinicians in narrowing down the identification of patients with pneumonia, but treatment becomes delayed rather than early prevention.

Artificial intelligence and electronic medical records have enabled ML algorithms to be more widely used in individualized medicine to support clinical decisions (35). XGBoost models have shown advantages in predicting pneumonia due to their impressive predictive accuracy and ease of use (14). Chen et al. (36). based on the XGBoost algorithm, developed a risk prediction model for postoperative pneumonia in patients after liver transplantation. According to Li (37), XGBoost was used to build a machine learning model for predicting stroke-associated pneumonia with a 0.84 model evaluation precision. In our study, the XGBoost model has a sensitivity of 88.5%. In contrast, the specificity is 80.0%, meaning that only 11.5% of actual patients are not identified as high-risk (false-negative rate), yet 20% of the patients who do not have pneumonia are incorrectly labelled as high-risk patients (false-positive rate). This level of exposure to unnecessary interventions may be acceptable for a safe, inexpensive, and well-tolerated intervention (e.g., enhanced surveillance or pulmonary hygiene protocols) to prevent pneumonia in high-risk patients. In contrast, high sensitivity is essential in clinical practice because of the consequences of misclassification of actual patients with pneumonia. We can combine our models’ and doctors’ opinions to further improve system performance to overcome the effects of false negatives and positives.

Given these results, it appears that not all low-risk-level patients are pneumonia free. We report low-risk groups in several patients with pneumonia (Supplementary Table S1). This result reflects the incomplete clinical presentations initially observed in some patients. In the early admission stage, false-negative patients with clear consciousness and stable vital signs did not require emergency surgery or transfusion therapy, and the model identified them as a low pneumonia risk group. However, all three patients had injuries other than chest trauma, such as lower limb fractures and abdominal organ injuries, and their subsequent treatment and long-term immobilization increased the risk of pneumonia. On the other hand, one false-positive patient who was severely injured was predicted to have a high risk of pneumonia by the XGBoost model. The patient died during the early stages of hospitalization, and the development of pneumonia may not be effectively observed during the disease. Importantly, for patients with incorrect predictions in this model, we can incorporate dynamic clinical features (e.g., 24 and 48 h of admission) into the model for multiple predictions. Three or more clinicians will also combine their clinical experience to determine the likelihood of pneumonia in patients with flail chest (38). Clinicians could identify them effectively by incorporating clinical knowledge.

In addition, a predictive model for a disease needs to be interpretable, and this interpretability can promote understanding and acceptance of the model's predictions by physicians (28). Therefore, we used SHAP values to estimate how important predictor variables were to the model to gain insight into model interpretability. By taking the average of the absolute value of each feature's SHAP value and observing the importance to the overall model, we found that the more crucial baseline variables in the pneumonia prediction model were systolic blood pressure, pH value, blood transfusion, and Injury Severity Score (ISS). The first three indicators reflect the internal environment (39). The Injury Severity Score (ISS) is a common anatomical-based trauma score that can effectively assess the injury's severity and predict clinical outcomes such as the likelihood of survival and ICU length of stay (40, 41). However, it depends exclusively on anatomical factors of injuries, not synthesizing the mechanism of injury and physiological factors (42).

ML models have demonstrated outstanding performance in predicting diseases and clinical conditions, and can be used to guide treatment decisions and interventions (35). For example, a machine learning model can generate a probability for each patient based on their characteristics. We observed that more severe trauma problems, such as higher ISS, rib fracture number, or intubation, increased the risk of pneumonia, consistent with clinical experience and previous data (5, 31). Consistent with previous studies, a more unstable physiological environment, e.g., lower pH value, systolic blood pressure, and haemoglobin, also increased the risk of pneumonia (39, 40, 43). As a result, the interpretable machine learning model built demonstrates good prediction results.

The study may have several limitations. First, it was a single-centre study with a limited sample size. Given that only four percent of patients with rib fractures develop flail chest, the sample size of flail chest patients in one medical centre was small (1). To mitigate the small sample size, we performed cross-validation (k = 3) of every model. However, selection bias was inevitable, and the model's efficacy for predicting pneumonia should be verified in future studies with larger sample sizes. Second, our study was conducted retrospectively. We can only suggest associated and correlated factors, not identify the main factors contributing to pneumonia. Furthermore, there is a possibility that our findings could be biased due to the long duration of patient enrolment. Third, whereas the missing numerical variables were imputed with weighted k-nearest neighbours, essential variables such as CRP and PH value had some missing values. Additionally, the accuracy may be improved by adding more variables currently not collected, such as comorbidity (40) (e.g., COPD, type 2 diabetes mellitus) and other predictors, such as smoking, alcohol usage, aspiration, and prehospital measures (44, 45). Fourth, we used variables within 48 h as independent variables instead of time-varying variables to predict the occurrence of pneumonia that cannot reflect the dynamic changes in of hemodynamic and metabolic variables in trauma patients. Finally, further external validation is needed. Fortunately, our team will continue to collect data on a prospective multicentre cohort study to refine the model.

We successfully established seven ML models to predict the risk of pneumonia during hospitalization in flail chest patients. The XGBoost model has shown better performance than the LR, RF, SVM, ANN, GBDT, and NB models. We anticipate that it is a convenient risk stratification tool that clinicians can use to identify individualized treatment options for patients with multiple rib fractures. To our knowledge, this is the first machine learning based study to provide a novel approach for predicting pneumonia in flail chest patients. However, further external validation is required to test the generalization of our model.