This study was conducted to develop a machine learning-based model for predicting the risk of lung infections following chemotherapy in lung cancer patients. The retrospective study included a cohort of 502 lung cancer patients who had undergone chemotherapy, aged 18 years and above, and had completed at least one cycle of treatment. Data encompassing demographic details, medical history, chemotherapy specifics, and blood test results were extracted from the hospital’s electronic medical record system. The SMOTE algorithm is used to solve the category imbalance problem. The Boruta algorithm and LASSO regression performed feature screening to identify the features most associated with the risk of lung infection. Subsequently, a range of ML models, including LR, RF, GNB, MLP, SVM, and KNN, were developed and refined by applying a 10-fold cross-validation methodology. The performance of these models was assessed using various metrics, including accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, Kappa score, AUC, calibration curve, calibration curves, Clinical Impact Curve and confusion matrix. To enhance the transparency and interpretability of the model, the SHAP method was employed to interpret the predicted results and elucidate the impact of each feature on the predictions, thereby offering a practical reference for clinicians. Figure 1 explains the overall workflow of the proposed system more clearly.

This retrospective study examined data from lung cancer patients at The Central Hospital of Shaoyang between January 2020 and December 2023. The study included adult patients aged 18 years and older who had not experienced lung infections within a week before receiving chemotherapy. Patient records with missing or abnormal data were excluded to maintain data quality. The study’s rigorous inclusion and exclusion criteria aimed to ensure the completeness and reliability of the information on included cases, thus providing a high-quality database for evaluating the risk of lung infections in lung cancer patients after chemotherapy. Inclusion criteria: (i) adult patients aged ≥18 years, (ii) patients diagnosed with lung cancer and treated with chemotherapy, (iii) patients who did not have any lung infection before chemotherapy, and (iv) patients with complete clinical information; Exclusion criteria: (i) patients with mental illness or intellectual disability, (ii) patients with missing or abnormal data, and (iii) exclusion of patients with a combination of other tumors.

The study encompassed 36 predictors related to demographic factors (gender, age), lifestyle habits (history of alcohol consumption, history of smoking), medical history (history of diabetes, history of hypertension, history of coronary heart disease), physical characteristics (BMI), disease severity (stage at diagnosis, histologic features, presence or absence of pleural effusion), treatment information (cycles of chemotherapy, number of hospitalizations), and laboratory values (leukocytes, erythrocytes, hemoglobin, platelets, percentage of neutrophils, percentage of lymphocytes, percentage of monocytes, NLR, NMR, neutrophil-platelet count ratio (NPR), indirect bilirubin, alanine aminotransferase, glutamine aminotransferase, total bilirubin, direct bilirubin, total protein, albumin, globulin, white globule ratio, urea, creatinine, uric acid, and CEA). Of these, gender, age, history of alcohol consumption, history of smoking, history of diabetes mellitus, history of hypertension, history of coronary artery disease, BMI, tumor typing, cycles of chemotherapy, number of hospitalizations, and the presence or absence of pleural effusions were the data before the last chemotherapy session. The other laboratory data were obtained after the last chemotherapy. A brief description of the study variables is given in Table 1 .

The diagnostic criteria for pulmonary infection in patients with lung cancer following chemotherapy encompass a body temperature exceeding 38°C, the presence of clinical symptoms indicative of pulmonary infection (e.g., cough and expectoration), the identification of moist rales in the lungs, and the visualization of a distinct infectious focus on CT imaging. Should a lung cancer patient meet at least three of these criteria within 14 days post-operation, a diagnosis of post-chemotherapy lung infection is warranted.

The SHAP is a technique utilized to interpret predictions generated by ML models, particularly those that are intricate and incorporate numerous features (24). The fundamental principle underlying this method involves the computation of the incremental impact of individual features on the model’s output, enabling interpretation of the model’s behavior at both a global and local scale. This is achieved through the development of an additive explanatory model that considers all features as contributors, thereby facilitating the calculation of the average incremental impact of each feature across all feasible feature combinations to derive a SHAP value for each feature, which provides both global and local interpretations, helping to understand which features are the main influences on model predictions, as well as the predictions of individual samples—factors, as well as the prediction results for a single sample (25).

All data analyses in this study were performed using SPSS (17.0), R language (version 4.3.2), Matlab (version R2021a), and Python (version 3.7). The initial analysis of the data set involved the application of descriptive statistics. Data points adhering to a normal distribution were represented as mean ± standard deviation, while those deviating from normal distribution were represented as median (quartiles). Subsequently, the independent samples t-test was employed to compare two groups with normally distributed data. In contrast, the Mann-Whitney U test was utilized to compare two groups with non-normally distributed data. For count data, frequencies and percentages were used to characterize group variances, while the chi-square test or Fisher’s exact probability method was employed to assess inter-group discrepancies. We solved the problem of sample imbalance by oversampling a small number of classes and thereby solving the sample imbalance problem through the SMOTE algorithm based on Matlab software. To construct the predictive model, the dataset was partitioned randomly into a training subset comprising 70% of the total data and a test subset comprising 30%. Subsequently, six ML algorithms were employed to train the model using the training subset data. During the model training process, a 10-fold cross-validation method is used to optimize the model parameters and prevent the occurrence of overfitting phenomenon. LASSO regression analysis was conducted utilizing the glmnet package [4.1.7] in R to analyze cleaned data and derive coefficient values of variables, logarithmic values of lambda, and regularized values of L1, followed by data visualization. The Boruta algorithm was implemented using Boruta 8.0.0 [4.1.7] in R. Interpretability analysis was carried out using the Python libraries shap=0.43.0. Statistical significance levels were established at P<0.05.

This study assembled a cohort of 502 lung cancer patients who did not have lung infections before undergoing chemotherapy. The median age of the patients was 65 years (range: 58-71 years), with 404 (80.48%) being male and 98 (19.52%) being female. We used the SMOTE algorithm for data imbalance. The original data of 502 cases contained 404 non-infected cases, 98 infected cases, and 19.52% of infected cases, and the processed data of 808 cases contained 404 non-infected cases, 404 infected cases, and 50.00% of infected cases. A comparison of baseline characteristics between the two groups revealed statistically significant differences in chemotherapy cycles, hospitalizations, WBC, pulmonary embolism, Gender, CREA, Histology, alcohol consumption, smoke, CHD, NEUT, LYM, NMR, NPR, IBIL, TBIL, and NLR (P < 0.05), as shown in Table 2 .

A total of 808 patients undergoing chemotherapy for lung cancer after data imbalance were divided into a training group consisting of 565 patients and a test group consisting of 243 patients, following a ratio of 7:3. Statistical analysis revealed no significant differences between the two groups ( Table 3 ). Utilizing the Boruta algorithm, an extension of the RF algorithm, enabled the identification of the actual feature set by accurately estimating the importance of each feature. The Boruta algorithm identified 35 key factors, including Drink, Smoke, Chemotherapy cycles, Hospitalizations, PE, NEUT, LYM, MONO, NLR, and NMR, etc ( Figure 2A ). In contrast, LASSO regression serves as a compression estimation method that accomplishes variable selection and complexity adjustment through the formulation of an optimization objective function incorporating penalty terms. In this study, LASSO regression was utilized to identify characteristic factors such as Gender, Drink, Smoke, Chemotherapy cycles, PE, NEUT, NLR, NMR, and AST ( Figures 2B, C ). Through a comparative analysis of the outcomes obtained from LASSO regression and Boruta algorithm screening, we identified a common subset of feature variables selected by both methods. These selected features were ultimately utilized in the construction of the model and consisted of Gender, Drink, Smoke, Chemotherapy cycles, PE, NEUT, AST, NLR, and NMR ( Figure 2D ).

In the training dataset, the RF model exhibited superior predictive performance with an AUC of 1.00, indicating a high level of accuracy in prediction. In contrast, the AUC values for the remaining five models were as follows: 0.888, 95%CI(0.863-0.911) for LR, 0.822, 95%CI(0.791-0.852) for GNB, 0.792, 95%CI(0.760-0.825) for MLP, 0.719, 95%CI(0.681-0.758) for SVM, and 1.000, 95%CI(NaN- NaN) for KNN ( Figure 3A ). The F1 scores for these models were as follows: LR 0.816, RF 0.998, GNB 0.756, MLP 0.736, SVM 0.679, and KNN nan. In the test set, the AUC values for LR, RF, GNB, MLP, SVM, and KNN were 0.876(95%CI(0.806-0.953)), 0.923(95%CI(0.866-0.979)), 0.817(95%CI(0.726-0.909)), 0.777(95%CI(0.674-0.880)), 0.709(95%CI(0.590-0.828)), and 0.837(95%CI(0.750-0.923)), respectively ( Figure 3B ). The corresponding F1 scores were 0.791, 0.837, 0.747, 0.716, 0.658, and nan for LR, RF, GNB, MLP, SVM, and KNN, respectively. The forest plot comparing the AUC scores of the six ML models is presented in Figure 3C . In this study, the accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and kappa value of each model were computed and compared ( Figures 3D, E ). While the RF model exhibited exceptional performance on the training set, the Logistic Regression model was ultimately selected as the optimal model due to concerns regarding potential overfitting.

The results of the univariate logistic analysis are summarized in Supplementary Table 1 . 12 variables were statistically significant: Gender, Drink, Smoke, CHD, Chemotherapy cycles, Hospitalizations, PE, NEUT, LYM, NLR, NMR, and CEA. Table 4 presents the coefficients and odds ratios (OR) for the nine predictor variables included in the model. The logistic equation was as follows:

y = - 2.954 - 0.424×Gender - 0.049×Drink + 1.754×Smoke + 0.395×Chemotherapy cycles + 1.417×PE + 0.083×NLR + 0.017×NMR + 0.009×AST - 0.008×NEUT. In this study, we evaluated the prediction accuracy and calibration of the model by calibration curve analysis of the training and test sets. The calibration curve results showed that the model in the training set had high prediction accuracy with a Somers’ D coefficient of 0.777 and an area under the ROC curve of 0.888, indicating that the model had excellent discriminative ability ( Figure 4A ). In addition, the logistic regression calibration slope of the training set model was close to the ideal value of 1.000, with an intercept of 0.000, showing excellent calibration. The Brier score of 0.134 reflected the high reliability of the model predictions. In contrast, the model in the test set maintained a high discriminative power with an area under the ROC curve of 0.876, although there was a slight decrease in prediction accuracy (Somers’ D coefficient of 0.751) ( Figure 4B ). The decision curve for the training set ( Figure 4C ) shows that the model provides significantly higher net gains than the baseline strategy when the threshold probabilities are between 0.1 and 0.9. On the test set ( Figure 4D ), the model similarly demonstrates good net returns, especially in the range of threshold probabilities from 0.1 to 0.85, where it maintains a high level of net returns. The confusion matrix results show the difference in the model’s performance on different datasets. In the training set ( Figure 4E ), the model correctly identified 320 true negatives and 283 true positives, and misidentified 42 false positives and 82 false negatives, with a true positive rate (sensitivity) of 77.5% and a true negative rate (specificity) of 88.4%. In the test set ( Figure 4F ), the model correctly identified 32 true negatives and 27 true positives and misidentified 10 false positives and 12 false negatives, for a true positive rate of 69.2% and a true negative rate of 76.2%. Finally, we plotted clinical impact curves (CICs) to assess the net gain in clinical utility and applicability of the model with the highest diagnostic value. The clinical impact curves ( Figures 4G, H ) provide information on the ability of the models to predict high-risk patients at different cost-benefit ratio thresholds. The curves for both the training and test sets show that when the threshold probability is greater than the 65% predictive score probability value, the predictive model’s determination of those at high risk of developing an infection in the lungs after chemotherapy is highly matched to those who actually develop an infection, confirming that the predictive model is clinically highly effective.

This study assessed the relative significance of various factors influencing the susceptibility to lung infections following chemotherapy in patients with lung cancer. Figure 5A visually represents this ranking, with each point denoting a sample and the color gradient from blue to red indicating the magnitude of the sample eigenvalues. The vertical axis displays the importance ranking of features, along with the correlation and distribution of each eigenvalue with the SHAP value. The impact of the top nine features in the importance ranking on prediction outcomes is illustrated in Figure 5B . Specifically, Chemotherapy cycles, Smoke, and PE exhibit positive contributions to the predictive results, while NEUT demonstrate negative influences on the model’s output. Figure 5B illustrates the hierarchical significance of features in the logistic regression model. The vertical axis displays individual features in descending order of importance, while the horizontal axis represents average SHAP values. The analysis reveals that Chemotherapy cycles, Smoke, PE, NMR, and NLR are the top five features ranked by importance, indicating their critical influence on the presence of a lung infection. To enhance comprehension of the model’s decision-making process at the individual level, we conducted a detailed interpretability analysis on two representative samples, as illustrated in Figures 5C, D . By visualizing the SHAP values of these samples, we could discern the impact of each feature on the model’s predictions for these specific instances.

In this study, two nomograms were constructed, integrating nine important predictor variables such as alcohol consumption, smoking, and chemotherapy cycle to visually assess the risk of lung infection after chemotherapy. Figure 6A shows an interactive nomogram with a score of 3.51 for the example patient, corresponding to a 94.5% probability of infection, providing a quick and easy-to-interpret risk assessment. Figure 6B illustrates a dynamic nomogram with different risk profiles derived from 10 combinations of variables.