Objective To explore the recognition capabilities of electronic nose combined with machine learning in identifying the breath odor map of benign and malignant pulmonary nodules and Traditional Chinese Medicine (TCM) syndrome elements. MethodsThe study design was a single-center observational study. General data and four diagnostic information were collected from 108 patients with pulmonary nodules admitted to the Department of Cardiothoracic Surgery of Hospital of Chengdu University of TCM from April 2023 to March 2024. The patients' TCM disease location and nature distribution characteristics were analyzed using the syndrome differentiation method. The Cyranose 320 electronic nose was used to collect the odor profiles of oral exhalation, and five machine learning algorithms including random forest (RF), K-nearest neighbor (KNN), logistic regression (LR), support vector machine (SVM), and eXtreme gradient boosting (XGBoost) were employed to identify the exhaled breath profiles of benign and malignant pulmonary nodules and different TCM syndromes. Results(1) The common disease locations in pulmonary nodules were ranked in descending order as liver, lung, and kidney; the common disease natures were ranked in descending order as Yin deficiency, phlegm, dampness, Qi stagnation, and blood deficiency. (2) The electronic nose combined with the RF algorithm had the best efficacy in identifying the exhaled breath profiles of benign and malignant pulmonary nodules, with an AUC of 0.91, accuracy of 86.36%, specificity of 75.00%, and sensitivity of 92.85%. (3) The electronic nose combined with RF, LR, or XGBoost algorithms could effectively identify the different TCM disease locations and natures of pulmonary nodules, with classification accuracy, specificity, and sensitivity generally exceeding 80.00%.ConclusionElectronic nose combined with machine learning not only has the potential capabilities to differentiate the benign and malignant pulmonary nodules, but also provides new technologies and methods for the objective diagnosis of TCM syndromes in pulmonary nodules.
Objective To construct a "disease-syndrome combination" mathematical representation model for pulmonary nodules based on oral microbiome data, utilizing a multimodal data algorithm framework centered on dynamic systems theory. Furthermore, to compare predictive models under various algorithmic frameworks and validate the efficacy of the optimal model in predicting the presence of pulmonary nodules. MethodsA novel multimodal data algorithm framework centered on dynamic systems theory, termed VAEGANTF (Variational Auto Encoder-Generative Adversarial Network-Transformer), was proposed. Subsequently, based on a multi-dimensional integrated dataset of "clinical features-syndrome elements-microorganisms", all subjects were divided into training (70%) and testing (30%) sets for model construction and efficacy testing, respectively. Using healthy individuals and patients with pulmonary nodules as dependent variables, and combining candidate markers such as clinical features, lesion location, disease nature, and microbial genera, the independent variables were screened based on variable importance ranking after identifying and addressing multicollinearity. Missing values were then imputed, and data were standardized. Eight machine learning algorithms were then employed to construct pulmonary nodule risk prediction models: random forest, least absolute shrinkage and selection operator (LASSO) regression, support vector machine, multilayer perceptron, eXtreme gradient boosting (XGBoost), VAE-ViT (Vision Transformer), GAN-ViT, and VAEGANTF. K-fold cross-validation was used for model parameter tuning and optimization. The efficacy of the eight predictive models was evaluated using confusion matrices and receiver operating characteristic (ROC) curves, and the optimal model was selected. Finally, goodness-of-fit testing and decision curve analysis (DCA) were performed to evaluate the optimal model. ResultsThere were no statistically significant differences between the two groups in demographic characteristics such as age and sex. The 312 subjects were randomly divided into training and testing sets (7∶3), and prediction models were constructed using the eight machine learning algorithms. After excluding potential problems such as multicollinearity, a total of 301 clinical feature information, syndrome elements, and microbial genera markers were included for model construction. The area under the curve (AUC) values of the random forest, LASSO regression, support vector machine, multilayer perceptron, and VAE-ViT models did not reach 0.85, indicating poor efficacy. The AUC values of the XGBoost, GAN-ViT, and VAEGANTF models all reached above 0.85, with the VAEGANTF model exhibiting the highest AUC value (AUC=0.923). Goodness-of-fit testing indicated good calibration ability of the VAEGANTF model, and decision curve analysis showed a high degree of clinical benefit. The nomogram results showed that age, sex, heart, lung, Qi deficiency, blood stasis, dampness, Porphyromonas genus, Granulicatella genus, Neisseria genus, Haemophilus genus, and Actinobacillus genus could be used as predictors. Conclusion The "disease-syndrome combination" risk prediction model for pulmonary nodules based on the VAEGANTF algorithm framework, which incorporates multi-dimensional data features of "clinical features-syndrome elements-microorganisms", demonstrates better performance compared to other machine learning algorithms and has certain reference value for early non-invasive diagnosis of pulmonary nodules.