Dielectric property measurements for the rapid differentiation of thoracic lymph nodes using XGBoost in patients with non-small cell lung cancer: a self-control clinical trial.

BACKGROUND

One of the important criteria for thoracic surgeons in making surgical strategies is whether the thoracic lymph nodes (LNs) are metastatic. Frozen section (FS) is widely used as an intraoperative diagnostic method, which is time-consuming and expensive. The dielectric property, including permittivity and conductivity, varies with different tissues. The extreme gradient boosting (XGBoost) is a powerful classifier and widely used. Thus, this study aims to develop the rapid differentiation method combining dielectric property and XGBoost, and assess its efficacy on the thoracic LNs in patients with non-small cell lung cancer (NSCLC).

METHODS

This was a single center self-control clinical trial with paraffin pathology section (PPS) results as gold diagnosis. The LNs from the pathologically diagnosed patients with NSCLC were recruited, which were measured by open-ended coaxial probe for the dielectric property within 1-4,000 MHz after removal from the patients and then were sent to perform FS and PPS diagnosis. The XGBoost combining with dielectric property was developed to differentiate malignant LNs from benign LNs. The classified efficacy was determined using the receiver operator characteristic (ROC) curve and area under the curve (AUC).

RESULTS

A total of 204 LNs from 67 NSCLC patients were analyzed. The mean values of the two parameters differed significantly (P<0.001) between benign and malignant LNs. The AUC for permittivity and conductivity were 0.850 [95% confidence interval (CI): 0.786 to 0.915; P<0.001] and 0.887 (95% CI: 0.828 to 0.946; P<0.001), respectively. The AUC was 0.893 (95% CI: 0.834 to 0.951; P<0.001) when the two parameters were combined. After the application of the XGBoost, the AUC was 0.968 (95% CI: 0.918 to 1.000; P<0.001), and the accuracy was 87.80%. Its sensitivity was 58.33% and the specificity was 100%. When the Synthetic Minority Oversampling Technique (SMOTE) algorithm was used, the AUC was 0.954 (95% CI: 0.883 to 1.000; P<0.001) and the accuracy was 92.68%. Its sensitivity was 83.33% and the specificity was 96.55%.

CONCLUSIONS

This method might be useful for thoracic surgeons during surgery, for its relatively high efficacy in rapid differentiation of LNs for patients with NSCLC.