Combining tumor habitat radiomics and circulating tumor cell data for predicting high-grade pathological components...

Lung cancer is the leading cause of cancer incidence and mortality. Early surgical resection significantly improves patient prognosis. Studies have shown that high-grade components in lung adenocarcinoma (LUAD) severely impact patient outcomes. Therefore, early prediction of these high-grade components is crucial for clinical surgical decision-making.

We delineated tumor subregions using k-means clustering and excluded features with low reproducibility by calculating the intraclass correlation coefficient (ICC). Stratified sampling ensured a consistent sample distribution between the training and testing datasets, and Borderline synthetic minority over-sampling technique (BorderlineSMOTE) addressed the data imbalance in the training dataset. Normality tests were conducted, followed by feature selection using independent sample t tests, Mann‒Whitney U tests, and Spearman rank correlation. Principal component analysis (PCA) of reduced dimensionality and model integration were performed using a stacking approach. Model predictive performance was evaluated using the area under the curve (AUC), and significant differences between models were assessed using the DeLong test.

The combined Habitat-circulating tumor cell (CTC) model showed the best predictive performance for high-grade components in both the training and validation datasets, achieving AUCs of 0.98 [95 % CI: 0.95-1.00] and 0.91 [95 % CI: 0.82-1.00], respectively. In the training dataset, the combined model's AUC of 0.98 [95 % CI: 0.95-1.00] was notably higher than that of the single CTC model, which achieved an AUC of 0.75 [95 % CI: 0.64-0.85], and the single sub-region model, which had an AUC of 0.94 [95 % CI: 0.88-1.00]. Decision-curve analysis demonstrated maximal net benefit at threshold probabilities of 0.2-0.4. In the independent cohort (n = 29), AUC reached 1.00 [95 % CI: 1.00-1.00].

Combining habitat radiomics and CTC-related clinical models allows for more precise prediction of high-grade pathological components, aiding in clinical preoperative decision-making.