Integrative habitat analysis and multi-instance deep learning for predictive model of PD-1/PD-L1 immunotherapy efficacy in NSCLC patients: a dual-center retrospective study

Abstract

Background: PD-1/PD-L1 immunotherapy represents the primary treatment for advanced NSCLC patients; however, response rates to this therapy vary among individuals. This dual-center study aimed to integrate habitat radiomics and multi-instance deep learning to predict durable clinical benefits from immunotherapy.

Methods: We retrospectively collected 590 NSCLC patients from two medical centers who received PD-1/PD-L1 inhibitor immunotherapy. Patients from the GMU center were divided into a training cohort (n = 375) and an internal validation cohort (n = 161) for habitat analysis and multi-instance deep learning model development. Patients from the YJ center formed an external testing cohort (n = 54) for model validation. We implemented a DenseNet121-based architecture extracting radiomics features from triplanar (axial/coronal/sagittal) tumor sequences to construct a 2.5D deep-learning dataset. Then, we fuse 2.5D features through multi-instance learning. Additionally, we use K-means clustering to divide the tumor VOI into three subregions to extract radiological features for building a Habitat model. Finally, we use the Extra-Trees classifier to construct MIL, Habitat, and Combined models, the Combined model integrating age factors into the analysis. The primary endpoint was durable clinical benefit. Finally, a separate PD-L1 expression dataset was used to compare the predictive performance of imaging models against PD-L1 status (positive/negative) and expression levels (high/low) to identify the optimal model for predicting immunotherapy clinical benefit.

Results: The Combined model combining Habitat, MIL, and patient age demonstrated robust DCB prediction with AUCs of 0.906(95% CI: 0.874-0.936), 0.889(95% CI: 0.826-0.948), and 0.831 (95% CI: 0.710-0.927)in training, validation, and testing cohorts respectively. Comparative analysis revealed all imaging models outperformed PD-L1 expression status (positive/negative) and levels (high/low) in predicting therapeutic response, with Habitat analysis showing superior performance to MIL alone. Notably, peritumoral structural features emerged as significant predictors of treatment efficacy.

Conclusion: This non-invasive predictive framework provides clinically actionable insights for immunotherapy stratification, potentially overcoming limitations of current biomarker testing while highlighting the prognostic value of spatial tumor heterogeneity analysis.

Keywords: Deep learning; Habitat analysis; Immunotherapy; Multiple instance learning; Non-small cell lung cancer; PD-L1 expression.

Fig. 1

Patient inclusion and exclusion criteria

Fig. 2

The overall research design process

Fig. 3

Illustrates the CH scores, Silhouette Score, and DB Score for different cluster counts, highlighting the influence of cluster number on segmentation effectiveness

Fig. 4

Coefficients of 10-fold cross-validation (A). MSE of 10-fold cross-validation (B). The histogram of the Rad-score is based on the selected features (C)

Fig. 5

The receiver operating characteristic curves, calibration curves, and decision curve analysis of all signatures in the training cohort (A, D, G), validation cohort (B, E, H), and test cohort (C, F, I)

Fig. 6

Visualization of subregions for habitat analysis

Fig. 7

Heat map of lung cancer in axial sequence (A), Coronal sequence (B), and Sagittal sequence (C)

Fig. 8

The ROC of Radiomics Models and PD-L1-Based Models