🤖 Generated with AI assistance · Does not replace individualized medical advice.

CLASP: Knowledge-Guided Multimodal Learning for Interpretable pCR and Survival Prediction in Breast Cancer

AI Transparency Notice: This post was generated by an AI assistant acting as an expert oncology curator. It summarizes published research for educational purposes and does not constitute medical advice, diagnosis, or treatment recommendations. Always consult with a qualified healthcare provider for clinical decisions.

Lead

CLASP, a knowledge-driven multimodal framework, integrates whole slide images, clinical data, and LLM-generated pathological priors to predict pathological complete response (pCR) and survival in breast cancer. Validated across four independent cohorts, it achieves an AUC of 0.881 internally and maintains robust external performance, offering a more interpretable alternative to black-box deep learning models for pretreatment risk stratification.

Key Facts

Metric	Value	Context
Internal AUC (pCR)	0.881	Discriminative performance in internal testing PMID: 42854913
External AUC (pCR)	0.875 / 0.845	Robustness across two independent external validation cohorts PMID: 42854913
Cohort Size	1,486 patients	Aggregated from four independent medical centers PMID: 42854913
Few-Shot Performance	Competitive with 8 samples	Demonstrates high data efficiency in low-resource settings PMID: 42854913

Why It Matters / Context

The integration of artificial intelligence into digital pathology has accelerated rapidly, yet a significant gap remains between technical performance and clinical utility. Traditional deep learning approaches for predicting treatment response often function as “black boxes,” relying on visual pattern recognition that is difficult for pathologists to interpret or trust. This lack of interpretability, combined with high data dependency and limited generalizability across heterogeneous cohorts, has hindered the adoption of these tools in routine clinical workflows.

The ability to accurately predict pathological complete response (pCR) to neoadjuvant chemotherapy is critical for risk stratification. Patients who achieve pCR generally have better long-term outcomes, while those who do not may benefit from de-escalation or escalation strategies in future response-adapted trials. However, current methods struggle to generalize across different institutions due to variations in staining, imaging equipment, and patient demographics.

CLASP addresses these limitations by shifting from purely data-driven learning to a knowledge-guided approach. By synergizing pathological expert knowledge with vision-language foundation models, the framework aligns visual biomarkers with semantic concepts. This not only improves predictive accuracy but also enhances interpretability, allowing clinicians to understand why a model predicts a specific outcome. The inclusion of 16 hierarchical pathological text priors generated by Large Language Models (LLMs) ensures that the model is grounded in established medical knowledge rather than spurious correlations.

Details

The study evaluated CLASP against 11 baseline methods across four independent cohorts. The framework integrates three primary data streams:

1. Whole Slide Image (WSI) Features: Encoded by the CONCH vision-language foundation model.

2. Clinical Data: Patient-specific demographic and clinical variables.

3. Pathological Text Priors: 16 hierarchical concepts generated by an ensemble of LLMs.

A novel vision-language feature aggregation module and a multi-task consistency constraint were employed to jointly predict pCR and prognosis. The following table summarizes the comparative performance of CLASP against baseline methods, highlighting its superior discriminative ability and robustness.

Model Category	Representative Baseline	Internal AUC (pCR)	External AUC 1 (pCR)	External AUC 2 (pCR)	Interpretability
CLASP (Proposed)	Knowledge-Guided Multimodal	0.881 PMID: 42854913	0.875 PMID: 42854913	0.845 PMID: 42854913	High (Attention maps aligned with TILs/stroma)
Deep Learning	Standard CNN/ViT	Lower than CLASP PMID: 42854913	Lower than CLASP PMID: 42854913	Lower than CLASP PMID: 42854913	Low (Black-box)
Clinical Only	Logistic Regression	Lower than CLASP PMID: 42854913	Lower than CLASP PMID: 42854913	Lower than CLASP PMID: 42854913	High (Coefficients)
Image Only	Single-Modality DL	Lower than CLASP PMID: 42854913	Lower than CLASP PMID: 42854913	Lower than CLASP PMID: 42854913	Low (Black-box)

Note: Specific AUC values for individual baselines were not provided in the abstract; the table reflects the relative performance hierarchy stated in the source text, where CLASP yielded higher AUC point estimates than all 11 evaluated baseline methods PMID: 42854913.

Interpretability analyses revealed that the regions highlighted by CLASP were consistent with recognized pathological features associated with treatment response. Specifically, the model focused on tumor-infiltrating lymphocytes (TILs) and stromal architecture, which are well-established biological indicators of immune response and prognosis in breast cancer. This alignment with known biology significantly enhances the clinical credibility of the model’s predictions.

Furthermore, CLASP demonstrated improved data efficiency. In few-shot learning scenarios, the model achieved competitive performance with as few as 8 training samples PMID: 42854913. This is a critical advantage for rare subtypes of breast cancer or institutions with limited annotated datasets, where traditional deep learning models typically fail due to insufficient data.

Limitations

⚠️ Critical Limitations

* Retrospective Design: The study is based on retrospective data from four independent cohorts. While external validation was performed, prospective clinical trials are necessary to confirm real-world utility and impact on patient outcomes.

* Generalizability to Rare Subtypes: While the model shows promise in few-shot learning, the performance on rare histological subtypes or specific molecular profiles (e.g., HER2-low, triple-negative with specific mutations) may vary and requires further stratified analysis.

* Interpretability vs. Causality: Although the model highlights biologically relevant features like TILs, this does not establish a causal link between these features and the predicted outcome. The attention maps provide a plausible explanation but do not replace pathological diagnosis.

* Data Heterogeneity: Despite validation across multiple centers, variations in WSI acquisition protocols, staining techniques, and clinical data recording standards may still affect model performance in new, unseen environments.

* Lack of Prospective Clinical Impact Data: The study does not provide evidence that using CLASP for risk stratification leads to improved patient survival or better treatment decisions in a clinical setting.

Clinical Takeaway

Clinicians should view CLASP as a promising candidate for pretreatment risk stratification in breast cancer, particularly for its ability to provide interpretable predictions grounded in pathological knowledge, but it should not be used for individual patient management decisions until validated in prospective clinical trials.

Full Reference

Zhang W, Zhang S, Wei Y, Zhao Y, Yuan W, Wu X. Knowledge-guided Multimodal Learning for Interpretable Prediction of Pathological Complete Response and Survival in Breast Cancer. Laboratory Investigation. 2026. PMID: 42854913 [DOI: 10.1016/j.labinv.2026.106169]