Multimodal artificial intelligence predicts PIK3CA mutation in breast cancer from digital pathology and clinical...

Area of Science:

• Oncology
• Computational Biology
• Medical Imaging

Background:

• Accurate PIK3CA mutation detection is crucial for guiding PI3K-targeted therapies in breast cancer.
• Current sequencing methods are not universally accessible, and single-modality prediction models show limited performance.

Purpose of the Study:

• To develop a multimodal deep learning framework integrating whole-slide imaging (WSI) and structured clinical data for improved PIK3CA mutation prediction in breast cancer.

Main Methods:

• A multimodal deep learning framework (MPM) was developed, integrating a histopathology model (H-optimus-0, CLAM-SB MIL) with a clinical model (XGBoost).
• The framework utilized a decision-level late fusion strategy, combining outputs from WSI and clinical data analysis.
• Performance was evaluated using AUC, with interpretability assessed via attention heatmaps and SHAP analysis across TCGA and external cohorts.

Main Results:

• The multimodal framework (MPM) outperformed single-modality models, achieving an AUC of 0.745 on TCGA data.
• Stable performance was observed across external cohorts (AUCs of 0.695, 0.690, 0.680).
• SHAP analysis identified molecular subtype as a key clinical predictor, while attention maps pinpointed mutation-associated morphological regions.

Conclusions:

• The multimodal framework effectively integrates morphological and clinical data for robust PIK3CA mutation prediction.
• The model demonstrates strong multicenter adaptability and biological interpretability, supporting its potential as a clinical decision-support tool.
• This approach offers a promising, accessible alternative to molecular testing for PIK3CA mutation status in breast cancer.