PSA-1DCNN: Multimodal Biomarker and Text Integration for Lung Cancer Diagnosis.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Oncology

Background:

• Integrating diverse healthcare data like clinical narratives and biomarkers is challenging due to format and semantic heterogeneity.
• Existing deep learning methods struggle to effectively fuse clinical text and biomarkers, limiting precision medicine applications.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework, PSA-1DCNN, for effective multimodal data integration in lung cancer detection.
• To investigate optimal strategies for fusing clinical text and biomarker data to enhance diagnostic accuracy.

Main Methods:

• Proposed a Parallel Self-Attention 1D Convolutional Neural Network (PSA-1DCNN) combining self-attention for text and 1D CNNs for biomarkers.
• Explored four distinct fusion strategies to optimize cross-modal information integration.
• Validated the model on MIMIC-III and MIMIC-IV datasets, comparing against established baselines like ClinicalBERT, LSTM, and 1D-CNN.

Main Results:

• PSA-1DCNN significantly outperformed state-of-the-art methods in lung cancer detection.
• Achieved a high F1-score of 98.4% on the MIMIC-IV dataset.
• Demonstrated strong cross-version generalization from MIMIC-IV to MIMIC-III and provided clinically relevant insights via SHAP interpretability.

Conclusions:

• The PSA-1DCNN framework offers a scalable and interpretable solution for integrating heterogeneous healthcare data modalities.
• This approach advances precision oncology by enabling more accurate diagnostics and personalized treatment strategies.
• Highlights the potential of combining advanced AI with multimodal data for improved patient outcomes.