Multimodal AI-driven Biomarker for Early Detection of Cancer Cachexia

Area of Science:

• Oncology
• Artificial Intelligence
• Biomarker Development

Background:

• Cancer cachexia is a severe metabolic syndrome linked to poor cancer outcomes.
• Early detection is hindered by a lack of standardized biomarkers.
• Existing detection methods lack generalizability and early detection capabilities.

Purpose of the Study:

• To develop and validate a multimodal AI-based biomarker for early cancer cachexia detection.
• To leverage foundation models and large language models (LLMs) for enhanced predictive accuracy.
• To create a scalable and clinically deployable tool for personalized oncology.

Main Methods:

• Trained a multimodal AI biomarker on clinical, radiologic, laboratory, and unstructured clinical note data.
• Utilized foundation models and LLMs for feature extraction and prediction.
• Integrated diverse data modalities including clinical text and CT image embeddings.

Main Results:

• Achieved 92% prediction accuracy by integrating clinical text and CT image embeddings.
• Demonstrated improved survival prediction with increasing data modality integration.
• Framework accommodates missing data, ensuring scalability and real-world applicability.

Conclusions:

• The AI biomarker offers a clinically applicable, scalable, and trustworthy tool for early cancer cachexia detection.
• Patient-specific predictions and uncertainty estimation enhance clinical reliability and personalized care.
• This approach facilitates integration into existing oncology workflows using standard-of-care data.