Adaptive AI-driven clinical decision-making in oncology through digital twins and large-scale imaging biomarkers.

Abstract

e13698 Background: Small deviations in clinical attributes particularly biomarkers play a crucial role in oncology disease staging, affecting prognosis, diagnosis, and treatment decisions. This clearly emphasizes the necessity for substantially precise and specific biomarker tests, as well as more sophisticated interpretative models to translate these subtle variations into meaningful clinical decisions. In addition, such subtle variations highlight the significance of personalized approaches to cancer treatment. Similarly, the low specificity of some biomarkers may lead to overlooked cancer cases or late detections, emphasizing the need for improved testing methods. This problem will be aggravated once high-dimensional (HD) data resources need to be integrated. Here, we used a scenario to integrated HD-to-LD (lower dimensional) data through digital twin (DT) to assess adaptability of decision-making in oncology leveraging our own defined mapping, for personalized cancer treatment. Methods: DT model is designed and used for large-scale multimodal imaging data across lung and breast cancer diseases to enhance adaptability of overall survival (OS) and diagnostic prediction. We integrated Parzen-Rosenblatt constrained isometric mapping through DT simulation disease prognosis and diagnosis, advancing robustness in precision oncology through data-driven decision-making. Results: We have comprehensively tested the proposed DT model across 1,713 cases spanning from multiple datasets 211 cases, 57,836 CT/PET images, 422 CT cases, 88 CT scans for none-small cell lung cancer (NSCLC), 780 breast cancer with ultrasound images, and 212 breast cancer screening with Mamo+Thermography images. Initial embedding accuracies were 78.5% (±4.4), 88.4% (±1.4), and 61.4% (±11.4), for three lung cancer data, 80.3% (±5.5), for breast cancer with ultrasound, and 82.9% (±2.3) for breast cancer mamo+thremo, validated under biomarker fluctuations (±10%) in LD space using PR-isometric embedding to ensure model fidelity (±4%) to patient variability. Conclusions: The results demonstrating the success of DT model in handling large-scale multimodal data sets the stage for transformative changes in robust cancer management and treatment strategies. Our results indicate that DT can be used in cancer care and management for redefining personalized healthcare and improving overall patient outcomes.