An Agentic Multimodal Sensing Architecture for CT-Guided Wearable and Respiratory Monitoring in Oncology Care

Oncology care increasingly depends on heterogeneous sensing streams generated by computed tomography (CT), radiotherapy planning systems, wearable devices, home respiratory sensors, patient-reported outcomes, and clinical records. These data streams are often processed separately, limiting their value for longitudinal, context-aware review. This study proposes OncoSense-Agent, a reliability-aware agentic multimodal sensing architecture for CT-guided respiratory monitoring in oncology care. The architecture links CT-derived anatomical evidence with wearable physiology, respiratory symptoms, functional assessment, treatment context, and explainable human-in-the-loop review-priority generation. To move beyond a purely conceptual design, we implemented a lung-focused proof-of-concept with six bounded software agents: Imaging Reliability, Wearable Monitoring, Respiratory Review, Treatment Context, Multimodal Fusion, and Explainability. The prototype used real nnU-Net v2 3D lung segmentation metrics from 139 patients with complete bilateral lung CT data as the imaging anchor, while wearable, respiratory, symptom, and treatment-context channels were introduced as deterministic overlays for controlled validation. OncoSense-Agent changed review-priority assignment relative to CT-only assessment in 78/139 cases (56.1%), assigned 111/139 cases (79.9%) to high-priority or high-uncertainty tiers, and showed increasing Safety Gate activation as CT quality declined. Three illustrative cases demonstrate hidden respiratory deterioration, wearable data-quality uncertainty, and treatment-context risk not captured by CT-only assessment. The prototype does not establish clinical diagnostic accuracy, but demonstrates operational, auditable, reliability-aware multimodal review-priority generation for clinician-supervised oncology monitoring.