Stable measurement framework for rheumatoid arthritis assessment using late fusion of thermal radiomics and lightweight convolutional representations

Reliable assessment of rheumatoid arthritis (RA) using infrared thermography requires a measurement framework capable of characterizing the dynamic thermal response to controlled excitation while limiting sensitivity to acquisition conditions, inter-subject variability, and computational instability. This study presents a measurement-oriented Active Dynamic Thermography (ADT) framework in which a dual-stream late-fusion stage combines radiomic descriptors with convolutional representations of the acquired thermal response. The measurand was defined as the spatio-temporal thermal response of the hand during post-stimulus rewarming, interpreted as an indirect functional indicator of superficial microvascular and subcutaneous vascular-bed activity associated with inflammatory processes. Controlled cold excitation provided a standardized thermal perturbation for observing vascular recovery dynamics, while radiometric verification and protocol standardization were used to reduce non-biological acquisition variability. The proposed framework was evaluated on thermographic recordings acquired from 187 patients with clinically confirmed RA and 292 healthy controls. To assess the robustness of the diagnostic output to cohort partitioning, experiments were repeated across ten independent subject-disjoint realizations. FUSION achieved a mean accuracy of 91.0%, balanced accuracy of 91.4%, Matthews correlation coefficient (MCC) of 0.819, ROC-AUC of 0.954, and PR-AUC of 0.901. Compared with the evaluated standalone baselines, FUSION showed higher mean performance and lower descriptive inter-seed variability. Ranking and distributional analyses suggested improved stability. Paired Wilcoxon comparisons showed significant improvements over the standalone CNN baseline for several discriminative metrics after adjustment, whereas differences relative to the stronger INR baseline did not reach the adjusted significance threshold. The results support the feasibility of a measurement-oriented ADT framework in which complementary thermal descriptors are integrated within a computational decision stage. The fusion component should therefore be interpreted as part of the broader thermal measurement chain rather than as the measurement principle itself. Further test-retest, inter-device, and external validation is required to establish physical repeatability, reproducibility, and clinical generalizability.