Assessing Stand-to-Sit Kinematics via mmWave Radar: A Real-to-Sim Robust Bidirectional State-Space Model

Continuous monitoring of the Stand-to-Sit (STS) transition serves as a critical indicator of lower-limb frailty in the elderly, for which millimeter-wave radar provides an ideal privacy-preserving, device-free sensing solution. However, robustly distinguishing between safe Controlled Sits (CSs) and dangerous Uncontrolled Descents (UDs) is severely hindered by the prohibitive cost of subjective expert scoring for fine-grained labels, alongside the pervasive “Clever Hans” effect where existing deep models overfit static environmental clutter rather than learning intrinsic human kinematics. To circumvent these bottlenecks, we formulate STS evaluation as a dynamic boundary detection problem and propose SCA-BiMamba, a linear-complexity bidirectional State-Space Model that utilizes actual fall events as extreme kinematic surrogates for UDs. This forces the network to learn a strict physical boundary between CS and physiological failure without subjective grading. Furthermore, we establish a stringent Real-to-Sim diagnostic audit as a core methodological contribution. By projecting models trained on noisy real-world data onto pure-kinematics simulations—incorporating stochastic temporal phase shifts, kinematic overlaps, and unified physiological tremors—we explicitly quantify feature disentanglement. This protocol serves as a formal ‘probing test’ to expose the ‘Clever Hans’ effect, ensuring the model relies on invariant human physics rather than transient environmental artifacts. Extensive experiments demonstrate that SCA-BiMamba achieves highly robust classification on real-world data (averaging 94.2% Macro F1 with 100.0% Uncontrolled Descent Recall), and achieves a highly robust 99.4% ± 1.1% Macro F1 in the simulated zero-shot transfer. We emphasize that this optimal performance reflects the successful abstraction of extreme kinematic boundaries, rather than a flawless resolution of all clinical complexities. Concurrently, it exhibits strict resistance to shortcut learning and sustains robust real-world scalability using merely 20% of the training data, thereby establishing a promising privacy-preserving boundary-based radar motion classification framework for distinguishing controlled sitting from extreme instability surrogates.