Multi-Objective Evaluation of Wearable-Assisted Radar-Only Rehabilitation Monitoring

Millimetre-wave radar offers contactless exercise sensing, while wearable inertial measurement units (IMUs) can supply privileged training supervision without being required at inference. We study radar-only exercise recognition on the mRI dataset (20 healthy subjects; ten exercises) and distinguish descriptive historical architectures from matched RTPD-Net versus constrained fine-tuning (CFT) comparisons. Historical RTPD-Net, PARETO-RKD, and AURORA-RKD were trained using different budgets, seeds, or corruption settings and cannot establish a preferred method. For a fixed 40-segment test split, the original evaluation reported mean corrupted macro-F1 changing from 0.5332 to 0.6945; a paired re-evaluation of saved checkpoints, using 11 corrupted conditions and five realisations each, instead obtained 0.5370 versus 0.6374 (difference 0.1004). The subject-cluster bootstrap interval, conditional on the same four subjects, was [0.0683, 0.1283], but the exact two-sided subject-level prediction-swap test gave p=0.125 (four independent test subjects), so statistical significance is not established. Five independently generated subject-partition seeds (25 fully retrained folds) produced mean corrupted macro-F1 of 0.5375 for the fold-local RTPD reference versus 0.5449 for the validation-gated policy (mean difference 0.0074); fine-tuning was accepted in 5 of 25 folds. These results demonstrate that the validation gate usually retains the reference and that the observed adaptation gains are not established as broadly useful; they also indicate that robustness benefits are heterogeneous and that larger independent-subject evaluations are necessary.