Delivering Assistance on Time: Over-the-Air Fusion of Movement Intention for Wireless Stroke Rehabilitation

Some of the few rehabilitation methods that move chronic stroke patients past their recovery plateau share one design rule. The robot helps only when the patient is already trying to move. In a pilot trial, a hand exoskeleton that assisted only once the patient's own joint torque exceeded 20% of their maximum produced gains that exceeded the clinically meaningful threshold in every treated patient (Efficacy of robot-assisted fingers training in chronic stroke survivors). By contrast, usual care in chronic patients yields almost nothing: a pooled 0.9-point gain on a 66-point motor scale (Upper limb motor recovery in chronic stroke). That exoskeleton was a single wired device. Home and multi-device rehabilitation instead spreads the evidence of intention across several wireless body-worn sensors, and the assistance is useful only inside a short window after the attempt begins. Wireless links fail in exactly the way that hurts such a window. Standard digital links either arrive intact or, below a signal-quality threshold, drop out entirely (the "cliff effect"), and retransmissions spend the time budget.

We propose to treat this timing window as the quantity the communication system is designed for. The idea is to fuse intention evidence from several sensors by letting their analog radio signals add up in the air. This is over-the-air computation: simultaneous transmissions on one frequency arrive as a weighted sum, so the receiver gets the fused signal in one channel use. The sensor-side encoders and actuator-side trigger are trained end to end through a simulated noisy, fading channel, with a loss that rewards triggers landing inside the window and penalizes false triggers. The claim is narrow and testable. When no single sensor can detect intention reliably, this design should deliver more on-time, correct assistance per unit of radio energy than local detection with voting or digital streaming. When one sensor suffices, it should not, and we will show that too.

Core Claim

The communication-for-learning literature has already shown that analog, channel-aware transmission degrades gracefully where digital transmission collapses. AirNet sends network weights as analog symbols and recovers model accuracy at 15 dB, where digital compression baselines fail even at 35 dB (AirNet). AnalogDeepCMC treats the wireless link as a non-trainable noisy layer inside an autoencoder and gains 19.3% downlink rate at 10 dB (CNN-Based Analog CSI Feedback). Both motivate the analog design, but both optimize average reconstruction or accuracy. Neither has a deadline.

AirGNN embeds fading directly into the graph aggregation step, computes neighbor sums over the air, and proves SGD converges when trained through the random channel (Graph Neural Networks Over the Air). AirTNN extends this to higher-order topologies (Topological Neural Networks over the Air). We reuse this mechanism. What we change is the graph and the objective. The graph becomes a patient's body-sensor network, and the objective becomes event timing rather than node classification or flocking.

The closest rehabilitation work, Federated Joint Learning of rehabilitation robot networks, learns trajectories across clinics but assumes a lossless network and is evaluated only in simulation (Federated Joint Learning of Robot Networks). It addresses slow, cross-site model sharing. We address the fast, within-patient control loop, which is where a timing window exists.

Formally, at each 10 ms step $t$ sensor $i$ sends an analog feature $x_{i,t}=f_\theta(s_{i,t})$ computed from its local signal $s_{i,t}$. The actuator-side receiver observes

$$y_t=\sum_i h_{i,t}\,x_{i,t}+n_t,$$

where $h_{i,t}$ is the random fading gain and $n_t$ is Gaussian noise. A trigger network $g_\phi$ maps recent $y_t$ to a firing probability. We train $\theta,\phi$ through sampled channels on

$$\mathcal{L}=-\sum_{k}\log P\big(\text{fire in }[\tau_k,\tau_k+D]\big)+\lambda\,\mathbb{E}[\text{fires outside any window}],$$

where $\tau_k$ is the onset of the $k$-th true movement attempt, $D$ is the window length, and $\lambda$ trades missed help against unwanted help. The file contains no physiological value for $D$, so we sweep it rather than fix it.

Research Program

The work has three parts.

The first is a benchmark. We build distributed intention-detection tasks from existing recordings by splitting body keypoints or sensors into virtual wireless nodes, one per body segment. Onset labels are defined from movement kinematics, as a proxy for the torque-threshold trigger. Sources are KiMoRe rehabilitation exercise kinematics (200 patients including upper-limb hemiplegia, as used in Federated Joint Learning) and, if access can be arranged, the four-IMU seven-movement protocol of Estimation of Fugl–Meyer Sub-Scores. We label onset as the first sample where segment velocity exceeds a patient-normalized threshold. We then evaluate detection on the preceding, weaker signal, so the task is genuinely early and genuinely distributed.

The second part is the method and its channel simulation. Channels use Rayleigh fading with additive noise across an SNR sweep, plus erasures for the digital arms. The file supplies no body-area channel model; the COST 2100 model it lists targets indoor massive-MIMO cells. We therefore use generic fading and flag this as a threat to validity.

The third part is a bench test. Healthy volunteers wear commodity radios on several segments, and we measure real channel traces during reaching and grasping. The trained models are replayed on those traces. Clinical efficacy is out of scope. The question here is engineering: does the timing window survive the channel?

A secondary analysis follows from a tension in the literature. Training through channel noise acts as a regularizer (AirGNN). Separately, trunk and pelvic compensation can masquerade as recovery (Wearable-informed generative digital avatars). We will test on held-out patients whether channel-noise training makes the trigger rely less on compensatory trunk signals, measured by attribution to the trunk node.

Baselines

The naive reference is continuous assistance with no trigger, a fixed-schedule assist. It gives 100% "delivery" but a maximal false-assist rate. It anchors the false-trigger axis and represents non-contingent help, the kind the plateau literature suggests is ineffective.

The strongest competitor is local detection with robust digital voting. Each sensor runs its own detector, sends a one-bit decision in a short, heavily coded packet, and the actuator fires on an OR or k-of-n rule tuned on validation data. This is textbook distributed detection. It is cheap, and it avoids the cliff effect because one-bit packets can be protected heavily.

The second competitor is digital streaming of quantized features with retransmission. It follows the MRC retransmission mechanism of Decentralized Inference with GNNs: nodes send quantized features in time slots, the receiver retransmits until a reliability target is met, and a central detector fuses the results. Lee et al. report near-zero misclassification when their robustness target is met but give no numeric latency. We found no reported result on Lacuna for trigger latency or contingent delivery of any method under wireless impairments, so all baseline numbers will come from our own runs.

All arms share the same training data, the same detector backbone capacity (a small temporal convolution per node), the same hyperparameter search budget, and the same energy budget per event, counted in channel uses times transmit power.

Two ablations isolate the proposal's ingredients. The first is analog over-the-air fusion trained with reconstruction loss: AnalogDeepCMC-style, with the trigger trained afterwards on reconstructed features. It removes the timing loss while keeping the analog channel. The second is the timing loss with digital orthogonal transmission, which keeps the objective but removes superposition.

Beating the strongest baseline means a higher area under the contingent-delivery-versus-SNR curve at matched false-trigger rate and matched energy, with the margin exceeding the across-seed spread on held-out patients.

Why not just…?

Why not just detect locally and send one well-protected bit? This is the real threat. If one body segment carries enough early evidence, local detection plus a strongly coded bit will match or beat our method at lower complexity. The proposal only pays off when evidence is weak and spread across nodes, because thresholding each node discards the sub-threshold information that joint fusion keeps. We test this directly by varying how evidence is distributed, from single-segment tasks like finger flexion to multi-segment reaching. We expect, and will report, a crossover.

Why not just stream features over standard digital links with retransmission? Retransmission is effective for reliability, as Lee et al. show (Decentralized Inference with GNNs), but each round consumes part of a fixed window. Digital schemes also fail abruptly below a threshold SNR (AirNet). The SNR sweep with deadline accounting will show whether this matters at realistic SNRs, or only at SNRs so low that nothing works.

Why not just wire everything onto the device, as the exoskeleton trial did? For a single self-contained exoskeleton, that is the right answer, and this proposal is not needed. Susanto et al. sense torque on the device itself (Efficacy of robot-assisted fingers training). The scope here is multi-wearable home setups, where the sensors sit on segments other than the actuator. We state this limit rather than claim generality.

Why not just use an analog autoencoder and train the trigger afterwards? This is the reconstruction-loss ablation. If it matches the timing-loss arm, the contribution reduces to "use analog over-the-air fusion", and we will say so.

Proposed Experiments and Decisive Checks

At high SNR all arms should reach similar delivery rates. If they do not, the pipeline is broken.

The main result is the delivery-versus-SNR curve at a fixed false-trigger rate, for window lengths swept over a range. It is reported separately for single-segment and multi-segment tasks, on leave-patients-out splits of KiMoRe.

The decisive prediction has two sides. On multi-segment tasks at moderate-to-low SNR, over-the-air fusion with the timing loss beats local voting at equal energy. On single-segment tasks, it does not. If it also fails to beat local voting on multi-segment tasks, the core claim is false.

The two ablations then attribute any gain to superposition, to the timing objective, or to both.

Finally, the bench traces test whether simulated fading ranks the methods the same way real body-worn channels do. A ranking reversal on real traces would mean the simulation results should not be trusted.