Wearable rehabilitation systems increasingly want to compute brain-network measures at the bedside. One example is how efficiently motor regions exchange information while a stroke patient imagines moving a limb. Computing these measures on a small battery-powered chip, rather than a workstation, means trading numerical accuracy for energy. Today that trade is set by engineering convention, for example "match the floating-point reference to some number of bits." It is not set by what the clinician actually uses, which is a correlation between the network measure and a recovery score. This proposal asks how much numerical error a clinically meaningful brain-network biomarker can absorb. It then uses that tolerance, rather than bit-exactness, to choose the chip's supply voltage and its forward body bias. Forward body bias is a voltage applied to the transistor's substrate that lowers the switching threshold. The chip then runs faster at a given supply, at the cost of more leakage current.
The motivating observation is that the clinical correlation is itself very noisy. The time-varying connectivity study of post-stroke motor imagery, Time-Varying Effective Connectivity for Describing the Dynamic Brain Networks of Post-stroke Rehabilitation, reports $r = 0.61$, $p = 0.035$ between network global efficiency and Fugl-Meyer motor scores in 12 stroke patients. A standard Fisher-transform interval on that estimate spans roughly 0.06 to 0.88. Hardware error far smaller than this sampling uncertainty is clinically invisible. The chip can therefore run at a voltage and bias point much more aggressive than a bit-exact target allows. The novelty is to make this tolerance explicit, derive an operating-point rule from it, and test whether the rule still pays off after cheap algorithmic simplifications are accounted for.
Core Claim
Suppose hardware approximation adds a per-subject error to each patient's global efficiency (GE). If that error is independent of the patient's true GE, the observed clinical correlation shrinks by the classical attenuation factor
$$ r_{\text{obs}} = r_{\text{true}} \sqrt{\frac{\sigma_{GE}^2}{\sigma_{GE}^2 + \delta^2}}, $$
where $\sigma_{GE}$ is the between-patient standard deviation of GE and $\delta$ is the standard deviation of the hardware-induced error. Take a tolerance of $\delta/\sigma_{GE} = 0.2$. The correlation then loses only about 2% of its value, far inside its confidence interval. We call $\delta/\sigma_{GE}$ the clinical-fidelity budget.
The claim has two parts. First, this budget is loose enough that the energy-optimal operating point for the connectivity pipeline sits in a different place from where a bit-exact target would put it. That point combines near-threshold supply, reduced arithmetic precision, and forward body bias used only to meet the real-time deadline. Second, a body-bias controller that tracks the budget online saves energy beyond what algorithmic simplification alone achieves. Algorithmic simplification here means fewer channels, lower autoregressive order, or coarser windows run on a conventional low-power core.
The attenuation formula only covers random error. Fixed-point rounding and timing faults can instead introduce subject-dependent bias. For example, a patient with sparse, weak connections might be systematically underestimated more than a patient with dense connections. Such bias can change the correlation in either direction. The program measures this directly and does not assume it away.
Research Program
The pipeline under study is the one in Time-Varying Effective Connectivity for Describing the Dynamic Brain Networks of Post-stroke Rehabilitation. It fits a time-varying multivariate adaptive autoregressive model to 64-channel EEG. From that fit it computes the Adaptive Directed Transfer Function, a frequency-domain measure of which channel drives which. It then thresholds the resulting directed graph and computes global efficiency over 0–6 s motor-imagery windows. That paper supplies the workload and the clinical target. It does no hardware work.
The first stage is a software fidelity study. We re-implement the pipeline in parameterized fixed point and sweep word length, autoregressive order, channel count, and window stride. Each configuration is scored on two things. One is $\delta/\sigma_{GE}$. The other is the correlation with Fugl-Meyer scores, including its bootstrap distribution. Bootstrapping over patients separates hardware-induced shifts from resampling noise. This stage also tests the bias concern directly: we regress hardware error on true GE and on graph density.
The second stage builds a hardware energy model. The model is a small RISC-V cluster with a matrix unit in 22 nm or 28 nm fully-depleted silicon-on-insulator (FD-SOI), the process family where body bias has a strong effect. Its structure follows the PULP-style cluster of Performance-aware predictive-model-based on-chip body-bias regulation strategy for an ULP multi-core cluster in 28 nm UTBB FD-SOI. That work regulates body bias with an on-chip PID controller fed by ring-oscillator monitors, and it supplies our control method. Its controller targets a frequency; ours targets a clinical-fidelity budget.
We calibrate the model against published FD-SOI behavior. The Cost of Application-Class Processing reports a 30% speed boost from forward bias at 0.5 V. Arnold reports an 18× leakage reduction from reverse bias at 0.5 V while retaining configuration state. The energy model is built from gate-level simulation across supply and body-bias corners of a commercial FD-SOI library. If a PULP-class FD-SOI board is accessible, we validate on silicon.
Two constraints from prior work shape this stage. Dynamic Body Biasing for Performance and Leakage Optimization shows that excessive reverse bias triggers band-to-band tunneling leakage. The idle-mode bias range must therefore be bounded, not pushed to the rail. The roughly 6-second boot calibration and the temperature-tracking error reported by Di Mauro et al. mean the controller cannot assume perfect knowledge of speed. Its safety margin must be part of the fidelity budget.
The third stage is the controller itself. Each EEG window is a burst of dense matrix work followed by idle time until the next window. The controller chooses supply, body bias, and precision per window. It minimizes energy subject to finishing before the next window and to a running estimate of $\delta/\sigma_{GE}$ staying under budget.
That running estimate comes from shadow computation. On a small random fraction of windows, the chip recomputes GE at full precision and nominal bias. It compares this against the fast result, and the controller tightens or relaxes the operating point accordingly. Between bursts it applies reverse bias. The shadow check is the ingredient that makes the method specific to the clinical target rather than generic dynamic body biasing.
As a secondary, software-only generalization check, we apply the same budget analysis to thresholded functional-connectivity classification on ADHD-200 (Peking site), following Individual classification of ADHD patients by integrating multiscale neuroimaging markers and advanced pattern recognition techniques. That paper reports 76.15% leave-one-out accuracy with $p < 0.008$ feature selection. The check asks whether reduced-precision correlation matrices change which edges survive thresholding. It says nothing about edge hardware, since that workload is offline fMRI.
Baselines
The naive reference point is the floating-point pipeline on a general-purpose processor at nominal voltage with no body bias. It defines the clinical reference, $r = 0.61$ as reported by Xu et al., and the energy ceiling. Every configuration is compared against its GE values patient by patient, on identical EEG epochs.
The strongest algorithm-only competitor is the best simplified pipeline on a conventional low-power core using only dynamic voltage and frequency scaling. The simplifications include reduced channel montage, lower model order, coarser stride, and fixed-point arithmetic. We give it the same search budget over algorithmic knobs as our method gets over the joint space, and hold it to the same fidelity budget. Our method beats it only if it reaches equal $\delta/\sigma_{GE}$ and an overlapping bootstrap correlation distribution at meaningfully lower energy per window. We pre-register "meaningfully lower" as at least 25%. That margin is chosen so it cannot be explained by energy-model calibration error.
The strongest existing hardware method is closed-loop, performance-targeted body-bias regulation as in Di Mauro et al. That work reports a 15% improvement in global energy consumption at 0.7 V and 170 MHz, with up to 2× leakage reduction. The dynamic fine-grain body biasing of Mitigating Parameter Variation with Dynamic Fine-Grain Body Biasing is a related reference, reporting 28–42% leakage reduction at fixed frequency. We run the Di Mauro-style controller on the same simulated cluster, the same workload, and the same deadline, with a bit-exact precision target. The difference between it and our controller isolates the value of targeting clinical fidelity instead of frequency.
No reported result was found on Lacuna for effective-connectivity or ADTF computation on any embedded or body-biased hardware. The energy comparison therefore rests on our own runs of the baselines, not on quoted numbers.
The key ablation removes the shadow-computation feedback. The joint operating point is chosen once, offline, from the stage-one fidelity study, and then frozen. If the frozen point performs as well as the adaptive controller across temperature and aging corners, the online component is unnecessary. The honest conclusion would then be that the contribution is the budget analysis, not the controller.
Why not just…?
Why not just run the pipeline on a commercial microcontroller with standard voltage scaling? This is the most serious threat. A 64-channel EEG at sub-kilohertz sampling with a 0–6 s window is a modest workload, and a simplified pipeline may fit comfortably on an off-the-shelf part. Teodorescu et al. show that adding dynamic body bias to voltage scaling cuts total power a further 21–36% by removing worst-case margins. Near-threshold, Di Mauro et al. show that voltage scaling alone suffers from speed-estimation errors that body bias reduces. Neither result shows that the absolute savings matter for this particular workload. The algorithm-only baseline above is built to settle this. If it lands within 25% of our energy, we report that body bias is unnecessary here and that the fidelity budget alone is the useful result.
Why not just simplify the algorithm — fewer channels, lower model order — and skip hardware work entirely? Algorithmic simplification likely yields the largest single saving, and we expect it to. The question is whether hardware knobs add a further saving on top. The stage-one study maps the algorithmic frontier first. Hardware knobs are credited only for improvement beyond that frontier. There is a real risk here as well. Removing channels may bias GE in a subject-dependent way, for example by dropping lesioned-hemisphere electrodes, so simplification is not free either. The bias regression tests this.
Why not just stream the EEG to a phone or the cloud? For offline studies this is the right answer, and we do not argue against it. The case for local computation is weaker than for images. Arnold and Hyperdrive argue that off-chip data movement dominates energy for dense sensor streams, but raw EEG is low-bandwidth. Local computation matters when neurofeedback must close the loop within a motor-imagery trial. It also matters when the headset must run untethered for a full session. We state this as an application assumption, not a proven need. Our energy model includes a radio-transmission baseline so readers can see where the crossover lies.
Proposed Experiments and Decisive Checks
The first decisive check is whether the fidelity budget is actually loose. On the stroke cohort, we sweep precision and algorithmic knobs and find the cheapest configuration with $\delta/\sigma_{GE} \le 0.2$. We confirm that its bootstrap correlation distribution with Fugl-Meyer scores overlaps the floating-point one. This requires access to data comparable to the Xu et al. cohort (12 stroke patients, 9 controls, 64-channel motor-imagery EEG, Fugl-Meyer scores). Its public availability is not established in our sources. If needed, we will seek it from the authors or collect a matched cohort. If GE turns out fragile, for example if thresholding makes edge counts jump discontinuously under small errors, the premise fails early and cheaply.
The second check tests whether error behaves like noise. If hardware error correlates with true GE or graph density, the attenuation formula does not apply. The budget must then be set empirically on the correlation itself, which needs far more subjects than 12 to estimate reliably.
The third check is the energy comparison per window. It sets our controller against the algorithm-only baseline, the performance-targeted body-bias controller, and the frozen-point ablation. We evaluate across temperature corners and aged-transistor corners, since adaptation should matter most there.
The fourth check is the ADHD-200 edge-survival analysis. It asks whether the budget idea transfers to thresholded correlation matrices, measuring how many selected edges and how much of the 76.15% accuracy survive reduced precision.
A positive outcome is a documented case in which a clinical correlation, not a numerical tolerance, sets a chip's operating point, with savings beyond algorithmic simplification. A negative outcome on the first baseline is still informative: it would show that for this biomarker, the value lies in the statistical budget rather than the silicon.
References
- Time-Varying Effective Connectivity for Describing the Dynamic Brain Networks of Post-stroke Rehabilitation
- Performance-aware predictive-model-based on-chip body-bias regulation strategy for an ULP multi-core cluster in 28 nm UTBB FD-SOI
- The Cost of Application-Class Processing: Energy and Performance Analysis of a Linux-Ready 1.7-GHz 64-Bit RISC-V Core in 22-nm FDSOI Technology
- Arnold: an eFPGA-Augmented RISC-V SoC for Flexible and Low-Power IoT End-Nodes
- Optimal body bias selection for leakage improvement and process compensation over different technology generations
- Individual classification of ADHD patients by integrating multiscale neuroimaging markers and advanced pattern recognition techniques
- Mitigating Parameter Variation with Dynamic Fine-Grain Body Biasing
- Hyperdrive: A Systolically Scalable Binary-Weight CNN Inference Engine for mW IoT End-Nodes