Automated drug-dosing systems for shock already exist, but the signals they use to decide whether to give more fluid are only valid in patients on a ventilator. They also depend on an arterial catheter, a bedside monitor and a desktop computer. This note proposes a different signal: the state of the smallest blood vessels near the surface of the body, read by a flat, lensless camera worn on the skin or inside the mouth. A lensless camera replaces the lens with a patterned mask and a computer that undoes the blur. The key move is to skip that "undoing" step entirely. A wearable chip cannot afford to reconstruct an image, and a dosing controller does not need one. What the controller needs is a handful of numbers describing how many capillaries are carrying blood and how fast. We propose to train a tiny estimator that maps raw mask-encoded sensor readings directly to those numbers. The sensor is split into identical tiles that all share one estimator, which makes the computation small enough for a microwatt-scale chip. Because the estimator is so small, it can also be trained across hospitals without moving patient data.
The opportunity matters for two reasons. First, the best closed-loop resuscitation system in the literature reports that its fluid-responsiveness inputs fail in spontaneously breathing patients. Second, systemic blood pressure can look adequate while peripheral tissue is underperfused. If a reconstruction-free microvascular index tracks fluid responsiveness as well as the ventilator-dependent signals do, it would open closed-loop titration to non-intubated patients during transport. If it does not beat a simple pulse-oximeter-style optical sensor, the experiments are designed to show that clearly, and we would stop.
Core Claim
Consider a lensless array in which each tile sits in its own optically isolated chamber behind an identical phase mask. Every tile then sees the tissue through the same forward operator. Write the raw readings of tile $k$ at time $t$ as $y_{k,t} = A x_{k,t} + n_{k,t}$, where $x_{k,t}$ is the unknown patch of tissue, $A$ is the shared mask operator and $n$ is sensor noise. We claim that a microvascular perfusion index can be estimated as
$$\hat p_t = g\Big(\tfrac{1}{K}\sum_{k=1}^{K} h_\theta(y_{k,t-W:t})\Big)$$
where $h_\theta$ is one small network shared by all $K$ tiles and applied to a short window of $W$ frames, and $g$ is a light readout. We further claim that this estimate is accurate enough to stand in for pulse pressure variation and stroke volume variation in a titration loop. The estimator never inverts $A$.
The reason this should work is that perfusion is mostly a temporal statistic: red cells moving through capillaries make the tissue patch flicker. Second-order statistics pass through a linear mask without inversion, because the temporal covariance of $y$ equals $A\,\mathrm{Cov}(x)\,A^\top$ plus noise. Flow-related changes in $\mathrm{Cov}(x)$ therefore remain visible in the raw readings. Image-domain problems such as stitching artifacts at tile boundaries and blur at depth matter far less when the target is a pooled scalar rather than a picture.
Relation to Prior Work
Two lensless designs supply the hardware ideas. The Bio-FlatScope (In vivo imaging with a flat, lensless microscope; In vivo lensless microscopy via a phase mask generating diffraction patterns with high-contrast contours) is a contour-pattern phase mask fabricated on a sensor. It keeps fine detail above the noise floor in low-contrast tissue and resolved capillaries under 20 µm in human oral mucosa. Its reconstruction, however, is offline desktop deconvolution, and its vascular data were purely observational. The Block-wise Lensless Compressive Camera tiles small sensors in isolation chambers behind one coded aperture so that a single sensing matrix is reused. It suffers visible stitching artifacts and was only tested on static images. We borrow the contour mask from the first and the identical, isolated tiles from the second. We differ from both by discarding reconstruction and evaluating on physiological dynamics. Array-camera systems such as Neural array meta-imaging motivate distributed apertures for short optical track, but they still reconstruct images.
The titration side motivates the problem. ReFit (Autonomous precision resuscitation during ground and air transport of an animal hemorrhagic shock model) titrates fluids and vasopressors in 15-minute cycles from mean arterial pressure, heart rate, pulse pressure variation, stroke volume variation and dynamic elastance. Its authors state that the variation-based signals require mechanical ventilation with fixed tidal volume, and a disconnected USB cable froze the system in flight. We reuse ReFit's controller logic and protocol unchanged and replace only its fluid-responsiveness inputs. This isolates the sensing question.
The hardware side constrains the design. The 40 nm arrhythmia accelerator (A 10.60 μW 150 GOPS Mixed-Bit-Width Sparse CNN Accelerator for Life-Threatening Ventricular Arrhythmia Detection) shows that a pruned, low-bit network over 512-sample one-dimensional windows can run at 10.60 μW. The research file notes that two-dimensional image inversion exceeds wearable thermal and energy limits. Our tile-shared estimator is deliberately sized to that accelerator's class of workload. At the block-wise camera's compressive ratio of 0.1, an 8×8 tile yields only about six numbers per frame, so a window of frames is comparable in size to one electrogram window. We do not fabricate silicon. Instead, we report operation counts and simulated energy against that chip's published figures.
The federated side supplies methods. FedHealth: A Federated Transfer Learning Framework for Wearable Healthcare freezes lower layers and personalizes upper layers per user, with second-order feature alignment; it reports a 5.3% absolute gain over a non-personalized federated model on smartphone activity data. FedMetaMed: Federated Meta-Learning for Personalized Medication in Distributed Healthcare Systems shows FedAvg collapsing to 66.49% out-of-distribution accuracy on dermoscopy versus 89.51% for its method. Its two-model client is too heavy for a wearable, and despite its name it never touches dosing. Privacy-Protected Causal Survival Analysis Under Distribution Shift down-weights dissimilar sites with a sparse weighting objective, which we borrow for combining sites. In our setting, skin pigmentation and tissue thickness change the effective operator seen by each patient. The shared $h_\theta$ is a natural object to federate, and the readout $g$ is a natural object to personalize. None of these federated methods has been applied to raw optical measurements or to a control target.
Research Program
The work proceeds in three stages, each of which can stop the project.
Stage one is simulation. We build a forward model of the isolated-tile contour-mask array, using the mask design principles and the roughly 50–80 µm axial resolution reported for the Bio-FlatScope. Synthetic capillary beds are generated with controllable vessel density, red-cell velocity, depth, scattering and pigmentation. We train $h_\theta$ with ground-truth perfusion labels and test whether reconstruction-free estimates match those computed from perfect images. We also characterize where accuracy breaks as depth and scattering grow.
Stage two is a human perturbation study on oral or sublingual mucosa, the site where the Bio-FlatScope already resolved capillaries. A bench lensless array is mounted alongside a ×4 objective that provides a reference capillary density and flow. Healthy volunteers undergo reversible perfusion changes: postural tilt, arm-cuff occlusion and release, and cold exposure. Each volunteer is treated as a federated client, and recruitment spans two or more sites with different skin-type distributions.
Stage three is an offline porcine study embedded in the ReFit hemorrhagic-shock protocol. The lensless array records continuously alongside ReFit's ventilator-dependent signals. Because those animals are ventilated, pulse pressure variation and stroke volume variation are valid there and serve as ground truth for fluid responsiveness at each 5-minute reassessment. Only if the index matches them do we propose a follow-on arm with spontaneously breathing animals, where the index would drive the unchanged ReFit controller.
Baselines
The naive reference point is the raw-pixel temporal variance, which needs no learning: average the per-pixel flicker over the sensor window and map it to perfusion with a one-parameter fit. Thanks to the covariance identity above it is a legitimate inversion-free estimator. If our learned estimator cannot beat it, the learning adds nothing. It runs on identical recordings, and its single parameter is fit on the same training split.
The strongest sensing competitor is a point optical sensor of the multi-wavelength type in A Non-Invasive and Highly Accurate Multi-Wavelength Light Near-Infrared Glucose Sensor Using A Multilevel Metric Learning–Back Propagation Network. That sensor reported 9.98% mean absolute relative difference for glucose, which shows that compact spectral sensors with a learned readout can be accurate. We emulate it by summing all pixels of our array into one channel, so optics, placement and noise are identical, and we give it the same network capacity and tuning budget. No reported result was found on Lacuna for fluid-responsiveness prediction from a point optical sensor. Beating it means statistically lower error against the objective-microscope reference in stage two, and higher discrimination of responders in stage three.
The strongest accuracy competitor is reconstruct-then-measure. We run the Bio-FlatScope reconstruction pipeline (Wiener deconvolution with Tikhonov regularization) offline, then compute capillary density and flow from the image. It is the most accurate thing anyone currently does with these sensors, and we grant it unlimited desktop compute. We count the comparison as won if our estimator comes within a pre-registered margin of its error, at an operation count low enough to sit within the accelerator class of the 10.60 μW chip.
The clinical reference is ReFit's ventilator-dependent inputs. Its reported outcomes in the file are 100% survival in 12 of 12 managed animals and initial stabilization in about 42 minutes during transport. Beating it is not the goal. The goal is agreement at reassessment points: the controller's fluid-give decisions computed from our index should match those computed from pulse pressure and stroke volume variation.
On the federated side, centralized pooling is the upper bound. FedAvg is the lower federated baseline, with FedAvg's 66.49% out-of-distribution figure from FedMetaMed as context rather than a target. A FedHealth-style shared extractor with a personalized head is the strong lightweight competitor. All are trained for the same number of rounds and the same on-device compute.
The key ablation removes tile sharing, training a separate estimator per tile. A second ablation removes temporal windows and uses single frames. Together they test whether the identical-operator structure and the dynamics are what make the small model work.
Why not just…?
Why not just use a pulse-oximeter-style point sensor, since it is cheaper and already wearable? This is the most serious threat, and the pixel-summing baseline answers it directly on identical hardware. The argument for spatial resolution is that a point sensor mixes arterial pulsation, venous pooling and capillary flow into one number. Capillary recruitment, meaning the fraction of vessels carrying blood, is a spatial quantity. If the summed-channel baseline matches the tiled estimator in stages two and three, we will report that and recommend point sensors. The proposal would then reduce to a negative result about imaging for titration.
Why not just reconstruct the image on a phone or edge server and compute features there? Split computation in the style of the federated split-learning work in the file is plausible, and the reconstruct-then-measure baseline bounds its accuracy. Its costs are continuous radio transmission of raw frames, on a device where batteries already exceed half of wearable mass per Lightweight holographic near-eye display system with self-charging capability using solar energy. It also adds a tether failure mode of the same kind that froze ReFit in flight. If on-device estimation loses substantial accuracy relative to reconstruction, split computation wins, and the experiment will show the size of that gap.
Why not just keep ventilator-dependent signals and restrict closed-loop dosing to intubated patients? That works where it applies. ReFit's own report shows the signals become invalid without ventilation, which excludes most transport and ward patients. Stage three only claims value if our index reproduces ReFit's decisions where ground truth exists.
Why not just centralize the training data? Centralized pooling is included as the upper bound. The point of federation here is not accuracy. Optical recordings of mucosa are identifiable health data, and the statistical work cited in the file (distributed result computation) shows decentralized aggregation can match pooled analysis. If federated training falls far below the pooled bound, that gap is itself the finding.
Proposed Experiments and Decisive Checks
Direct titration benchmarks are missing: searches of Lacuna for medication titration, closed-loop glucose, hemodynamic shock and MIMIC datasets returned no results, so the data must be collected. The benchmarks we do anchor to are the Bio-FlatScope mucosa and resolution measurements, the block-wise camera's compressive ratios of 0.1 and 0.05, the ReFit porcine protocol, and the arrhythmia accelerator's power and latency figures.
The first decisive check is in simulation. Perfusion error should rise smoothly with depth and scattering. If the reconstruction-free estimator loses its advantage at exactly the depths where the Bio-FlatScope's resolution degrades, that defines the usable anatomy.
The second check is stage two. Against the ×4 objective reference, we report correlation and mean error for each estimator. Leave-one-site-out evaluation then tests the federated variants on held-out skin types. The pre-registered success condition is that the tiled estimator beats both pixel variance and the summed-channel point sensor, and stays within margin of reconstruct-then-measure.
The third check is stage three. At each ReFit reassessment we compute the area under the curve for predicting fluid responsiveness, as labelled by the ventilator-dependent signals, along with agreement with the controller's actual dosing decisions. We also report robustness during ground and air transport segments, where vibration is the obvious way an optical sensor fails.
Finally, we report operation counts, memory, and simulated energy per estimate, scaled against the 10.60 μW, 35 μs accelerator. The claim of wearability stands or falls on this number, not on a fabricated chip.
References
- In vivo imaging with a flat, lensless microscope
- In vivo lensless microscopy via a phase mask generating diffraction patterns with high-contrast contours
- Block-wise lensless compressive camera
- Neural array meta-imaging
- Autonomous precision resuscitation during ground and air transport of an animal hemorrhagic shock model
- A 10.60 μW 150 GOPS Mixed-Bit-Width Sparse CNN Accelerator for Life-Threatening Ventricular Arrhythmia Detection
- FedHealth: A Federated Transfer Learning Framework for Wearable Healthcare
- FedMetaMed: Federated Meta-Learning for Personalized Medication in Distributed Healthcare Systems
- Privacy-Protected Causal Survival Analysis Under Distribution Shift
- A Non-Invasive and Highly Accurate Multi-Wavelength Light Near-Infrared Glucose Sensor Using A Multilevel Metric Learning–Back Propagation Network
- Lightweight holographic near-eye display system with self-charging capability using solar energy
- A Protocol for the Secure Linking of Registries for HPV Surveillance