Skin Spectra Over Time: Tracking Differential Toxic Exposure Without Needles

Heavy-metal contamination in urban soil and household dust is routinely declared "safe" for the average adult, yet the children living in those same rooms ingest contaminated dust hand-to-mouth and end up with roughly four times the health risk of their parents — a child-specific risk score of 0.078 versus a low adult score, measured in the same Wuxi neighborhoods, in the same soil (Study on the risk of soil heavy metal pollution in typical developed cities in eastern China). We cannot actually watch what this chronic, stratified exposure does to individual children, because the tools that connect external contamination to internal biology — serial blood draws for lead, cadmium, and inflammatory markers — are too invasive and too burdensome for repeated pediatric screening. Meanwhile, two adjacent literatures have matured in isolation. Hyperspectral imaging — cameras that record a full reflectance spectrum (dozens of contiguous wavelength bands) at every pixel, rather than three broad RGB colors — can already read microcirculatory oxygenation, hemoglobin state, and tissue water content through skin, label-free, in seconds (Hyperspectral imaging for perioperative monitoring of microcirculatory tissue oxygenation and tissue water content in pancreatic surgery — an observational clinical pilot study). And trajectory modeling has shown that the rate and direction of a biomarker, not its value at one moment, carries the prognostic signal: in-hospital albumin trajectories predict survival after myocardial infarction where the admission value fails (Association between albumin changes and prognosis in older patients with acute myocardial infarction), and inflammatory trajectories leave a "legacy" of elevated metabolic risk even while actively improving (Association Between High-Sensitivity C-Reactive Protein Trajectories and the Incidence of Metabolic Syndrome). Nobody has joined these: every medical hyperspectral study is a single-timepoint snapshot, and every exposure-risk study stops at environmental chemistry.

This proposal builds the missing link: a longitudinal cohort study in a heavy-metal-impacted urban district, in which children's skin spectra are measured repeatedly and cheaply, modeled as continuous-time trajectories, and validated against sparse blood-metal panels. We do not claim skin optics can read lead concentrations directly. The claim is narrower and testable: chronic metal exposure produces subclinical hematological and microvascular sequelae — anemia-driven hemoglobin and oxygenation shifts, perfusion and inflammation changes — that live precisely in the visible and near-infrared absorption bands hyperspectral imaging already resolves, and that a trajectory model can separate from normal growth.

Core Claim

Repeated low-cost spectral measurements of skin, reconstructed from smartphone RGB images using learned spectral priors and calibrated by periodic clinic-grade pushbroom scans, when modeled with an additive Gaussian-process latent trajectory model, will (a) stratify children by internal metal burden more accurately than area-level environmental hazard scores, and (b) predict which children's blood lead or hemoglobin will worsen by the next visit more accurately than the strongest existing alternative — a single-snapshot hyperspectral classifier of the kind that already works for ICU sepsis. The mechanistic novelty is applying trajectory decomposition, previously tested only on scalar clinical scores and brain MRI, to cutaneous spectral time series; the clinical novelty is the first non-invasive longitudinal biomarker channel for differential toxic exposure.

Research Program

The study runs in a city with documented soil contamination, recruiting roughly 120 children aged 2–10 plus one caregiver per household (the adult contrast arm mirrors the child-versus-adult design of the Wuxi and Kathmandu studies, which found ingestion-driven child risk multiples of adult risk from identical environments — Study on the risk of soil heavy metal pollution in typical developed cities in eastern China, Non-carcinogenic and Carcinogenic Risk Assessment of Heavy Metals Exposure to Indoor Dust in Kathmandu, Nepal). Each household's exposure stratum is fixed by dust and soil sampling with the same atomic-absorption chemistry those studies used, deliberately avoiding the mismatched foreign default exposure parameters the Kathmandu authors flag as a source of distorted intake estimates. Every quarter, caregivers capture palm RGB images at home; every six months, each child visits the clinic for a pushbroom hyperspectral scan under the controlled protocol of the Hyper-Skin dataset (Specim FX10, 400–1000 nm, chin-rest-free adaptation for children) plus a venous blood panel: lead, cadmium, hemoglobin, and high-sensitivity C-reactive protein. The palm is the measurement site because the ICU sepsis study found it consistently outperformed the finger for both diagnosis and mortality prediction (AI-powered skin spectral imaging enables instant sepsis diagnosis and outcome prediction in critically ill patients). Blood draws serve as sparse ground-truth labels, not as the monitoring channel — the attrition arithmetic of serial invasive follow-up is brutal, with one inflammation cohort collapsing from 58,994 candidates to 4,312 analyzable participants (Association Between High-Sensitivity C-Reactive Protein Trajectories and the Incidence of Metabolic Syndrome).

The sensing pipeline has three stages. First, an RGB-to-spectrum reconstruction network is initialized from Hyper-Skin, the 330-cube, 51-subject facial skin dataset with 28 camera response functions, which established both that reconstruction priors transfer badly from natural scenes to skin and that skin-specific training fixes this (spectral angle error dropping from roughly 0.60 to 0.10) (Hyper-Skin: A Hyperspectral Dataset for Reconstructing Facial Skin-Spectra from RGB Images). We fine-tune it on our own multi-phototype pediatric palm data, since Hyper-Skin is single-session and demographically narrow. Second, reconstructed spectra pass through a physiological encoder that maps each scan to a small vector of tissue parameters — oxygenation, perfusion and hemoglobin-derivative proxies, tissue water — the same quantities the perioperative and trauma hyperspectral studies extract from 500–1000 nm reflectance (Hyperspectral imaging for perioperative monitoring of microcirculatory tissue oxygenation and tissue water content in pancreatic surgery, Implementation of hyperspectral imaging in a trauma resuscitation room: a randomized controlled trial). Third, and this is the key ingredient, the parameter streams are modeled as continuous-time trajectories with an explicit decomposition:

$$z_i(t) = g(t) + u_i(t),$$

where $g$ is a population development trend shared by all children, $u_i$ is child $i$'s individual deviation, and $u_i$ carries an additive Gaussian-process prior over time. This is the construction of the Longitudinal Variational Autoencoder, which demonstrated on Health MNIST that an additive GP latent prior can pull apart two co-occurring temporal processes — aging versus disease — that an ordinary independent-prior model entangles (Longitudinal Variational Autoencoder). Our analogue: pull apart growth from exposure-driven deviation. Irregular quarterly-to-semiannual visit spacing and missed appointments are handled either by this GP formulation or, for trajectory forecasting, by interpolative multi-marginal flow matching, which generates smooth individual paths between irregularly spaced clinical observations and lifted a downstream disease classifier from 71.7% to 80.8% accuracy when used to extend Alzheimer's trajectories from 18 to 36 months (Longitudinal Flow Matching for Trajectory Modeling). A serious engineering risk is that smartphone cameras and lighting drift across two years of home sessions; we model sensor and calibration drift explicitly using the hyperspectral-sensor-shifts library (vvaquet/hyperspectral-sensor-shifts), and the semiannual clinic pushbroom scan doubles as a per-child calibration anchor.

Baselines

The do-nothing reference point is the regulatory-style environmental hazard assessment: apply the US EPA ingestion and dermal risk equations to our own household dust and soil measurements, exactly as the Wuxi study did, producing a per-child hazard index from environmental concentrations and default intake factors. This is the incumbent method in the field, and the Wuxi result — child risk 0.078, four times the adult value in the same soil — shows what it can and cannot do: it stratifies zones, not individuals. It will be run on identical environmental data, and the proposal beats it if individual-level prediction of blood-lead elevation from skin trajectories achieves materially higher AUROC against the blood panel than the hazard-index strata alone. The strongest existing method is the single-snapshot hyperspectral classifier: a ResNet14d convolutional network reading one palm hypercube, the architecture that diagnosed sepsis in the Heidelberg ICU cohort with an AUROC of 0.80 from palm hyperspectral imaging alone (rising to 0.94 combined with clinical data within 10 hours) (AI-powered skin spectral imaging enables instant sepsis diagnosis and outcome prediction in critically ill patients). We retrain and tune it on our first-visit spectra with the same nested cross-validation protocol and the same compute and hyperparameter budget as the trajectory model, tasking it with predicting which children will show top-quartile blood lead at the next draw. Beating it requires higher AUROC with non-overlapping confidence intervals, plus measurably earlier flagging — the snapshot model can only react to physiology that has already changed. Finally, two ablations isolate the proposal's ingredients: the same trajectory pipeline with an independent-prior VAE latent (no $g$/$u_i$ decomposition) tests whether disentangling growth from exposure deviation is what carries the gain, and the same trajectory pipeline run on raw RGB features rather than reconstructed spectra tests whether the spectral channel matters at all.

Why not just…?

Why not just use blood tests? They are the gold standard, and we use them — as semiannual labels. The alternative is quarterly venipuncture on toddlers, which the cohort-maintenance record of the hsCRP study (58,994 screened, 4,312 retained) suggests is optimistic even for consenting adults with annual visits. The proposal's value proposition is not replacing blood chemistry but extending its temporal resolution roughly sixfold and enabling compliance-sensitive cohorts the invasive route cannot reach at all.

Why not just use smartphone RGB pictures with a good vision model? Three-channel cameras collapse chemically distinct states to identical colors (the metamerism problem that motivates hyperspectral imaging in surgery and forensics), and they are structurally blind to the 700–1000 nm band where perfusion, water, and deep dermal structure live. Empirically, the Hyper-Skin authors found models without skin-specific spectral priors performed poorly on skin, with spectral angle error near 0.60, while skin-trained reconstruction brought it to about 0.10 (Hyper-Skin). Still, this is a real threat to the cheap-sensor design: if RGB reconstruction error drifts over sessions in a child-specific way, it could masquerade as a physiological trajectory. The reconstruction-stability experiment below measures exactly this, and if drift cannot be corrected by the semiannual pushbroom anchors, we say so and fall back to clinic-based scan sequences.

Why not just take one hyperspectral snapshot, as the sepsis study did? The Heidelberg result (AUROC 0.80 from a single palm scan) is genuinely strong, and we treat it as the principal competitor rather than arguing it away. But the trajectory literature gives concrete reasons to expect a snapshot to miss exposure biology: a "legacy effect" in which children whose inflammation is actively improving still carry a 40.8% elevated metabolic-syndrome hazard (HR 1.408) invisible at any single timepoint (Association Between High-Sensitivity C-Reactive Protein Trajectories and the Incidence of Metabolic Syndrome), and admission-value failures that serial albumin trajectories rescue (Association between albumin changes and prognosis in older patients with acute myocardial infarction). The head-to-head experiment settles whether that transfers to skin spectra; if the snapshot matches the trajectory model, the longitudinal apparatus is unjustified and we will report that.

Why not just rely on regulatory thresholds and hazard indices? Because they are calibrated to a standardized average adult and are zone-level, not person-level. Children in Wuxi faced four times adult risk at adult-safe concentrations; compliant commercial shellfish still carried acute neurotoxic pediatric doses (An Occurrence and Exposure Assessment of Paralytic Shellfish Toxins from Shellfish in Zhejiang Province, China); and adult-calibrated tolerable daily intakes fail infants outright (Exposure assessment and risk characterization of n-nitrosodimethylamine (ndma) in the diet of children from 6 to 36 months in Hanoi, Vietnam). Our do-nothing baseline makes this comparison quantitative rather than rhetorical.

Proposed Experiments and Decisive Checks

The first experiment is a signal-existence pilot before any cohort investment: compare skin spectral tissue parameters between adults with documented occupational metal exposure and matched controls, against concurrent blood panels. If hemoglobin-linked and perfusion spectral signatures show no coherent relationship to internal dose here — where effects should be largest — the physiological proxy hypothesis fails and the project should stop. The second experiment measures reconstruction fidelity and its temporal behavior: fine-tune the Hyper-Skin-initialized reconstruction network on our multi-phototype pediatric data and report spectral angle and structural similarity against paired clinic pushbroom scans, with skin-optimized Hyper-Skin results (SAM about 0.10, SSIM about 0.95–0.97) as the reference bar, and specifically test whether reconstruction error is stationary across sessions or drifts. The third and decisive experiment is the three-way prediction contest at identical compute and tuning budget: hazard-index strata, single-snapshot ResNet14d, and the GP-decomposed trajectory model, each predicting next-visit blood-lead elevation, with AUROC and time-to-flag as primary metrics. The fourth experiment tests the core latent claim directly: after regressing out age and body size, the individual deviation trajectory $u_i(t)$ should correlate with household exposure stratum, and children who relocate out of contaminated housing should show whether spectral trajectories exhibit the same persistence the hsCRP study found — a legacy signature that no snapshot method could detect by construction. The environmental hypersensing literature shows reflectance spectroscopy can already quantify the hazard side of this chain in soil (An Improved Gradient Boosting Regression Tree Estimation Model for Soil Heavy Metal (Arsenic) Pollution Monitoring Using Hyperspectral Remote Sensing); this program closes the loop by reading the receptor side, longitudinally, in the children the regulatory average was never built to protect.