Reading Uptake Curves, Not Snapshots: Smartphone Kinetics on Chelating Hydrogel Pads for Field Water Testing

Cheap paper test strips and phone cameras can already measure acidity and common urine markers. They fail at what matters most for drinking water in poor and rural areas: toxic metals at trace concentrations. Paper has almost no capacity to collect a scarce ion, so the color change is too faint to read. A second failure is less obvious. Ordinary users do not wait the prescribed time before photographing the strip, and a single photo taken at the wrong moment gives the wrong answer. Materials chemists, meanwhile, have built composite adsorbents that grab specific metals strongly and sometimes change color when they do, but these are read with benchtop spectrometers.

This note proposes a different test design. A small hydrogel pad loaded with a metal-binding, color-forming agent soaks up the target metal from the water sample. The phone records a short video while the color develops. The concentration is then inferred from the shape of the uptake curve rather than from one photo. Because the reading comes from how fast and how far the color moves, it does not depend on the user's timing, and it can be read before the pad reaches equilibrium. A Delphi panel decides what the device must report and when it must refuse to report. Delphi here means an anonymous, multi-round expert survey with statistical feedback between rounds, and the panel includes water technicians, health workers and chemists. The panel's job is to set action bands, the acceptable miss rate and the maximum assay time.

Core Claim

Treating the pad as a dynamic adsorption system turns the user's timing error from a source of bias into extra signal. If color intensity $S(t)$ is proportional to the amount adsorbed per gram, $q(t)$, and uptake follows the pseudo-second-order form widely used for chelating composites,

$$q(t) = \frac{k\,q_e^2\,t}{1 + k\,q_e\,t},$$

then fitting a video of $S(t)$ recovers both the equilibrium loading $q_e$ and the initial rate $k q_e^2$. Here $t$ is time since immersion and $k$ is a rate constant. The equilibrium loading maps to sample concentration through the pad's adsorption isotherm, and the initial rate rises with concentration under film-diffusion control. Each estimate therefore checks the other, and their disagreement is a built-in flag that something is wrong, for example competing ions or a damaged pad. The resulting hypothesis is that a kinetic reader on a chelating hydrogel pad will match a fixed-time, perfectly timed reader in accuracy and stay accurate under realistic timing errors, while a snapshot reader degrades.

The closest existing work reads static paper strips. The Smartphone Modulated Colorimetric Reader with Color Subtraction linearizes raw camera color using inverse camera response curves from the EMoR dataset, corrects uneven lighting, and subtracts the sample's own pigment from a cellulose pad. We reuse that radiometric pipeline wholesale, but replace the cellulose pad with a pre-concentrating chelating matrix and replace the snapshot with a time series. Smartphone-Based Colorimetric Analysis of Urine Test Strips for At-Home Prenatal Care supplies the motivating failure. Home users did not respect incubation windows, and leukocyte false positives rose. Its pipeline, like every other in the file, treats the pad as a still image.

Research Program

Phase 1: the pad. The pad is constrained by optics as much as by chemistry. That rules out the most popular composites. Biochar-hydrogel composites, reviewed in Recent Advances in Biochar-Based Hydrogel Composites, offer high capacity and easy handling, but biochar is black and would swamp any color change. Nitrogen-doped carbons such as the Cr(VI) sorbent in Two-Step Hydrothermal Reaction Enhances Removal of Cr(VI) have the same problem.

We therefore take only the encapsulation idea from that review. A thin, optically clear alginate or cellulose film carries a chelator. The starting chelator is the tannic-acid chemistry that Zhu et al. showed binds uranyl strongly under acidic conditions, where unmodified materials lose their binding sites. Whether the tannic-acid complex gives enough visible color contrast at trace levels is an assumption to verify first. If it fails, a conventional chromogenic chelator for Cr(VI) is substituted. In either case, batch isotherms and kinetics are measured against laboratory spectrophotometry before any phone is involved.

The pad is designed for a single use. This sidesteps the regeneration and biodegradation problems the review flags. Leaching of unreacted crosslinker is measured, because it is one of the review's named risks.

Phase 2: the reader. The reader runs on the phone and has four steps. It applies the radiometric linearization and color subtraction from the colorimetric reader paper to each frame. It tracks the pad across a 30–120 s video. It fits the uptake curve above to the frames. It then reports a concentration and an internal consistency score from comparing the equilibrium and rate estimates.

The fit is a two-parameter least-squares problem. It avoids the multi-minute latency that limits heavier computational optics such as the diffuser wavefront sensor, which needs about two minutes per measurement.

Phase 3: the Delphi panel. Delphi decides what the device is for, not what the chemistry does. A modified Delphi of three rounds follows the evidence-to-field structure of the NOGAPS project, which needed three rounds to reconcile research protocols with field constraints. The panel includes environmental chemists, municipal water operators, field technicians and community health workers. It rates candidate items such as:

  • the action bands to report (safe, retest, unsafe);
  • the width of the "retest" band in which the reader must abstain;
  • the acceptable false-safe rate;
  • the maximum tolerable assay time;
  • acceptable pad disposal requirements.

Items are retained by content-validity ratio and index, as in VISION-Cog and Wake et al.. The retained items become the reader's operating point: its threshold, its abstention rule and its video-length cap.

We acknowledge the known weakness that Delphi measures agreement, not truth. That is why the panel never rules on chemistry, only on decision costs. Those costs are genuinely a matter of stakeholder judgment.

Baselines

The naive reference is a single photo taken whenever the user chooses. It is read by hue matching against a reference card, as in the prenatal urine-strip study, using the same hydrogel pad. That study reports an average F1 of 0.81 for hue comparison and 0.70 for Euclidean distance on urine analytes. No reported result was found on Lacuna for any smartphone reading of trace metals on an engineered adsorbent pad, so our own naive baseline defines the floor.

The strongest existing method is the full colorimetric-reader pipeline (inverse camera response, lighting correction, color subtraction, fourth-order polynomial regression in a perceptual color space) applied to one frame at the prescribed time. That paper reports roughly a 4-fold reduction in average pH error relative to human reading and earlier mobile readers. It gives that result as a ratio, without an absolute error figure in the file. We run it two ways:

  • Oracle timing: the frame is taken at exactly the prescribed time.
  • Realistic timing: the frame is drawn from a timing-error distribution measured from our own users.

We run it on a commercial paper strip and on our pad, with the same phones, the same samples and the same calibration-set size. Its regression is fit on the same calibration data our kinetic fit uses.

Two ablations isolate the key ingredients. A "pad without kinetics" variant asks how much of the gain comes from pre-concentration. A "kinetics on paper" variant asks how much comes from the time series.

Beating the strongest baseline means two things. Under realistic timing, the kinetic reader must have lower concentration error and lower false-safe rate. Under oracle timing, it must be no worse. A gain under oracle timing alone would not support the claim.

Why not just…?

Why not just use commercial strips with a good phone reader? Paper has negligible capacity for trace extraction, so the color signal at regulatory-relevant concentrations should be near the noise floor. The paper-strip arm of the baselines measures this directly. If calibrated paper reading reaches the panel's action threshold, the composite pad is unnecessary and we will say so.

Why not just add a countdown timer to the app? This is the most serious alternative, because it is trivially cheap. The urine-strip study shows users ignore manufacturer windows even when instructed. Still, an enforced timer that only accepts a photo at the right second could close most of the gap. We include a timer-gated arm in the field study. If it matches the kinetic reader, the remaining case for kinetics rests on two things: shorter assays, by reading the initial rate before equilibrium, and the consistency flag for interfering samples. Both are tested separately.

Why not just use regulatory limits instead of a Delphi? A regulatory limit fixes one threshold. It does not say how wide the abstention band should be, how many false-safe results are tolerable, or how long a field technician will wait. If the panel's thresholds collapse to the regulatory value and its abstention band is negligible, the Delphi contributes little. We will report that outcome, comparing the reader's operating point under Delphi rules and under limit-only rules.

Proposed Experiments

The laboratory study uses spiked tap, well and turbid surface water with competing ions. It spans concentrations bracketing the chosen action threshold. Each sample is split for ICP-MS reference measurement. Capture uses at least four phones, including a premium model, to probe the "high-end phone paradox" reported by the colorimetric-reader authors: premium phones gave lower raw accuracy because their built-in image processing alters color. Each sample is filmed for the full uptake period, so that timing scenarios can be replayed offline from the same video.

Metrics are:

  • limit of detection;
  • mean absolute concentration error against ICP-MS;
  • false-safe and false-unsafe rates at the Delphi bands;
  • abstention rate;
  • error as a function of timing offset and of truncated video length.

The decisive check is the timing-robustness curve. It plots error against capture-time offset for the snapshot reader and the kinetic reader on identical videos. The claim fails if the two curves overlap.

A second check is whether the disagreement between the equilibrium-based and rate-based estimates rises on samples with interfering ions. That determines whether the consistency flag is real.

A small field pilot with non-specialist operators then compares naive, timer-gated and kinetic arms on paired samples. It also records Delphi-item acceptance among technicians.