Auditing Plasma-Activated SERS Substrates by What the Model Listens To

Fast optical tests can read a drop of saliva or urine in minutes. They work by shining a laser on a nanostructured metal surface and recording the enhanced vibrational "fingerprint" of the molecules sitting on it, a technique called surface-enhanced Raman spectroscopy (SERS). A neural network then turns the spectrum into a diagnosis or a concentration. These networks are accurate on held-out test sets, but nobody can easily tell whether they are reading the target molecule or the surrounding biofluid. In one published case, half of a model's distinct reasoning modes turned out to track interfering molecules or measurement artifacts. Meanwhile, the substrate itself is usually treated as a fixed given. Plasma surface activation, which bombards a surface with ionized gas to roughen it and graft polar chemical groups, is a cheap dry process for making substrates wet evenly and hold nanoparticles uniformly. It could make SERS substrates more reproducible. The same polar groups, however, may also grab more proteins and mucins from the sample.

This note proposes to settle that ambiguity with a specific kind of local explanation. For each spectrum, we perturb the amounts of physically meaningful spectral components: the target analyte, known interferents, and the background baseline. We perturb nothing else, and we fit a small piecewise-linear model to the network's response. The coefficients then answer a chemical question directly: how much does this prediction move when the dopamine signal moves, compared with when the serotonin signal moves? We call the resulting number a matrix-leakage score. We use it in two ways. The first is as a seconds-scale, per-sample flag in a rapid test. The second is as the acceptance criterion for choosing plasma treatment conditions. Our central claim is testable: measured on development data, the leakage score predicts which substrate and model combinations will fail when the sample matrix shifts to new donors, and it does so better than contact angle, spectral reproducibility, or in-distribution accuracy.

Core Claim

Two things are new, and they depend on each other. The first is a local surrogate built in the space of physical spectral components rather than individual wavenumber bins. The second is using that surrogate's interferent sensitivity as a design signal for substrate fabrication, not only as an after-the-fact explanation.

Concretely, we fit each spectrum $x$ as $x \approx b + \sum_k c_k s_k + r$. Here $s_k$ are reference spectra of the analyte and known interferents, $c_k$ their fitted amplitudes, $b$ a low-order baseline, and $r$ the leftover residual. We perturb the amplitudes $c_k$ and the baseline coefficients while holding $r$ fixed, so every perturbed spectrum is a physically plausible mixture. We then fit a surrogate that is anchored at the instance and may bend there:

$$g(\Delta c) = f(x_0) + \sum_k \left[ a_k^{+} \max(\Delta c_k, 0) + a_k^{-} \min(\Delta c_k, 0) \right].$$

In this expression, $f$ is the black-box network, $x_0$ the spectrum being explained, and $\Delta c_k$ the perturbation of component $k$. The coefficients $a_k^{\pm}$ are the local slopes on either side of $x_0$. The two slopes capture signal saturation, where one side is flat, which a single linear slope would average away.

The leakage score is

$$L(x_0) = \frac{\sum_{k \in \text{interferents, baseline}} \bar a_k \sigma_k}{\sum_{\text{all } k} \bar a_k \sigma_k},$$

where $\bar a_k$ is the mean absolute slope of component $k$ and $\sigma_k$ is that component's natural spread across donors. Weighting by $\sigma_k$ means we measure how much real-world variation in each component would move the prediction, not just raw sensitivity.

What Exists and What Is Missing

The closest work is Explainable Deep Learning Framework for SERS Bioquantification. It quantifies serotonin in urine with custom saturating CNN heads. It explains them with CRIME, which runs standard LIME on spectra and clusters the explanation vectors with a variational autoencoder. That analysis found six contexts: three driven by serotonin, two by dopamine or epinephrine, and one by an outlier. This paper motivates our proposal, because it shows matrix reliance is real and common. It also exposes a mismatch. The network was built with a saturating output because Raman signals saturate, yet it was explained with a single-slope linear surrogate. Sparse and Faithful Local Explanations with Piecewise Linear Surrogates shows that a linear surrogate averages opposing or saturating slopes toward zero. On Adult data, its instance-anchored two-segment surrogate improves ranking stability by 51% over LIME. We reuse that surrogate form, but move it from tabular features to spectral components.

Our choice of coordinates comes from two sources. The first is LAMP: Extracting Local Decision Surfaces from Large Language Models, which fits local linear surrogates in a chosen coordinate system rather than raw inputs. The second is density-aware work such as Density-Guided Robust Counterfactual Explanations on Tabular Data under Model Multiplicity. That paper shows that unconstrained perturbations drift into regions where the model has no training data. Independent bin masking on spectra has exactly this problem, because it breaks the coupling between the bands of a single molecule.

Two SERS virus papers show the scale and speed of the application, but neither offers any explanation. Multiplex Detection and Quantification of Virus Co-Infections Using Label-free Surface-Enhanced Raman Spectroscopy and Deep Learning Algorithms reaches 95.8% blind-test accuracy in a roughly 15-minute saliva assay. Extracting True Virus SERS Spectra and Augmenting Data for Improved Virus Classification and Quantification separates each spectrum into a concentration-invariant virus signature and background, and reports 91.9% accuracy in saliva. It also explicitly names substrate-to-substrate variability as unsolved. Its decomposition supplies the method we use to estimate components when clean references are unavailable.

On the plasma side, Surface Modification of Ultra-High-Molecular-Weight Polyethylene and Applications: A Review documents large wettability gains from plasma, with contact angles falling from about 149° to 13°. It also documents two costs: over-treatment damages the substrate, and treated surfaces lose their hydrophilicity over time. Oxygen Plasma Modified Carbon Cloth with C=O Zincophilic Sites as a Stable Host for Zinc Metal Anodes shows that oxygen plasma creates uniform carbonyl nucleation sites that make metal deposition more even. That is the property we hope transfers to nanoparticle anchoring. Electrospun Biodegradable Nanofibers Coated Homogenously by Cu Magnetron Sputtering Exhibit Fast Ion Release. Computational and Experimental Study plays a constraining role. Its plasma-anchored copper oxidized by about 55% within two hours of liquid contact, a warning that activated metal–polymer interfaces change chemistry in biofluids. None of these plasma works measures anything spectroscopic or model-related. They judge success by contact angle, bond strength, or electrochemistry.

Research Program

The program has three parts, and the method is developed before any new substrates are made.

The first part is the component-space surrogate. We use analyte and interferent reference spectra where they exist, such as serotonin, dopamine, and epinephrine in water. Where they do not exist, we use decomposition-derived components in the style of the true-virus-spectrum extraction. The perturbation scale for each component is set from its observed spread across donors, not by an arbitrary kernel width. This ties "locality" to physical variability and addresses LIME's known sensitivity to bandwidth. With five to fifteen components, a surrogate needs tens to low hundreds of network queries, not thousands of bin masks. We will measure the actual wall-clock latency on a laptop GPU and on a CPU.

The second part validates the surrogate where the ground truth is known. We train networks under controlled confounding. In training data, an interferent is made to correlate with the analyte, for example dopamine rising with serotonin. The test set breaks that correlation. A good audit must assign high leakage to the confounded model and low leakage to a model trained on decorrelated data, and it must do so from development data alone.

The third part is a plasma fabrication study with the audit in the loop. We use polymer and carbon backings (polypropylene film, PDMS, and carbon cloth) under oxygen plasma across a grid of power and time. Controls are untreated backings and a wet permanganate etch. Gold nanoparticles are deposited on each, and spectra are collected in water and in pooled and individual-donor urine and saliva spiked with analyte and interferents. Every substrate condition is also re-measured after storage for several days to capture the loss of hydrophilicity. For each condition we train the same network and record five quantities: contact angle, spectral coefficient of variation across spots and batches, in-distribution error, leakage score, and error on held-out donors. The honest open question is which way plasma pushes leakage. More uniform hot spots should reduce artifact-driven contexts. More polar groups may attract more matrix proteins and increase interferent signal. The audit is designed to measure that trade-off rather than assume its direction.

Baselines

The naive reference point has two parts: no substrate treatment, and in-distribution accuracy as the only quality signal. We report the error of the untreated substrate with the same network. We also ask how well in-distribution accuracy alone ranks substrate conditions by their error on held-out donors. The leakage score must beat this ranking to justify itself.

The strongest existing quantification result on this question is the sCNN in Explainable Deep Learning Framework for SERS Bioquantification. It reports an MAE of 0.11 µM and a mean percentage error of 3.52% for serotonin in denoised lyophilized urine, against 0.70 µM for partial least squares regression. We will reimplement its denoiser and network heads, train them on our data with the same splits and tuning budget as our model, and use them as the black box being audited. The audit is not meant to lower this error. Beating this baseline means one of two things. Either a plasma condition selected by the leakage score achieves lower held-out-donor error than the condition selected by in-distribution error, at equal fabrication budget, or the error stays comparable while the leakage score falls.

The strongest existing explanation baseline is CRIME from the same paper, meaning standard LIME on wavenumber bins plus clustering. We run it with the same query budget as our surrogate, and also with its native larger budget. We add two further comparisons: gradient saliency and SHAP computed on bins and then summed onto each component's bands. For explanation quality on SERS, no reported result for surrogate fidelity or stability was found on Lacuna, so these numbers will be produced here. Beating CRIME means three things: higher local weighted fidelity on held-out perturbations, better detection of the planted confound (area under the ROC curve), and more stable component rankings across repeated runs, all at equal or fewer queries.

Two ablations isolate the key ingredients. The first keeps component-space perturbation but uses a single linear slope, which isolates the value of the bend under saturation. The second keeps the piecewise surrogate but perturbs bin groups rather than components, which isolates the value of physically plausible perturbations.

Why not just…?

Why not just measure spectral reproducibility and accuracy across substrates? This is the most serious alternative. Coefficient of variation and held-out accuracy are cheap and already standard. They will be reported as competitors, not controls. The reason to expect them to fall short is the CRIME finding. A model can be accurate on its test distribution while half its reasoning tracks dopamine, epinephrine, or artifacts. A substrate that raises matrix signal uniformly can even lower the coefficient of variation. The fabrication study decides this directly. If spectral variation plus in-distribution error ranks conditions by held-out-donor error as well as the leakage score does, measured by Spearman correlation across plasma conditions with bootstrap intervals, then the audit adds nothing for substrate selection. In that case we would report that result and keep only the per-sample flagging use.

Why not just use gradient saliency, which is nearly free? Gradients are local derivatives at a single point. They cannot represent the two-sided saturation that motivated the saturating heads in the SERS quantification paper. On spectra they are also noisy across adjacent bins. We include saliency summed over component bands as a baseline in the confound-detection test. If it matches our surrogate there, the cheaper method should be preferred for deployment.

Why not just use an interpretable physical model, such as partial least squares regression or the true-virus-spectrum decomposition, and skip explanations? Partial least squares regression is interpretable, but it is six times worse in the reported urine result (0.70 versus 0.11 µM). The decomposition approach reaches 91.9% in saliva, below the 95.8% of the multiplex network, though on a different dataset and so not a clean comparison. Accurate deep models are what get deployed, which is why auditing them matters. We still train a decomposition-plus-gradient-boosting model on our data as a transparent reference. If it matches the deep model's held-out-donor error, the case for auditing a black box weakens, and we will say so.

Why not just skip plasma and use wet etching or colloidal films? The wet permanganate arm and the untreated arm are in the fabrication grid for exactly this reason. The polyethylene review reports that wet oxidants degrade the bulk polymer, and the nanofiber work abandoned wet functionalization for delicate substrates. Neither says anything about SERS matrix reliance. That comparison is what we measure.

Decisive Experiments

The first decisive check uses the planted confound, on the neurotransmitter-in-urine setting rebuilt in our laboratory, since no public Raman dataset was found on Lacuna. The leakage score computed on development data must separate confounded from clean models, and it must correlate with each model's error increase on the decorrelated test set more strongly than CRIME, saliency, or SHAP at matched query budgets. If the piecewise bend adds no fidelity over the linear ablation on saturating networks, that part of the method is dropped.

The second decisive check is whether leakage-guided substrate selection transfers. Across the plasma grid and its ageing time points, we rank conditions by leakage on pooled matrix. We then test the top choice against the conditions chosen by contact angle, by spectral variation, and by in-distribution error, on spectra from individual donors never seen in training.

The third check concerns real-time flagging for rapid tests. Per-sample leakage flags should raise the error rate among flagged samples well above that among unflagged ones on the held-out donors, with audit latency reported against the roughly 15-minute assay time as the budget. If plasma raises leakage at every useful setting, the conclusion is still actionable: current plasma recipes trade reproducibility for matrix reliance, and a treatment that blocks protein adsorption is needed.