Breath analyzers that use cheap metal-oxide gas sensors fail for a mundane reason: exhaled breath is nearly saturated with water vapor (92–94% relative humidity), and water competes with the target molecule, such as acetone, for the same sites on the sensor surface. Today these sensors are read as a single direct-current (DC) resistance. A single number cannot tell a change caused by water from a change caused by acetone, so people correct it afterwards with a separate humidity sensor and a regression. This proposal reads the same sensor with impedance spectroscopy instead: it measures the electrical response across many frequencies, each of which reflects processes on a different time scale. We then use sparse regression to keep the few frequencies whose response to acetone stays the same whatever the humidity. The selection penalty is the new part. Each frequency is charged according to how much its acetone sensitivity changes across humidity levels, so the model is pushed toward frequencies where water and acetone act through separable mechanisms.
The second contribution is a change of perspective. The usual way to interpret impedance data is to fit a hand-chosen equivalent circuit, a network of resistors and capacitors meant to mirror the material's physics. That choice is a judgment call, and different analysts can make it differently on the same spectrum. The analysts therefore behave like the disagreeing expert raters of diagnostic medicine. We will measure that disagreement directly and ask whether automatic frequency selection removes it without losing accuracy.
Core Claim
Consider a sensor calibrated at moderate humidity and then used at breath humidity. We claim that a sparse model restricted to humidity-invariant impedance frequencies estimates acetone concentration more accurately there than three alternatives: DC resistance with humidity regression, a pooled sparse model that simply adds humidity as a covariate, and analyst-chosen equivalent circuits. We further claim that it does so with a frequency set that is stable across resamples and free of analyst-to-analyst spread.
The test is extrapolation. Every method is trained without the 84% and 94% humidity conditions and evaluated on them. In-distribution accuracy is not the question, because dense models are known to do slightly better there. In the Parkinson's EEG study, Quantitative EEG (QEEG) Measures Differentiate Parkinson's Disease (PD) Patients from Healthy Controls (HC), a random forest on 79 features reached AUC 0.80 against 0.76 for a 6-feature LASSO model.
Why this gap is real
Two breath-sensing papers frame the problem. Assessment of an e-nose performance for the detection of COVID-19 specific biomarkers exposed a 32-sensor DC array to 16–94% relative humidity. Five sensors dominated the response through water alone. The acetone detection limit of the best sensor rose from 63 ppb in dry nitrogen to 208 ppb in a humid breath matrix. Its compression method, principal component analysis, blends the water-dominated channels into every component instead of discarding them. Online Accurate Detection of Breath Acetone Using Metal Oxide Semiconductor Gas Sensor and Diffusive Gas Separation is the strongest practical system in the file. It uses a diffusion reservoir to vent hydrogen and a multiple linear regression on an onboard humidity reading. Against mass spectrometry it reported R = 0.95, Bland–Altman limits of ±321 ppb, and a smallest standard deviation of 158 ppb at a 100-second window. Both papers motivate this proposal: they show that the remaining error sits in how the DC channel handles humidity.
Two papers from impedance work supply the physical premise. They show that frequency sweeps separate bulk, grain-boundary and interfacial processes because each has its own characteristic time constant. Piezoresistivity and AC Impedance Spectroscopy of Cement-Based Sensors does this for moisture-laden cement. Time Evolution of the Skin-Electrode Interface Impedance under Different Skin Treatments does it for a hydrating skin-electrode interface. The cement review also states the constraint we build on: no universal equivalent circuit exists, and an inappropriate choice gives unphysical parameters. That is the opening for the analyst-variability study.
The method draws on two sources. The first is LASSO, regression with an absolute-value penalty that sets most coefficients exactly to zero. The second is the debiasing step of Sparse Relaxed-Lasso Steering, which refits on the selected support with a mild ridge penalty to undo LASSO's shrinkage of true signals. That paper reached a 25-fold reduction in active features in an unrelated domain, generative image editing; we borrow only the estimator.
The multi-rater work informs how we treat analysts. More is more: leveraging multi-rater information for whole slide images grading via virtual expert panel found that ensembling all individual pathologist-trained models beat a consensus model: F1 78.2% versus 72.9%, and expected calibration error 0.050 versus 0.183. The Latent Doctor Model for Modeling Inter-Observer Variability models rater disagreement as a latent variable. Neither applies to physical measurements. We reuse their central lesson, that disagreement should be kept and measured rather than collapsed, and apply it to circuit-fitting analysts.
Research Program
We will build a bench rig around one commercial metal-oxide acetone sensor. It will be read in two ways at once. The first is DC resistance, to reproduce the existing pipelines. The second is a fast multisine impedance measurement from 1 Hz to 1 MHz at about 60 log-spaced frequencies, giving roughly 120 real and imaginary features per spectrum. A multisine excites many frequencies at once, so a full spectrum fits inside the 35–60 second saturation window that the e-nose study identified. We will also take spectra at several time points during each exposure, so the features carry transient kinetics as well as equilibrium values.
Gas will be generated by calibrated mixing in a factorial design. Acetone runs from 0 to 2 ppm, matching the dynamic range of the diffusive-separation benchmark, at relative humidity levels of 16, 30, 50, 70, 84 and 94%, with temperature logged. Later stages add ethanol and hydrogen as cross-interferents, and then real breath samples measured against a mass-spectrometry reference, as in the diffusive-separation study.
The estimator first computes, for each frequency feature $j$, a humidity heterogeneity score $s_j$. This is the variance across humidity levels of that feature's per-level acetone slope, divided by the mean squared slope. It is estimated only from training humidity levels. It captures the multiplicative effect that an additive humidity correction cannot remove: water changes how strongly the sensor responds to acetone, not just its baseline. The estimate $\hat\beta$ then solves
$$\hat\beta = \arg\min_\beta \sum_{h} \frac{1}{n_h}\sum_{i \in h} (y_i - x_i^\top \beta - c_h)^2 + \lambda \sum_j (1 + \gamma s_j)\,|\beta_j|,$$
where the symbols are defined as follows. The index $h$ runs over training humidity levels, and $n_h$ is the number of samples at level $h$. For sample $i$, $y_i$ is the acetone concentration and $x_i$ is its standardized spectral feature vector. The term $c_h$ is a per-level intercept that absorbs the humidity offset. The parameter $\lambda$ sets overall sparsity, and $\gamma$ sets how strongly heterogeneous frequencies are penalized. At test time $c_h$ is predicted from the onboard humidity reading.
Adjacent frequencies are highly collinear, and plain LASSO picks among them arbitrarily. We therefore run stability selection: the fit is repeated over bootstrap resamples drawn in blocks by humidity level, and we keep frequencies selected in a majority of resamples. A relaxed ridge refit on that support follows. Both $\lambda$ and $\gamma$ are tuned by leave-one-humidity-level-out validation within the training range.
The analyst study recruits about six to eight independent impedance practitioners. Each receives the identical spectra and fits the circuit topology they judge appropriate, with whatever fitting tool they prefer. Each analyst's fitted parameters then feed a calibration model for acetone. We will report between-analyst agreement, as an intraclass correlation, both on the parameters and on the final acetone estimates. Our sparse method also contains human choices: the frequency grid, the stability threshold, and the validation scheme. We will report its sensitivity to each, so the comparison of subjectivity is two-sided.
Baselines
The naive reference point is DC resistance calibrated at a single humidity (50%) with no correction, applied unchanged at 84% and 94%. It shows how large the humidity problem is on our sensor.
The strongest existing practical method is the pipeline of Online Accurate Detection of Breath Acetone: DC resistance plus a co-located humidity sensor feeding multiple linear regression, with the same 100-second window. We reimplement it on our sensor and gas data, because its reported figures come from different hardware and 37 breath samples. Those figures are R = 0.95, Bland–Altman ±321 ppb, and SD 158 ppb. We use them as a sanity anchor, not as a number to beat directly. Beating this baseline means narrower Bland–Altman limits and lower root-mean-square error at held-out 84% and 94% humidity, with the gain surviving a paired bootstrap over samples.
The strongest statistical competitor is a pooled LASSO on the full spectrum that includes measured humidity and humidity-by-feature interaction terms as candidate predictors. It gets identical data, the same relaxed refit, the same stability selection, and the same tuning budget, with $\gamma = 0$ as the only difference. This is the ablation that isolates the key ingredient, the heterogeneity-weighted penalty. A second ablation replaces $s_j$ with randomly permuted scores, which tests whether any non-uniform weighting would help.
A dense nonlinear competitor gets the same features and tuning budget: a random forest, and partial least squares regression, the standard spectral-regression method in chemometrics. The EEG study suggests the forest may win in-distribution. Our claim concerns the held-out humidity levels.
The analyst baseline takes two forms: each individual analyst's circuit-based model, and a "full panel" average of all analysts' estimates, mirroring the virtual-panel result. The full panel is the fairest form of the circuit-fitting approach, and our method must match its accuracy while removing its spread.
Why not just…?
Why not just dry the breath with a desiccant and read DC resistance? Drying agents and Nafion tubes tend to strip polar volatiles, such as ketones, alcohols and aldehydes, along with water. The diffusive-separation study avoided desiccants for exactly that reason and for portability. We will run a dried-sample DC arm on the bench, measuring acetone recovery against the mixed concentration. If recovery is near complete and the error is competitive, drying is a real alternative for bench use, and we will say so.
Why not just add a humidity sensor and regress, as already done? This is the main practical threat. It works when humidity shifts the baseline additively. Competitive adsorption is expected to change the acetone sensitivity itself, which a linear correction calibrated at lower humidity extrapolates poorly. The e-nose study's more than threefold rise in detection limit in humid breath is consistent with that. The held-out 84% and 94% test decides the question directly.
Why not just run LASSO on the spectrum with humidity as a covariate? This is the most serious threat to the specific contribution, because it is our method with $\gamma = 0$. If it matches the weighted version at held-out humidity, the honest conclusion is that the gain comes from reading the sensor by impedance, not from the invariance penalty. We would then report the paper that way. Selection stability, measured as Jaccard overlap across humidity-block bootstraps, is a secondary criterion: an interaction-heavy pooled model may predict comparably but choose frequencies inconsistently.
Why not just fit equivalent circuits, as the field does? Circuit parameters are physically interpretable, which sparse frequency bins are not automatically. The cement review, however, concedes that topology choice is non-unique. The analyst study measures what that costs. If analysts largely agree and the panel matches our accuracy, circuit fitting is vindicated for this sensor. We will also check whether the selected frequencies sit near the time constants of fitted circuit elements, which would give the sparse model a physical reading.
Proposed Experiments
Stage one is the factorial bench study, about 20 replicate exposures per humidity-by-concentration cell. All methods train on 16–70% humidity and are tested on 84% and 94%. Metrics are root-mean-square error in ppb, Bland–Altman limits, detection limit at 94% computed as in the e-nose study, the number of selected frequencies, and bootstrap Jaccard stability. The decisive check is whether the weighted method beats both the DC-plus-humidity regression and the $\gamma = 0$ LASSO on held-out humidity. A secondary check is whether its frequencies fall in a band separable from the frequencies that dominate analyte-free humid blanks, which would support the hypothesis that water and acetone act on distinct time scales.
Stage two adds ethanol and hydrogen interferents under the same held-out-humidity split, and checks that the selected set is unchanged or degrades gracefully. Stage three is the analyst study on stage-one spectra, reporting the intraclass correlation of analysts' acetone estimates beside the sensitivity of our method to its own design choices.
Stage four is a small breath cohort with a mass-spectrometry reference. It compares the frozen stage-one models with the reimplemented diffusive-separation pipeline on Bland–Altman agreement. The question is whether the bench advantage transfers to real breath, where humidity, temperature and the mix of other volatiles all vary together.
References
- Quantitative EEG (QEEG) Measures Differentiate Parkinson's Disease (PD) Patients from Healthy Controls (HC)
- Assessment of an e-nose performance for the detection of COVID-19 specific biomarkers
- Online Accurate Detection of Breath Acetone Using Metal Oxide Semiconductor Gas Sensor and Diffusive Gas Separation
- Piezoresistivity and AC Impedance Spectroscopy of Cement-Based Sensors: Basic Concepts, Interpretation, and Perspective
- Time Evolution of the Skin-Electrode Interface Impedance under Different Skin Treatments
- Sparse Relaxed-Lasso Steering: Automatic Sparse Autoencoder Feature Selection for Precise Image Editing
- More is more: leveraging multi-rater information for whole slide images grading via virtual expert panel
- The Latent Doctor Model for Modeling Inter-Observer Variability