Breath Change Scores for Lung Nodule Surveillance

Breath tests for lung cancer look excellent in the hospital where they were built and much worse everywhere else. A large multi-center study measured hundreds of volatile compounds in exhaled breath with a real-time mass spectrometer and reported a training AUC of 0.963. At external hospitals that figure fell to 0.771, and the authors blamed ambient air in the hospital, background contamination, and differences between patient groups (Development and validation of a screening model for lung cancer using machine learning). We propose to stop scoring a single breath sample and instead score how a person's breath changes between visits. Each sample would be measured against a room-air sample taken at the same visit and against the same person's earlier breath. The setting is patients who already return for repeated chest CT scans because a lung nodule was found. In that setting, the breath change score can be combined with the nodule's growth rate measured from the same scans.

The intuition is simple. Most of what breaks transportability is either a hospital-level offset, such as the chemicals in that hospital's air, or a person-level baseline, such as diet, metabolism, or smoking history. Both stay roughly constant from one visit to the next, so both largely cancel when we take a within-person, room-air-referenced difference. A growing tumor does not cancel. The question that matters clinically is whether the change score adds information beyond what CT growth already shows, early enough to change a decision.

Core Claim

We claim two things for patients under nodule surveillance. First, longitudinal breath change scores that are referenced to room air lose much less accuracy across hospitals than single-visit breath levels. Second, fused with CT volume growth, they separate malignant from benign nodules at an earlier surveillance visit than CT growth alone, at matched specificity. The new element is not breath screening, and it is not longitudinal CT. It is treating breath as a within-person trajectory with an explicit environmental reference. That turns the two-visit structure that surveillance already provides into a way to cancel confounders.

How the Four Directions Fit

High-resolution aerosol mass spectrometry is the measurement core. Proton-transfer-reaction time-of-flight mass spectrometry with end-tidal sampling can detect volatile compounds at parts-per-trillion levels without separating them first. High resolution also matters for a second reason. It separates fragments that share a nominal mass but differ in exact mass, as with CO₂⁺ versus N₂O⁺ at m/z 44 in aerosol work. Assigning a peak to the same compound across instruments therefore depends on aligning the mass axis precisely.

Spatial normalization enters as a lens, and once concretely. The lens is that atlas registration maps each subject into shared coordinates so that "the same location" means the same thing for everyone. We treat each hospital's instrument and air in the same way, as a coordinate distortion to remove before patients are compared. The concrete step is a registration of the m/z axis: drift-corrected peak alignment to a shared reference peak list, so that a feature means the same compound at every site. On the CT side, serial scans are registered to each other so that the same nodule is measured consistently. This is the role that affine template alignment plays in Keep Your Friends Close & Enemies Farther, which registers CT volumes to a template so that anatomically matched patches can be identified. We borrow the alignment step, not the contrastive objective.

Longitudinal tracking supplies the signal. Serial measurement acts as a low-pass filter on noisy markers, as shown for transient D-dimer spikes (Lei Zhuang et al.), pulsatile insulin (Kirsten Mehlig), and white-coat blood pressure (Yuanyuan Wang et al.). Kinetic parameters often beat raw levels, as in the log volume change rate (LVCR) of AI-based non-invasive profiling of the tumor immune microenvironment.

Multimodal integration is where the decision is made. We fuse the two trajectories, breath change and CT growth, rather than two single snapshots.

The Breath Change Score

Let $b_{ijt}$ be the intensity of aligned breath feature $j$ for patient $i$ at visit $t$. Let $a_{ijt}$ be the same feature in a room-air sample taken at that visit. The referenced level is

$$z_{ijt} = \log b_{ijt} - \beta_j \log a_{ijt},$$

where $\beta_j$ is a per-feature coefficient. It is fit only on benign-nodule and control visits, and it captures how much room air leaks into end-tidal breath. The change score, written to match the LVCR form, is

$$\Delta_{ij} = \frac{z_{ijt} - z_{ij,t-1}}{\Delta t},$$

with $\Delta t$ the time between visits. Any site- or person-specific additive term in $\log b$ that stays constant across visits drops out of $\Delta_{ij}$. That is the formal basis for expecting a smaller gap between internal and external performance.

The fusion model takes three inputs: the vector of change scores, the nodule LVCR, and baseline nodule size. It uses gradient-boosted trees and a small logistic model, because surveillance cohorts are small and the cited breath study already found XGBoost adequate over 318 features. As a secondary arm, we embed CT with a time-aware transformer, in the style of Time-distance vision transformers in lung cancer diagnosis from longitudinal computed tomography, which handles irregular gaps between scans either by adding a time-gap encoding to patch embeddings or by scaling attention with time distance. We then append the breath change scores to its pooled representation.

Research Program

The file lists no public dataset of serial breath samples, so the program needs new data. We would run a prospective multi-site cohort of patients with indeterminate nodules who are already scheduled for surveillance CT. At each visit, buffered end-tidal breath and a paired room-air sample are taken minutes apart. Outcomes come from pathology or from long-term stability. Hospitals are chosen to differ deliberately in ambient air quality.

Before collecting any data, we run a semi-synthetic stress test on any available single-visit breath data. We inject structured site offsets and person-level offsets of realistic size and confirm that $\Delta_{ij}$ removes them while a single-visit model degrades. This is a sanity check on the mechanism, not evidence of clinical value.

The m/z registration step is validated separately. We measure a shared reference gas mixture on every instrument and report peak-assignment agreement across sites before and after alignment.

Baselines

The naive reference point is nodule size at the first visit with no breath data and no growth term. It is what a clinician has before any follow-up, and every method must beat it.

The strongest existing breath method is the six-compound XGBoost panel from Development and validation of a screening model for lung cancer. Its reported external AUC is 0.771, against 0.963 internally. We refit it on our cohort's first-visit samples using the same feature extraction, the same hyperparameter search budget as our model, and the same leave-one-site-out splits. Our method beats it if two things hold. Its leave-one-site-out AUC is higher, and its gap between internal and external AUC is at most half the panel's gap on our data. Our cohort is restricted to nodule patients, a harder population than cancer versus healthy controls, so we compare within our data rather than against 0.963.

The strongest CT-only competitor is LVCR together with baseline size, plus the time-aware transformer from Time-distance vision transformers. The research file does not include that paper's numbers, so there is no reported result found on Lacuna for TeViT or TaViT on our task. It will be trained on the same scans with the same budget. Beating it means higher sensitivity at 90% specificity at the second visit. It also means a shorter median time to a correct call for malignant nodules.

Three ablations isolate the key ingredients. The first removes the room-air reference by setting $\beta_j = 0$, so only within-person differencing remains. The second removes differencing, so only room-air-referenced single-visit levels remain. The third turns off m/z registration. If either single ingredient alone matches the full score, the claim narrows to that ingredient.

Why not just…?

Why not just apply statistical batch correction to single-visit breath data? Harmonizing feature distributions across sites, for example with per-site location and scale adjustment, is the most credible threat. It needs only one visit, which is cheaper. Its weakness is that it treats site as the only nuisance. It cannot remove person-level baselines, and it needs labeled or balanced data from each new site. Our leave-one-site-out experiment runs it as an additional baseline with site parameters estimated from that site's benign visits only. If it closes the transport gap as well as change scores do, we report that and drop the longitudinal breath claim in favor of the cheaper method.

Why not just subtract the room-air spectrum? This is our second ablation. It removes ambient contamination but leaves differences between people, which the snapshot literature suggests are large for metabolic markers. The comparison between that ablation and the full score settles how much each source of confounding contributes.

Why not just use CT growth, which surveillance already measures? This is a real threat. Growth may already be decisive by the second scan, which would leave breath nothing to add. The proposal is designed around that threat: the primary endpoint is decision timing, not AUC at the final visit. If the fused model does not improve sensitivity at matched specificity at the second visit, the honest conclusion is that breath is redundant for nodule surveillance, and we would say so.

Why not train one end-to-end multimodal network on raw spectra and CT volumes? Surveillance cohorts number in the hundreds. Unconstrained models would likely learn the site and contamination shortcuts we are trying to remove. The time-aware transformer arm with breath scores appended tests a learned fusion while keeping the confounder cancellation explicit.

Decisive Checks

The primary analysis is leave-one-site-out evaluation. We report AUC and sensitivity at 90% specificity for each model, plus the gap between internal and external performance. The claim survives only if three conditions all hold. The full change score shows a smaller external drop than both the six-compound panel and site-harmonized snapshots. Fusion with LVCR beats CT-only models at the second visit. Neither single ablation reproduces the full effect.

A negative control tests whether the cancellation is real. For stable benign nodules, $\Delta_{ij}$ should be centered at zero at every site, while single-visit levels should show site-dependent shifts. If change scores at benign visits still differ by site, then the hospital effects are not constant over time and the cancellation argument fails.