Within-Patient Reference Spectra for Terahertz Tumor Genotyping Across Patients, Sites, and Ancestries

Terahertz time-domain spectroscopy measures how tissue slows and absorbs very-low-energy light, between roughly 0.1 and 10 terahertz. Signals in this range track water content, packing density, and weak bonds between molecules. A recent study used it to predict whether a glioblastoma carries a TERT promoter mutation in minutes rather than the roughly ten days sequencing takes. It reported an area under the ROC curve (AUC) of 0.908. However, those 180 tissue slices came from only 9 patients, and the validation set was drawn at the slice level. Genetic epidemiology has learned the hard way that a biomarker rule tuned on one homogeneous group can fail silently in another. That field's standard fix for population confounding is to compare people against their own relatives, not against strangers.

We propose to carry that fix into terahertz diagnostics. Each tumor spectrum would be read relative to a reference spectrum of normal-appearing tissue from the same patient, prepared in the same session. The classifier would then be audited for failures across patient, hospital site, and genetically estimated ancestry, with the patient treated as the unit of generalization. The claim is concrete. Much of the between-patient variation in dielectric properties is baseline variation that has nothing to do with the mutation, such as hydration, section handling, and tissue architecture. A slice-level classifier can partly memorize this baseline. Subtracting a within-patient reference should remove most of it. What survives is a signal whose transfer across sites and ancestries we can test directly rather than assume.

Core Claim and What Is New

The motivating paper is Rapid diagnosis of TERT promoter mutation using Terahertz absorption spectroscopy in glioblastoma. It extracted six dielectric parameters over 0.2–1.4 THz, reduced 492 spectral points to 14 with LASSO, and classified with a random forest. Refractive index and dielectric constant carried the discrimination; the absorption coefficient did not. That detail matters. Refractive index and dielectric constant are exactly the quantities most sensitive to tissue density and water content, which also vary from patient to patient and site to site. Mutation status is a per-patient label, so the effective sample size in that study is nine, not 180. The reported AUC therefore cannot separate "detects the mutation's tissue consequences" from "recognizes which patient a slice came from."

The genetics literature supplies the lens. Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing showed a concrete failure of a phenotype threshold tuned on White patients (BMI < 30). Background disease shifts the phenotype distribution in South Asian and African-Caribbean patients, so the threshold misses cases in those groups. Group-specific thresholds doubled detection from 8.8% to 16%. We borrow two things from that literature. The first is the diagnostic question: does a fixed decision threshold keep its meaning across groups? The second is the remedy used when group membership is only a stand-in for background differences: condition on the individual's own baseline, as within-family genetic designs do.

No paper in our search links terahertz spectroscopy to ancestry or population-level generalization. The closest work is the single-center, slice-split TERT study. We differ from it in three ways. We make the patient the unit of evaluation. We add a same-patient reference measurement as the key ingredient. We run a pre-registered audit that separates site effects from ancestry effects.

Research Program

We will build a multi-center frozen-section cohort of glioblastoma patients with routine TERT promoter sequencing, targeting three sites and on the order of several dozen patients per site. For each patient we scan tumor sections and matched normal-appearing or peritumoral sections using the same 100 µm protocol and the same six dielectric parameters as the motivating study. Germline DNA, already collected alongside tumor sequencing, gives genetically estimated ancestry proportions. Admixture-aware tools such as XGMix: Local-Ancestry Inference with Stacked XGBoost could supply this; they report 97.98% accuracy on simulated three-way continental admixture. We use these proportions only as audit covariates, never as model inputs to the primary classifier.

The method is deliberately simple. For each dielectric parameter at each frequency, we express the tumor value as its difference from the patient's reference value. We then fit the same LASSO-plus-random-forest pipeline on the differences. The reference replaces generic spectral normalizations such as standard normal variate scaling, which are surveyed in Signal Processing and Machine Learning Techniques for Terahertz Sensing: An Overview. Those normalizations rescale each spectrum by its own statistics and cannot tell a patient's baseline apart from a tumor effect.

The analysis has three parts. First, we measure how much of the raw spectral variance is explained by patient, by site, and by mutation status. Second, we test whether a classifier can identify the patient, the site, or the ancestry group from reference tissue alone. Any such ability is a nuisance signal that a raw classifier could exploit. Third, we test whether classifier errors depend on ancestry after adjusting for site.

Baselines

The naive reference point is a classifier that predicts the training-set mutation prevalence for every patient. It fixes the AUC floor at 0.5 and gives a calibration floor. We also run a "patient-fingerprint" control: the raw pipeline trained on reference tissue only, which contains no tumor. Any above-chance AUC it achieves under slice-level splits is direct evidence of leakage.

The strongest existing method is the LASSO-plus-random-forest pipeline of Rapid diagnosis of TERT promoter mutation using Terahertz absorption spectroscopy in glioblastoma. It reports AUC 0.908 and specificity 1.0 on a 55-slice validation set. We reimplement it on our data with the same six parameters, frequency band, and LASSO feature budget. We evaluate it first under slice-level splits, to reproduce the published setting, and then under leave-patient-out and leave-site-out splits. The within-patient reference model must use identical features, tuning grids, and nested cross-validation budget. Beating this baseline means one of two things under leave-site-out evaluation. Either the reference model achieves higher patient-level AUC with non-overlapping bootstrap intervals, or it matches that AUC while clearly reducing worst-group AUC gaps and calibration error.

We add a deep-learning competitor following Discrimination of Wheat Varieties by Terahertz Time-Domain Spectroscopy and Convolutional Neural Network. That paper reshaped spectra into 32×32 matrices, and its seven-layer network reached 97.8% accuracy against 90.6% for LS-SVM. Those numbers come from wheat cultivars, not patients. We include this model because high-capacity networks are the most likely to memorize patient baselines. It gets the same tuning budget. We also include a pooled model given ancestry and site as extra inputs, and a group-specific-threshold model in the style of Jones et al. Finally, the key ablation keeps everything in our model fixed except the reference. The reference spectrum is replaced by a population-average reference spectrum, so any gain can be attributed to using the patient's own baseline.

Why not just…?

Why not just split by patient and keep the original pipeline? Patient-level splitting is necessary, and we do it for every model. But it only exposes the leakage; it does not remove the baseline variation driving it. If the patient-split raw pipeline matches the reference model under leave-site-out evaluation, the reference adds nothing and we will report that. The population-average-reference ablation is designed to settle this.

Why not just collect a larger, more diverse cohort and pool it? This is the real threat. Pooling many patients may let a flexible model learn to ignore baseline variation on its own. It mirrors the genetics argument that representation alone fixes transfer. Against it, Jones et al. found that a single pooled threshold was still misaligned even in a cohort of 1,706 people. That is evidence, though from another domain, that pooling does not guarantee a consistent decision rule. We test pooling head-on by training on two sites and testing on the third. We run this at several training-set sizes, to see whether the pooled model's gap to the reference model closes as data grow.

Why not just use group-specific thresholds, as in monogenic diabetes? Group-specific thresholds need enough patients per group per site, and they treat ancestry labels as the cause of the shift. If the shift is really hydration or processing, ancestry is only a proxy. The per-patient reference is the individual-level version of the same correction. We compare the two directly. If group thresholds match the reference model, ancestry-linked baseline differences are real, and that itself would be a publishable finding.

Why not just sequence the tumor? Sequencing stays the ground truth. The terahertz model is only useful intraoperatively, where roughly ten days of turnaround is too slow. That use makes silent failure in underrepresented groups more dangerous, not less.

Proposed Experiments and Decisive Checks

The primary metric is patient-level AUC under leave-site-out evaluation. Secondary metrics are expected calibration error, sensitivity at the specificity chosen on training sites, and worst-ancestry-group AUC. All come with patient-level bootstrap intervals.

The first decisive check is the slice-versus-patient gap for the published pipeline. A large drop would show that the 0.908 figure reflects patient identity. The second is the fingerprint test. If reference tissue alone predicts site or patient identity well above chance, and the within-patient differences do not, the mechanism is confirmed.

The third check is the ancestry audit. We fit a logistic model of per-patient error on ancestry proportions, adjusting for site. If there is no ancestry effect after the reference correction, while the raw model shows one, the reference has absorbed the confounding. If an effect persists after correction, ancestry-associated tissue differences exist that no baseline subtraction removes, and group-aware calibration becomes necessary.

We will report a negative outcome as prominently as a positive one: the reference model fails to beat the patient-split raw pipeline and pooling closes the gap. For practitioners, that outcome would mean the fix is simply more patients and honest splits.