Reading the Host's Rhythms to Tell Infection from Colonization

Modern sequencing finds microbial DNA almost everywhere it looks. A bronchoalveolar lavage run through metagenomic sequencing, which reads all DNA in a sample and matches it to microbial genomes, detects a pathogen far more often than sputum culture. It also detects organisms that are only living quietly on the airway, coating a breathing tube, or left over from an old infection. Deciding whether a positive result means disease is now the hard part of the diagnosis, and current tools answer it with a single snapshot: a read count, a blood marker, or a clinician's tally of fever and oxygen needs.

The patient, however, is continuously monitored. Bedside ECG, pulse oximetry and ventilator waveforms record how the body regulates itself across milliseconds, breaths, minutes and days. We propose to test whether the difference between invasive infection and harmless carriage leaves a scale-specific mark in these signals in the hours before a sample is taken. Specifically, we ask whether infection suppresses slow, minute-to-hour regulatory fluctuations while adding fast beat-to-beat irregularity, a pattern colonization should not produce. If such a mark exists, it should help most where molecular results are least informative: the low-read and culture-positive-but-uncertain cases.

Core Claim

Among patients who already have a positive microbiology result, continuous physiological signals decomposed by time scale carry information about whether that result reflects invasive disease. That information is not contained in read counts, static laboratory values or hourly vital-sign summaries. It is concentrated at identifiable scales, which we hypothesize to be beat-level variability and multi-hour autonomic trend rather than instantaneous heart rate. The new element is the prediction target. Earlier physiological models predicted whether a patient would become septic. Our model is conditioned on microbial detection and asks whether the detected organism is causing the illness. A second new element is a test of which temporal scale does the work, and whether that answer holds regardless of architecture.

Relation to Prior Work

Three studies motivate the gap by showing where snapshot methods break. Sun et al. separated Pneumocystis pneumonia from colonization with a read cutoff of more than 37 reads (AUC 0.964) and a nomogram built from HIV status, lymphocytes, 1,3-β-D-glucan and lactate dehydrogenase (AUC 0.892). Still, 19 confirmed cases fell below the cutoff, and the authors note that read counts do not transfer across sequencing platforms. Chiotos et al. found that in ventilated children, positive tracheal aspirate cultures were not significantly associated with ventilator-associated infection (OR 1.5, 95% CI 0.9–2.3). They also found that the charted signs that prompt cultures did not predict culture results. Ross et al. showed that a 45-transcript blood panel beats procalcitonin for bacterial infection (AUC 0.93 vs 0.83), yet all host assays failed in streptococcal pharyngitis, where carriage is common. All three rely on one time point. None uses the continuous record, and none can see a trajectory.

The sepsis-prediction literature supplies the closest physiological methods and constrains our design. The PhysioNet 2019 challenge and Liao et al. model hourly vitals against the Sepsis-3 label, which combines a culture order, antibiotics and a rise in organ-failure score. That label never checks whether the suspected infection was real, so these models learn clinical suspicion, not invasion. They also degrade across hospitals. We reuse their data and models as a transfer baseline but change the label entirely.

The hierarchical-signal literature supplies our method. HiMAE is a small convolutional U-Net pretrained by masking PPG. Its "resolution hypothesis" holds that different tasks peak at different network depths, and we reuse its per-depth probing to ask which scale separates infection from colonization. Its 10-second windows are too short for multi-hour autonomic change, so we stack it under a longer-horizon encoder. MINA decomposes ECG by frequency band, beat and rhythm (PR-AUC 0.9436 on PhysioNet 2017). We borrow its frequency-then-beat-then-rhythm structure for the ECG channel. Yu et al. showed with a Bayesian hierarchical model that acute stress suppresses minute-to-minute fluctuation while raising second-to-second residual noise. That result gives us a concrete, falsifiable scale signature to look for, and a non-deep model to test against. HEARTS and the heartbeat forecasting study rule out handing the problem to general-purpose large language models or flat time-series foundation models. Those models lose long-range structure, and that loss is exactly where we expect the signal to be.

Research Program

We would build a cohort, not borrow one. The file lists no dataset that pairs waveforms with adjudicated infection-versus-colonization labels. The cohort is ICU patients with a positive lower-respiratory result, from either tracheal aspirate culture or lavage sequencing, and archived bedside monitor waveforms for 48 hours before and 12 hours after sampling. Sampling time is the anchor. Two sub-cohorts target the documented failure points: ventilated patients with positive tracheal cultures (the Chiotos setting) and immunocompromised patients with Pneumocystis reads (the Sun setting). For the second group we record normalized read abundance, and spike-in-calibrated abundance where available.

Labels are the crux, and circularity is the main risk. A clinician's diagnosis already uses fever and oxygen requirement, so a model could score well by rediscovering the chart. The primary label therefore comes from a panel blinded to continuous monitoring. The panel works from histology, quantitative lavage culture, imaging and, where drawn, a host transcript panel. A secondary, outcome-based label is reported separately. We also include a control arm with non-infectious systemic inflammation, such as major surgery, pancreatitis or transfusion reaction, together with colonization. This arm lets us check that the model detects infection rather than inflammation in general.

The signal model has two versions, and the scientific claim does not depend on which one wins. The first is explicitly statistical. Following Yu et al., we fit per-patient variance at each scale $s$ (beat, breath, minute, hour) of heart-rate, pulse-amplitude and respiratory series. From these we compute a decoupling index:

$$D = \log \frac{\sigma^2_{\text{fast}}}{\sigma^2_{\text{slow}}},$$

where $\sigma^2_{\text{fast}}$ is residual variance at beat-to-breath scales and $\sigma^2_{\text{slow}}$ is variance of minute-to-hour fluctuations. Our hypothesis predicts that $D$ is higher in infection than in colonization at matched read abundance.

The second version is a learned hierarchy. A HiMAE-style encoder is pretrained on PPG and ECG segments. A MINA-style encoder handles the ECG channel. A long-horizon convolutional or state-space layer pools segment embeddings over the 48-hour window. Probes are attached at each depth. Both versions output a probability that is fused with log read abundance and the pre-sequencing clinical variables through a simple logistic layer, so the added value of physiology is directly readable.

Baselines

The naive reference point treats every positive result as infection, which is current practice in many units. The majority-class rule is also reported, so that sensitivity and specificity trade-offs are visible. The molecular baseline is the read-abundance cutoff refit on our cohort, following Sun et al., combined with their four-variable nomogram. On their own data this pairing reached AUC 0.964 for reads and 0.892 for the nomogram. Because we expect it to remain very strong overall, "beating it" is defined in the indeterminate stratum. That stratum covers read counts below the cutoff, plus all culture-positive ventilated patients, where the file reports no read-based result. Adding physiology must raise sensitivity at the nomogram's fixed specificity there, with a confidence interval excluding zero improvement.

The strongest physiological competitor is a sepsis-trained temporal convolutional network like Liao et al.'s. That class of model reported AUROC 0.924 on the Sepsis-3 task, a different label, and no reported result on infection-versus-colonization was found on Lacuna. We run it two ways: transferred zero-shot as a sepsis score, and retrained on our label from hourly aggregated vitals. It gets the same patients, splits and hyperparameter budget as our models. Alongside it we run gradient-boosted trees on hourly summaries.

The ablation that isolates the key ingredient keeps the learned hierarchy but restricts it in two ways. In one, it sees only one scale at a time. In the other, it sees the waveforms downsampled to hourly resolution. If the full model does not beat the hourly-resolution version, scale structure is not doing the work. All comparisons use patient-level splits, a held-out hospital where the cohort allows, and equal tuning trials.

Why not just…?

Why not just add procalcitonin or CRP to the culture result? Ross et al. report procalcitonin at AUC 0.83 with 68% sensitivity for bacterial infection. Their host assays, including the stronger transcript panel, failed specifically when carriage was common, which is our target population. Where these markers were drawn, we include them in the fused model. The question is not whether physiology replaces them but whether it adds to them, and the fused-model coefficients answer that.

Why not just use hourly vitals with XGBoost or a temporal network, as in the sepsis challenge? This is the most serious threat, because hourly heart rate, temperature and oxygen saturation already encode fever and oxygen need. It is also the best-developed tooling. Chiotos et al. found that exactly these coarse signs did not track culture results, which suggests that coarse summaries miss the distinction. That is suggestive, not decisive. The hourly-resolution ablation settles it directly. If hourly models match the waveform models in the indeterminate stratum, we will report that the signal lives at coarse scales and that waveforms are unnecessary.

Why not just compute classical multiscale HRV or entropy features and skip deep learning? This could plausibly work as well, and the design accepts that. It is why the first model version is the explicit variance-decoupling model rather than a straw man. The claim under test is about which scales carry information, not about architecture. If the Bayesian index matches the learned hierarchy, the deliverable becomes a cheap, interpretable bedside index. The deep model is only justified if it finds structure the index misses, which the per-depth probes would localize.

Why not just raise read-count thresholds or recalibrate them per platform? Recalibration fixes portability but not the 19 below-threshold Pneumocystis cases, and it has nothing to say about culture-only positives. Our stratified evaluation reports how the recalibrated cutoff performs in the low-read stratum specifically.

Decisive Checks

Four results would decide the question. The first is whether, at matched read abundance, infected and colonized patients differ in $D$ and in the learned scale-specific embeddings. The second is whether that difference is already present 24–48 hours before sampling, before antibiotics or sedation changes can drive it. The third is whether the non-infectious inflammation controls score like colonization rather than like infection. If they score like infection, the model detects systemic stress, not infection, and the claim fails. The fourth is whether performance survives transfer to a held-out hospital, given the site shift documented in PhysioNet 2019. Metrics are AUROC and AUPRC overall and in the indeterminate stratum, sensitivity at fixed specificity, calibration, and estimated antibiotic-days avoided among true colonizers. A null result on the first two checks would be informative in itself: it would indicate that bedside physiology cannot resolve the ambiguity that sequencing created, and that effort belongs with host transcriptomics instead.