Broad-spectrum antibiotics save a patient from one infection while gutting the protective bacterial community that would hold off the next one. Longitudinal 16S ribosomal RNA gene sequencing — reading a barcode region of a universal bacterial gene to census a microbial community without culturing it — shows that this damage persists: after meropenem clears a systemic E. coli infection, 18 genera remain significantly disrupted a month later, with protective Lactobacillus and Bacteroides depleted and opportunists like Clostridioides and Enterococcus elevated (Absence of gut microbiota restoration following meropenem treatment in a systemic infection model). Meanwhile, clinical surveillance flags patients as vulnerable before the pathogen even colonizes them, but only from prescription records and vitals, never from the microbial state itself (Clostridioides difficile infection surveillance in intensive care units and oncology wards using machine learning). And control theory has mature machinery for acting on delayed, noisy epidemic states under hard safety limits — but it has never been wired to a genomic sensor (Safety-Critical Control of Compartmental Epidemiological Models With Measurement Delays).
The proposal: build a closed-loop hospital infection controller in which the epidemic "plant" couples ward-level transmission to per-patient microbiome states estimated from irregularly delayed, compositionally biased 16S amplicon streams, and prove how much sequencing latency, compositional ambiguity, and taxonomic misclassification the safety guarantees can absorb before they break. Nobody has closed this loop; the nearest works are each missing one side of it — the EHR risk model has no molecular input and no actuator, the hospital-network monitoring optimizer assumes perfect instant sensors, and the microbiome studies are purely descriptive.
Core Claim
A safety-critical controller that treats within-host dysbiosis as a dynamic state variable — not a static risk flag — can hold ward-level colonization below a hard capacity limit while reducing unnecessary interventions, provided the sensor pipeline is redesigned around three sequencing-specific distortions: multi-day irregular latency, relative (compositional) rather than absolute abundance, and genus-level ambiguity in short-read data. The claim is quantitative: we will characterize exactly how much each distortion degrades the Input-to-State Safety guarantee that the controller otherwise provides, and show that a two-tier sensing design (cheap short-read screening, full-length confirmation only near the safety boundary) recovers most of the guarantee at a fraction of the sequencing cost.
The design differs from all proximate work. Ötleş et al. built a passive risk score from electronic health records with no genomic input and no actuation; we make the microbial state itself the observed variable and act on it (Clostridioides difficile infection surveillance in intensive care units and oncology wards using machine learning). Adhikari et al. optimized where to place perfect sensors in a 96,281-agent hospital contact network but never modeled control interventions or the microbiome (Fast and near-optimal monitoring for healthcare acquired infection outbreaks); we reuse their network as the simulation substrate and remove the perfect-sensor assumption. Ribeiro et al. documented the post-antibiotic "microbial scar" but controlled nothing (Absence of gut microbiota restoration following meropenem treatment in a systemic infection model); their data calibrates our plant. Ouyang et al.'s Kalman-filtering package smooths delayed molecular surveillance signals but explicitly stops short of any actuator (dlmwwbe: Dynamic Linear Model for Wastewater-Based Epidemiology); we extend that style of state-space smoothing to compositional amplicon vectors and then close the loop.
Research Program
The plant. We extend a compartmental hospital epidemic model in which each patient carries, alongside the usual susceptible/colonized/infected status, a scalar dysbiosis score $d_p(t)$ summarizing the 16S profile (pathobiont over commensal abundance). Transmission to patient $p$ is modulated as
$$\beta_p(t) = \beta_0\,\big(1 - u_p(t)\big)\,\big(1 + \kappa\, d_p(t)\big),$$
where $u_p(t) \in [0,1]$ is the intervention intensity (contact precautions, antibiotic de-escalation, or probiotic/prebiotic restoration) and $\kappa$ governs how strongly ecological vacancy raises susceptibility. This encodes exactly what the observational literature implies: Ötleş et al. show a pre-colonization vulnerability window captured by antibiotic exposure (Clostridioides difficile infection surveillance in intensive care units and oncology wards using machine learning), and Ribeiro et al. show that window lasts at least 30 days with 18 genera disrupted (Absence of gut microbiota restoration following meropenem treatment in a systemic infection model). The dynamics of $d_p$ under meropenem and under restoration interventions are fitted to their public mouse dataset (BioProject PRJNA1139227, V3–V4 fecal profiles at 3 hours, 5 days, and 30 days), so the within-host model is calibrated, not invented.
The sensor. The controller never sees true states. It sees 16S profiles that arrive with irregular latencies of 24 hours to several days, as relative fractions on the unit simplex ($\sum_i x_i = 1$), at whatever taxonomic resolution the chosen amplicon region supports. Three corrections, each grounded in a cited failure mode. First, latency: a dynamic linear observer in the spirit of dlmwwbe: Dynamic Linear Model for Wastewater-Based Epidemiology, extended to multivariate amplicon vectors, produces latent-state estimates from delayed irregular samples. Second, compositionality: a relative surge in Clostridioides can mean absolute proliferation or merely commensal collapse, so the observer jointly estimates an absolute-biomass scale (a load factor shared across samples), and we test both the corrected and the naive version — the naive one is expected to produce the false actuation alarms that mirror the 226-false-positive failure of the EHR model in Ötleş et al. Third, resolution: a two-tier scheme in which routine screening uses short-read V3–V4 profiling, and full-length 16S (which Evaluation of 16S rRNA gene sequencing for species and strain-level microbiome analysis shows resolves nearly all sequences to species, and separates commensal E. coli K-12 from enterohemorrhagic O157:H7) is triggered only when the controller approaches the safety boundary. A control action is never committed on a genus-level readout alone.
The controller. A control barrier function defines the safe set $\mathcal{C} = \{x : h(x) \ge 0\}$ with $h(x) = C_{\max} - \text{expected ward colonizations}$, and a minimum-norm quadratic program picks the least-intensive intervention mix keeping the trajectory inside $\mathcal{C}$, following Safety-Critical Control of Active Interventions for COVID-19 Mitigation. Because the barrier evaluates a predicted state over the sequencing delay window, we adapt the predictor-feedback construction of Safety-Critical Control of Compartmental Epidemiological Models With Measurement Delays, whose Input-to-State Safety result bounds excursion when predictions are imperfect — exactly the tool needed to state how many days of sequencing lag, and how much compositional error, the guarantee tolerates before the safe set must be shrunk. Policy execution constraints follow A Data-Driven Model Predictive Controller to manage epidemics: The case of SARS-CoV-2 in Mauritius: interventions update at most weekly in discrete levels, because a controller commanding hourly probiotic dose changes is not deployable.
Baselines
The naive reference point is guideline-based surveillance with no closed loop: periodic swab screening as evaluated by Ötleş et al., which caught only 23.1% (9/39) of hospital-onset C. difficile cases with a positive predictive value of 9.4%, and blanket staff screening, which per Adhikari et al. needed 8 days to identify an outbreak when testing every nurse. This baseline anchors the "what hospitals do now" comparison.
The strongest existing competitor is the best open-loop pipeline assembled from published components: Adhikari et al.'s HaiEarlyDetect submodular sensor placement on the same hospital network, which identified outbreaks in 6 days at a budget of 1,000 swabs per day (95% detection rate at modest budgets via HaiDetect, preventing an estimated 74% of downstream infections), with interventions triggered by fixed thresholds on the detection output, and patient-level risk from the Ötleş et al. EHR logistic regression (AUROC 73.4%). This is the honest ceiling of what current methods deliver without feedback or molecular input. It will run on the identical simulator, identical agent network, and an identical measurement budget expressed in swab-equivalents per day, with the same tuning effort (grid search over thresholds matched to the grid search over controller gains) so neither side gets a compute or search advantage. Beating it means reducing cumulative ward colonizations below its matched-budget result while holding ward occupancy of colonized patients strictly under $C_{\max}$ with quantified excursions, and raising intervention precision above the EHR model's 3.8% PPV in matched prevalence conditions.
Two ablations isolate the proposal's ingredients. The macro-only ablation runs the identical barrier-function controller with the dysbiosis channel severed ($\kappa = 0$, interventions driven by colonization counts alone), isolating what the 16S state buys. The short-read-only ablation removes the full-length confirmation tier, isolating what taxonomic resolution buys; Evaluation of 16S rRNA gene sequencing for species and strain-level microbiome analysis predicts this matters most near the safety boundary, where acting on a genus-level Escherichia readout cannot distinguish a benign commensal from a virulent strain.
Why not just…?
Why not just use targeted qPCR or diagnostic swabs instead of 16S sequencing? For a single named pathogen, qPCR is faster and gives absolute counts, and we will include it as a fusion channel rather than a rival. But as the sole sensor it is purely reactive: the prospective cohort in Clostridioides difficile infection surveillance in intensive care units and oncology wards using machine learning shows swabs detected only 23.1% of eventual cases, and 8 of the 9 patients the EHR model caught early were swab-negative — the vulnerability window is ecological, occurring before the pathogen is detectable. A qPCR panel cannot measure the Bacteroides/Lactobacillus depletion that constitutes that window (Absence of gut microbiota restoration following meropenem treatment in a systemic infection model). The honest concession: for absolute biomass, qPCR is better than compositional 16S, which is precisely why the proposal explicitly tests whether the observer's absolute-load correction is needed or whether spiking in a qPCR channel is the cheaper fix.
Why not just use the EHR machine-learning model? This is the real threat — it is cheap, already validated, and already demonstrates pre-colonization detection. But it fired 226 false positives at a PPV of 3.8%, and, decisively, the EHR model and the swabs identified almost entirely disjoint patient sets, with the combination lifting sensitivity from 23.1% to 43.6% (Clostridioides difficile infection surveillance in intensive care units and oncology wards using machine learning). Non-redundant information means no single modality suffices. Our design therefore consumes EHR features as observer inputs alongside the 16S stream, and the decisive experiment is a fusion ablation: if EHR-plus-dysbiosis beats EHR-plus-swabs at matched budgets on colonization prevention and alert precision, the molecular channel earns its cost; if not, we will report that the EHR signal already captures the antibiotic-exposure proxy and the proposal collapses to a latency-analysis contribution.
Why not just use static thresholds and guidelines? The evidence is consistent that static policies lose on both sides. Open-loop historical policies on the Mauritius third wave produced 240 deaths and hospitalizations peaking at 525 against a 200-bed limit, while the model predictive controller cut deaths to 140 (a 40% reduction) and held the peak at 190 (A Data-Driven Model Predictive Controller to manage epidemics: The case of SARS-CoV-2 in Mauritius). Blanket nurse testing took 8 days versus 6 for adaptive allocation (Fast and near-optimal monitoring for healthcare acquired infection outbreaks). And stochastic optimal control consistently beat degree-based, uniform, and spectral heuristics on networked spread (Stochastic Optimal Control of Epidemic Processes in Networks). The specific new argument here is that dysbiosis is a slow, predictable state — a 30-day scar — so threshold rules tuned to fast clinical signals systematically mistime interventions on it; the macro-only ablation tests this directly.
Why not just do short-read V4 profiling and skip full-length sequencing? Cost is a fair objection, which is why full-length sequencing is only the on-demand confirmation tier, not the routine channel. But the resolution ceiling is documented, not speculative: V4 fails to classify 56% of sequences at species level, and short regions cannot separate benign from virulent strains within a genus (Evaluation of 16S rRNA gene sequencing for species and strain-level microbiome analysis). In a control loop, that ambiguity translates either into wasted interventions against commensals or into barrier violations from missed pathogens. The short-read-only ablation quantifies this: if the safety violation rate is statistically indistinguishable between the two tiers, the cheap design wins and we will say so.
Proposed Experiments
Three experiments, all simulation-grounded because no closed-loop hospital study exists and none should before the theory is tested. First, the plant calibration: fit $\kappa$ and the restoration-intervention effect on $d_p$ from the PRJNA1139227 longitudinal mouse data, with taxonomy-resolution behavior validated against the Johnson et al. benchmarks (56% V4 species failure; full-length resolution of E. coli K-12 versus O157:H7 — no reported numerical concordance figure beyond qualitative match to metagenomic whole-genome sequencing on Lacuna for that comparison). Second, the closed-loop comparison: run the calibrated plant on the Southwest Virginia hospital network (96,281 agents, 72,146 locations, 200 days) from Fast and near-optimal monitoring for healthcare acquired infection outbreaks), seeding antibiotic-driven dysbiosis trajectories per patient, and compare the dysbiosis-feedback controller, the open-loop submodular-plus-threshold baseline, the macro-only ablation, and guideline surveillance on cumulative colonizations, peak occupied-isolation capacity, intervention count, and alert precision, all at matched measurement budgets. Third, the distortion sweep — the paper's core scientific deliverable: vary sequencing latency from 1 to 7 days with irregular arrival, compositional noise magnitude, and classifier error rate, and measure the excursion of the system beyond $C_{\max}$ against the Input-to-State Safety bound predicted by the predictor-feedback analysis. If measured excursions track the theoretical bound, sequencing latency becomes a designable quantity — a hospital can buy the minimum turnaround its safety margin requires.
Two checks are decisive for the whole program. The latency sweep must show that the safety guarantee degrades gracefully (boundably) rather than catastrophically with realistic 24-hour-to-multi-day amplicon turnaround; if excursions blow past the bound, the compositional observer is mis-specified and the result is a negative one worth publishing. And the dysbiosis-feedback controller must beat the macro-only ablation specifically in scenarios where antibiotics are in heavy use, while matching it when they are not — a signature that isolates the 16S ingredient as the cause rather than generic controller tuning. As a stress test of the compositional correction, we will replay the Ribeiro et al. day-30 profiles through both the naive relative-abundance barrier and the absolute-corrected barrier: the naive version should fire alarms on commensal collapse (the very false-positive pathology quantified at PPV 3.8% by Ötleş et al.), and the corrected version should not. No reported result found on Lacuna for any closed-loop intervention outcome using 16S-derived state variables, which is the gap this proposal fills.
References
- Absence of gut microbiota restoration following meropenem treatment in a systemic infection model
- Clostridioides difficile infection surveillance in intensive care units and oncology wards using machine learning
- Safety-Critical Control of Compartmental Epidemiological Models With Measurement Delays
- Fast and near-optimal monitoring for healthcare acquired infection outbreaks
- dlmwwbe: Dynamic Linear Model for Wastewater-Based Epidemiology
- Evaluation of 16S rRNA gene sequencing for species and strain-level microbiome analysis
- Safety-Critical Control of Active Interventions for COVID-19 Mitigation
- A Data-Driven Model Predictive Controller to manage epidemics: The case of SARS-CoV-2 in Mauritius
- Stochastic Optimal Control of Epidemic Processes in Networks.