When to Switch the Last-Resort Antibiotic: A Competing-Risks Hazard Model of Delayed Nephrotoxicity

When a patient carries a highly resistant infection, clinicians reach for last-resort antibiotics — colistin, aminoglycosides, vancomycin combinations — that quietly damage the kidneys over days of accumulated exposure. The same prolonged dosing that endangers the kidney also hands the bacterium more time under selective pressure, including the inducible resistance documented for colistin. Hospital software today counts how many days of therapy were given, months after the fact; it does not tell a clinician on day 4 whether this particular patient's kidney risk is about to spike. And the standard statistical treatment of the problem — "did kidney injury happen by day 30, yes or no" — throws away exactly the information that would make a switching decision possible: when the risk rises. This proposal builds a survival model — one that estimates an instantaneous risk of an event per unit time rather than a single binary outcome — with two competing clocks running simultaneously: the clock of the untreated resistant infection and the clock of treatment-induced organ injury, both driven by the patient's own cumulative drug exposure so far. The clinical output is a specific switching time: the day on which the predicted hazard of continuing the drug overtakes the predicted hazard of stopping it.

Core Claim

The claim is that delayed antibiotic nephrotoxicity is a non-proportional, cumulative-exposure hazard that can be predicted day-by-day from routine intensive-care records, and that this prediction converts a vague stewardship exhortation ("de-escalate earlier") into a computable stopping rule. Three facts from the literature make this both possible and necessary. First, the best existing clinical analysis of vancomycin/piperacillin-tazobactam kidney injury, a retrospective cohort of 350 inpatients by Munshi et al., found a 25.4% rate of acute kidney injury and a significant association with therapy duration (4.65 days in the kidney-injury group versus 4.04 days without, p=0.01) — but used static logistic regression, which cannot say whether cutting therapy from five days to three would prevent injury, because it discards onset timing and treats early discharges as non-events P-1654. Second, the machinery for time-to-event modeling of delayed toxicity exists and works: Lei and Zhu embedded a Cox proportional hazards model into Bayesian dose-finding and, in their simulation Scenario 3, identified the correct maximum tolerated dose 74% of the time versus 72.9% for a logistic model and 51.3% for a hyperbolic-tangent model, while terminating uniformly toxic regimens early 100% of the time versus roughly 75% for standard designs A novel model of the continual reassessment method in Phase I trial — but it has only ever been tested on synthetic dose skeletons, never on real hospital data, never with multiple drugs, and never with a pathogen in the picture. Third, the naive fix — a standard Cox model with the proportional hazards assumption — is precisely the assumption that cumulative toxicity violates, since kidney hazard is near zero early and then climbs as tubular exposure crosses a threshold; Nagpal et al. showed that on data with crossing hazard trajectories the standard Cox model has an Expected Calibration Error of 0.0992 on the FLCHAIN cohort while their latent-mixture extension, Deep Cox Mixtures, reached 0.0200 Deep Cox Mixtures for Survival Regression. The gap this proposal fills is the one no paper occupies: a time-to-event model of delayed antibiotic toxicity, fitted to real intensive-care records, with the infection itself as a competing risk and cumulative exposure as a time-varying driver.

Research Program

The model, which we call the dual-clock hazard model, has two cause-specific hazard functions fitted jointly. For the toxicity clock we use a latent-mixture-of-Cox architecture in the style of Deep Cox Mixtures Deep Cox Mixtures for Survival Regression, because the population of patients on last-resort therapy is heterogeneous — some clear the drug fast, some accumulate it — and a mixture of component hazards handles the crossing and non-proportional trajectories that a single Cox model cannot:

$$h_{\mathrm{tox}}\big(t \mid z(t)\big) = \sum_{k=1}^{K} \pi_k\big(z(t)\big)\, h_{0,k}(t)\, \exp\!\big(\beta_k^{\top} z(t)\big)$$

Here $z(t)$ is a vector of time-varying covariates updated daily: cumulative vancomycin area-under-the-curve to date, days of aminoglycoside or colistin exposure, the running slope of serum creatinine, and concurrent nephrotoxic insults such as intravenous contrast (the one risk factor Munshi et al. isolated, with an odds ratio of 2.4, 95% CI 1.14–5.03 P-1654). Each $\pi_k$ is a gating probability over $K$ latent subpopulations, each $h_{0,k}$ is a spline-smoothed non-parametric baseline hazard, and the $\beta_k$ are component-specific log hazard ratios. The second clock, $h_{\mathrm{fail}}$, is the cause-specific hazard of infection-related death or treatment failure, estimated on the same covariates plus severity markers. Death and live discharge are handled as competing events through cumulative-incidence calculations, not as ordinary censoring — a deliberate correction for the bias the file identifies as endemic to critical-care cohorts, where the sickest patients die before their kidneys can fail.

The resistance crisis enters in two ways, one as mechanism and one as motivation. As motivation: colistin remains the final line of defense — 95.7% of carbapenem-resistant A. baumannii isolates in a 3,246-isolate Chinese surveillance program were still colistin-susceptible Antimicrobial Activity of Colistin Against Contemporary (2015 – 2017) P. aeruginosa and A. baumannii Isolates From a Chinese Surveillance Program — precisely because it is used sparingly, and sparing use is exactly what a well-timed switching rule enforces. As mechanism: sub-inhibitory colistin exposure induces an 8.5-fold increase in mcr-9 expression and pushes the minimum inhibitory concentration from 4 mg/L to 128 mg/L in carbapenem-resistant Enterobacter cloacae Molecular antibiotic resistance mechanisms and co-transmission of the mcr-9 and metallo-β-lactamase genes in carbapenem-resistant Enterobacter cloacae complex, meaning every day of continued therapy after the infection is clinically controlled buys toxicity risk and selection pressure for inducible resistance. Because serial MIC measurements are almost never available in routine records, we cannot observe induction directly; instead we include an "exposure-beyond-stability" covariate — days of therapy continued after the first objective stability signal — as a measurable proxy, and we flag unmeasured induction as an explicit limitation to be probed in the sensitivity analyses below.

The switching rule itself is simple once the two clocks are fitted: continue therapy only while the one-day-ahead predicted probability of infection failure under de-escalation exceeds the one-day-ahead predicted probability of kidney injury attributable to continuing. The output is a single date per patient, not a risk score the clinician must interpret. A secondary deliverable connects this to drug development: Morgan et al. revived the abandoned streptothricin class — purified Streptothricin F overcame pandrug-resistant K. pneumoniae in mice with a maximum tolerated dose of 200 mg/kg versus 10 mg/kg for the toxic natural mixture — but explicitly cautioned that rodent maximum-tolerated-dose protocols cannot capture the delayed renal toxicity that killed the class in the 1940s, and called for longitudinal modeling before human trials Streptothricin F is a bactericidal antibiotic effective against highly drug-resistant gram-negative bacteria that interacts with the 30S subunit of the 70S ribosome. We will show how the fitted hazard shape from our intensive-care data, combined with the Cox-based continual reassessment design of Lei and Zhu A novel model of the continual reassessment method in Phase I trial, supplies an observation-window and stopping-rule template for the first-in-human Streptothricin F trial, closing the preclinical-to-clinical gap the file identifies as untouched.

Proposed Experiments

The primary cohort is built from MIMIC-III v1.4 using the MIMIC-Extract pipeline, which converts raw intensive-care records into hourly clinical time series of 104 aggregated variables MIMIC-Extract: A Data Extraction, Preprocessing, and Representation Pipeline for MIMIC-III. We select adult admissions receiving at least 48 hours of a last-resort or high-nephrotoxicity regimen (vancomycin with piperacillin-tazobactam, aminoglycosides, or colistin), define the event as time to acute kidney injury onset from serial creatinine, and treat death and dialysis-free discharge as competing events. No reported result found on Lacuna for time-to-kidney-injury prediction on MIMIC-III, so the first experimental contribution is establishing that reference ourselves; the nearest anchors in the file are Katzman et al.'s DeepSurv, which achieved a concordance index of 0.643 on METABRIC versus 0.630 for a linear Cox model and 0.624 for a random survival forest Deep Survival: A Deep Cox Proportional Hazards Network, and the calibration gap on FLCHAIN between Deep Cox Mixtures (ECE 0.0200) and a standard Cox model (0.0992) Deep Cox Mixtures for Survival Regression. We report time-dependent concordance and Expected Calibration Error for kidney-injury onset, plus Schoenfeld-residual diagnostics demonstrating that the proportional hazards assumption fails in this cohort — the crossing-hazard behavior already observed in the critically ill SUPPORT population used by both Katzman et al. and Nagpal et al.

The second experiment validates externally on AmsterdamUMCdb, the European intensive-care database used by the strongest existing stewardship model Estimating Treatment Effects for Time-to-Treatment Antibiotic Stewardship in Sepsis, asking whether the fitted hazard shapes and calibration transfer across health systems. The third experiment is the policy test. We replay the cohort under three policies: the observed clinical policy, the switching date recommended by our dual-clock rule, and the antibiotic-timing recommendations of the T4 counterfactual model. The endpoints are intravenous-therapy days saved, predicted kidney-injury events averted, and mortality neutrality. The reference point for the observed policy comes from the Birmingham hospital stewardship study, which reconstructed 24,511 therapy episodes and found that although 36% of intravenous regimens were eventually switched to oral therapy, clinicians delayed conversion by an average of 3.6 days (95% CI 3.4–3.7) after patients met objective stability criteria Feasibility study of hospital antimicrobial stewardship analytics using electronic health records; a model whose recommended switch dates cluster 3–4 days earlier than actual practice, without raising predicted mortality, is directly attacking the delay that study documented. Mortality neutrality is judged against the precedent of Cariou et al., who showed that carbapenem-sparing empiric regimens with aminoglycosides achieved 30-day mortality of 7% versus 15% for carbapenems (p=0.16) with comparable time to cure (3.3 versus 3.8 days) Efficacy of carbapenem vs non carbapenem β-lactam therapy as empiric antimicrobial therapy in patients with extended-spectrum β-lactamase-producing Enterobacterales urinary septic shock: a propensity-weighted multicenter cohort study — evidence that earlier de-escalation is not inherently dangerous, but also a reminder that their modest sample of 156 patients could not detect late-emerging safety signals, which is exactly what our design is built to see.

Baselines

The naive reference point is the observed clinical policy itself, characterized by the Birmingham stewardship cohort's mean length of therapy of 5.8 days and the 3.6-day average delay in intravenous-to-oral switching after stability criteria were met Feasibility study of hospital antimicrobial stewardship analytics using electronic health records, together with the 25.4% kidney-injury rate under vancomycin/piperacillin-tazobactam reported by Munshi et al. P-1654. This baseline matters because it defines what happens with no predictive model at all, and beating it means reducing intravenous days and kidney-injury incidence without increasing mortality in the replay. The strongest existing methodological competitor is T4, the LSTM encoder-decoder with propensity matching that estimates counterfactual antibiotic-timing windows in sepsis Estimating Treatment Effects for Time-to-Treatment Antibiotic Stewardship in Sepsis; it is the only prior model that individualizes antibiotic timing, and on MIMIC-III it reported mortality under its recommended treatments often below 10% against a 15% average. We will run T4 on our exact cohort, data extract, and tuning budget, and beat it only if the dual-clock rule matches its mortality outcomes while additionally predicting kidney-injury onset timing — a capability T4 lacks entirely, since it models organ-failure score trajectories under a binary "antibiotic or not" intervention rather than hazards under specific drug classes. Within the statistical comparison, the strongest baselines are a static multivariable logistic regression replication of Munshi et al.'s design, a standard linear Cox model with the same time-varying covariates (the most serious simple competitor), and DeepSurv as the non-linear single-hazard deep model Deep Survival: A Deep Cox Proportional Hazards Network, all given identical covariates, splits, and hyperparameter search budget. The key ablation removes the time-varying cumulative-exposure covariates from our mixture model while keeping everything else fixed, isolating whether the gain comes from tracking accumulated drug exposure or merely from the mixture architecture; a second ablation collapses the mixture to a single Cox component to isolate the non-proportionality handling.

Why not just…?

Why not just run logistic regression on day-30 kidney injury? Because the clinical question is a stopping question, and logistic regression cannot answer it. Munshi et al.'s own results show the problem: they found therapy duration significantly longer in the kidney-injury group (4.65 versus 4.04 days, p=0.01) and isolated intravenous contrast as a risk factor (OR 2.4), but their static model cannot convert that association into a day on which to switch, cannot distinguish kidney injury on day 2 from day 14, and must either drop or misclassify patients discharged before the window closes P-1654. Our first experiment includes this baseline head-to-head; if it matches our model's calibration and timing resolution, the proposal should be abandoned, and we say so.

Why not just use a standard Cox model with time-dependent covariates? This is the most serious objection, and it is a real threat: if hazards in this cohort were proportional, a time-varying Andersen-Gill-style Cox model would capture nearly everything at a fraction of the complexity. The counter-evidence is that cumulative-threshold toxicity is the canonical non-proportional case, and the calibration penalty is large — on FLCHAIN, where survival trajectories cross, the standard Cox model's Expected Calibration Error was 0.0992 against 0.0200 for the mixture model Deep Cox Mixtures for Survival Regression, and crossing hazards were likewise documented in the critically ill SUPPORT population. But rather than argue from analogy, our design settles it empirically: the time-varying standard Cox model is a first-class baseline, and the Schoenfeld diagnostics plus the single-component ablation will show directly whether the mixture machinery earns its cost on this specific cohort. If the simple Cox model calibrates as well, we will report that honestly and let the contribution rest on the competing-risks formulation and the switching rule.

Why not just extend the T4 model to predict kidney injury too? T4 is genuinely close in spirit — it individualizes antibiotic timing with counterfactual estimation — and extending it is a plausible alternative. What it lacks structurally is the survival formulation: it predicts discrete organ-failure score steps under a uniform binary antibiotic intervention, with no right-censoring machinery, no event-onset timing, no class-specific exposure accumulation, and no resistance-relevant structure Estimating Treatment Effects for Time-to-Treatment Antibiotic Stewardship in Sepsis. Bolting a kidney-injury output onto it would still leave a yes/no prediction on a fixed horizon, reproducing the logistic-regression failure. Still, we treat T4 as the policy benchmark rather than dismissing it: the head-to-head replay on the identical cohort is the experiment that decides whether hazard-based switching adds value over counterfactual timing alone.

Decisive Checks

The proposal stands or falls on three checks. First, the proportional hazards diagnostics must show non-proportionality in the kidney-injury hazard of this cohort; if they do not, the mixture architecture is unjustified and the simpler Cox baseline should win the calibration comparison — we will report whichever result obtains. Second, the model must predict kidney-injury onset timing out-of-sample with calibration materially better than the static logistic and time-varying Cox baselines, benchmarked against the ECE gap of 0.0200 versus 0.0992 that the mixture approach achieved on FLCHAIN Deep Cox Mixtures for Survival Regression. Third, and decisively for the clinical claim, the replayed switching rule must shorten intravenous therapy by on the order of the 3.6-day delay documented in real practice Feasibility study of hospital antimicrobial stewardship analytics using electronic health records while keeping predicted mortality at or below the level achieved by T4's recommendations Estimating Treatment Effects for Time-to-Treatment Antibiotic Stewardship in Sepsis. A sensitivity analysis withholding the exposure-beyond-stability covariate will bound how much of the policy gain depends on the resistance-proxy ingredient that induction data cannot directly validate, and the Streptothricin F trial-design demonstration will show whether the fitted hazard shapes generalize from stewardship to first-in-human safety monitoring Streptothricin F is a bactericidal antibiotic effective against highly drug-resistant gram-negative bacteria that interacts with the 30S subunit of the 70S ribosome.