How Much Vasopressor Is Enough? Sizing an Exposure Budget for ICU Blood-Pressure Support

Critically ill patients whose blood pressure collapses are kept alive with vasopressors and inotropes, drugs that squeeze blood vessels or force the heart to contract harder. These drugs also cause arrhythmias, poor blood flow to organs, and extra strain on an already failing heart. Bedside practice treats them as an on/off response to a pressure alarm, with no explicit sense of how much total exposure a patient should absorb. We propose to borrow a question from energy and water engineering: how large should a storage buffer be, given that covering the last few rare peaks of demand costs far more than it returns? We apply it to cumulative vasoactive exposure. The "buffer" is a per-patient budget of drug exposure. The "demand" is the hourly risk of low blood pressure, read from invasive arterial and central venous lines. The question is where the trade-off curve between exposure spent and hypotension prevented bends.

The core claim is that the right budget depends on the dosing policy that spends it, and the right policy depends on the budget. Storage engineers call this a circular dependency and solve it by optimizing both together. Our hypothesis is that jointly choosing the budget and the dosing rule yields a hypotension-versus-exposure trade-off curve that dominates both observed clinician behavior and alarm-triggered rules. We further hypothesize that the curve has a knee, a point past which extra exposure buys little. If so, a per-patient exposure cap becomes a defensible clinical quantity rather than an implicit habit. We test this on hourly invasive-monitoring time series from MIMIC-III, a public ICU database.

Where the idea comes from and how it differs

The storage literature supplies the method. Zhiqi Wang et al. size household batteries by decoupling the control policy from physical capacity with an unbounded "virtual reservation." They then binary-search for the capacity at which marginal investment cost equals marginal revenue. Ye Wang 0005 et al. co-design a water tank's volume and its state-dependent pumping thresholds. They evaluate long-run cost from the stationary distribution of a discretized Markov chain instead of simulating decades. They tune both together with simultaneous perturbation stochastic approximation (SPSA), a gradient-free optimizer that perturbs all parameters at once. Against the utility's historical trigger-level control, this cut annual cost by 18%. The Distributed Real-Time Balancing Study uses Lyapunov drift-plus-penalty control. This method turns a long-run constraint into an hour-by-hour decision using a running "debt" counter and needs no forecasts. It reports a 1.7x cost improvement over greedy control. Diminishing-returns results motivate looking for a knee: Jinlong Tu et al. find that covering 60% of peak power captures over 95% of savings. We reuse these tools. What we change is the application: the stored quantity is a biological harm budget, and the demand process is autocorrelated, non-stationary patient physiology. Wang et al. explicitly flag that setting as outside their i.i.d., degradation-free assumptions.

The clinical side motivates the gap. Shirly Wang et al. built MIMIC-Extract, which turns MIMIC-III into hourly series of invasive arterial pressure, central venous pressure (CVP), heart rate, and vasopressor and ventilation status. It predicts vasopressor onset and offset with macro AUROC of 0.90 to 0.94. That task predicts what clinicians did, not what they should have done, and our work is normative rather than descriptive. Meiling Zhao et al. supply the clinical stakes. In 160 post-cardiac-surgery heart failure patients, conventional inotropic support carried a 22.5% complication rate versus 5% with levosimendan, and the trial used fixed dosing with no feedback. Exposure is therefore a real cost worth budgeting, and fixed protocols leave room for state-dependent control.

Invasive-hemodynamics work constrains us. Koshi Miyahara et al. and Noah Weingarten et al. show that true contractility can only be separated from loading conditions using simultaneous pressure and volume, for example with pressure-volume loops under vena cava occlusion. Routine ICU lines give pressure without volume. Our state estimate of "hemodynamic reserve" is therefore load-confounded. We use CVP as a crude filling-pressure proxy and report this as a limitation rather than claiming contractility-level control.

Research Program

We first build the cohort and the exposure measure. From MIMIC-Extract we take all ICU stays with any vasopressor hours and an arterial line, and we stratify the cardiac-surgery recovery unit as a prespecified subgroup. MIMIC-Extract's intervention records are, as far as the file indicates, on/off status. Our primary exposure is therefore vasopressor-hours, and the action is the hourly decision to give or withhold support. Recovering continuous doses from raw MIMIC-III infusion records is a secondary analysis, because dose granularity matters clinically.

Next we fit a dynamics model, which predicts next-hour mean arterial pressure (MAP) given the current state and action, on training patients. We also fit a discretized Markov chain over MAP, CVP, and recent-exposure bins. As in Ye Wang 0005 et al., the chain's stationary distribution gives a cheap long-run estimate of hypotension burden and exposure for any policy.

The default controller is drift-plus-penalty with a virtual budget queue:

$$Q_{t+1}=\max(Q_t+a_t-b,\,0),\qquad a_t=\arg\min_{a}\; V\,\mathbb{E}[h(\mathrm{MAP}_{t+1})\mid s_t,a]+Q_t\,a .$$

Here $a_t$ is the hour-$t$ vasopressor action and $b$ is the budgeted exposure per hour, which plays the role of storage capacity. $Q_t$ is accumulated over-budget exposure, $s_t$ is the patient state, and $h$ is a hypotension cost, for example the shortfall below 65 mmHg. $V$ weights pressure against budget.

Standard Lyapunov analysis gives average cost within $O(1/V)$ of the best budget-feasible policy, with budget overshoot growing as $O(V)$. That guarantee assumes the dynamics model is correct, which is exactly what we must stress-test.

Sizing then follows the Wang et al. logic. Sweeping $b$ traces a frontier of hypotension-hours against exposure-hours. The recommended budget $b^*$ is where the marginal hypotension-hours averted per extra exposure-hour falls below a clinician-stated exchange rate. We jointly tune $b$ and the controller's threshold parameters with SPSA, mirroring tank-plus-pump co-design. Finally we ask whether $b^*$ is patient-specific, meaning predictable from admission features. If it is, the deliverable is an individualized exposure cap.

The controller is deliberately swappable. If a penalized offline reinforcement-learning policy produces the same frontier, the contribution is the sizing and co-design question, not the Lyapunov machinery.

Baselines

The naive reference point is the observed clinician policy. This is the hypotension-hours and exposure-hours actually realized in held-out patients, evaluated identically through the dynamics model and directly from the data. Any proposed policy must at least match its hypotension burden at lower exposure.

The first engineered baseline is an alarm-trigger rule: start support when MAP falls below a threshold, stop after a fixed number of hours above it. Its two thresholds are tuned by the same SPSA budget we give our method. This is the clinical analogue of SA Water's trigger control, which lost 18% to co-design in Ye Wang 0005 et al..

The second is an imitation policy. We threshold the MIMIC-Extract onset/offset classifier (random forest or LSTM, as in Shirly Wang et al.) into actions. Its reported macro AUROC is 0.90 to 0.94 with AUPRC often below 0.50. No reported result was found on Lacuna for the policy value of any vasopressor dosing policy on MIMIC, so we will establish that number ourselves.

The strongest competitor is penalized offline reinforcement learning: fitted Q-iteration with reward equal to minus hypotension cost minus a per-hour exposure penalty. We run it on the same training split, dynamics model, and hyperparameter search budget, and sweep its penalty to produce its own frontier.

The key ablation is "sizing without co-design." We fix the trigger policy, sweep only $b$ by capping its total hours, and compare the resulting frontier and $b^*$ to the jointly optimized ones. This isolates the circular-dependency claim.

Beating a baseline means the proposed frontier lies below it at matched exposure, with bootstrap confidence intervals over patients excluding zero. The result must hold under at least two independently fit dynamics models and a doubly robust off-policy estimator.

Why not just…?

Why not just use a well-tuned alarm-trigger rule? It is interpretable and is what clinicians already do. In storage problems, trigger rules lose to state-dependent co-design: 18% in water (Ye Wang 0005 et al.) and 1.7x against greedy control in power (Distributed Real-Time Balancing Study). Whether that gap survives in physiology is an open empirical question. Our trigger baseline gets the same tuning budget. If it lands on the same frontier, we report that simple rules suffice and that sizing the cap is the only useful output.

Why not just train a penalized offline RL policy? This is the real threat. A dose penalty is a Lagrangian version of our budget, and with a good value function it could trace the same frontier. We therefore treat the controller as interchangeable and run RL inside the same co-design loop. The Lyapunov controller's remaining advantages are concrete. It enforces a running budget rather than an average penalty, it needs no learned long-horizon value function, and its single knob maps directly to the budget. If RL matches it at equal exposure, the paper's claim narrows to the frontier, the knee, and patient-specific caps, which stand regardless.

Why not just regress outcomes on observed vasopressor hours? Sicker patients receive more drug, so a naive dose-outcome curve slopes the wrong way. Prior work gives no way around this confounding by indication without a model of how pressure responds to treatment. Our dynamics model and doubly robust estimation are the minimum needed. We also report the naive regression to show how far it misleads.

Why not just deploy the existing vasopressor-onset predictor? It imitates historical practice, including the toxicity Meiling Zhao et al. documents, and its low AUPRC means unreliable timing on rare escalations. It is included as a baseline so the gap between imitation and budgeted control is measured, not asserted.

Decisive Checks

Three outcomes decide the proposal. First, the co-designed frontier must dominate both observed practice and the trigger rule at matched exposure, and must show a visible knee. If the frontier is roughly linear, there is no sizing story. Second, the co-design ablation must give a meaningfully different $b^*$ than sizing under a fixed trigger policy. Otherwise the circular dependency is irrelevant in this domain. Third, the ranking of policies must agree across both dynamics models and the doubly robust estimator. If the estimates disagree, the honest conclusion is that retrospective pressure-only ICU data cannot settle the question, given that pressure without volume is load-confounded (Koshi Miyahara et al.). That negative result would itself argue for collecting richer invasive measurements before any prospective trial.