Patients in shock die from two opposite errors: giving fluid to a heart that cannot use it (flooding lungs and tissue), and withholding fluid from a heart that can. Existing dynamic indices — heart-lung interaction measures like pulse pressure variation — tell us whether the heart is on the steep part of its filling curve, and machine-learning classifiers on arterial waveform data extend this, but all of them predict only the mechanical response to a bolus. None of them asks the question that actually determines outcome: is the microcirculation still starved, and would more flow fix it? Meanwhile, a separate physiology literature shows that the venous-to-arterial carbon dioxide difference — the gap in dissolved CO₂ between venous and arterial blood, which widens when venous blood sits stagnant in under-perfused tissue — is a sensitive marker of hidden tissue hypoperfusion, but it has only ever been used retrospectively, after the fluid has already been given. And a third, entirely separate literature on multi-agent navigation has built machinery for exactly the problem a bedside resuscitation controller faces: too many noisy signals, only a few of which matter right now. That machinery — attention mechanisms that dynamically rank and prioritize competing threats rather than treating all inputs equally — has never been applied to a physiological system.
This proposal brings the three together. The core claim is that fluid responsiveness prediction should be reframed as a dynamic prioritization problem over competing physiological threats, where the CO₂ gradient supplies a perfusion-debt signal that modulates which hemodynamic inputs the controller attends to. Concretely, we propose building a resuscitation policy whose input layer dynamically re-weights hemodynamic and metabolic signals — using rank-and-prune attention of the kind validated in large-scale multi-agent control — with the CO₂ washout gradient serving as the gating signal for whether fluid decisions should be driven by macro-hemodynamic pressure indices or by microcirculatory clearance needs. This is not merely adding a feature; it is a structural claim that the decision of which sensor to trust at this moment is itself learnable, and that the CO₂ gradient is the physiologically correct arbiter.
What Each Direction Contributes
Fluid responsiveness prediction is the application regime. The field has established three facts that shape any design: static filling pressures fail because they locate a patient on the horizontal axis of the Frank-Starling curve without revealing the slope (Diretrizes para tratamento da sepse grave/choque séptico: ressuscitação hemodinâmica); dynamic indices exploiting heart-lung interaction work well but break under spontaneous breathing, low tidal volumes, arrhythmia, and ARDS (Ability of respiratory pulse pressure variation to predict fluid responsiveness in ARDS: still an unanswered question?); and machine-learning classifiers on arterial pressure waveforms, such as the Random-forest MLFRA trained on 394 boluses in 58 swine (Improving the precision of shock resuscitation by predicting fluid responsiveness with machine learning and arterial blood pressure waveform data), improve on classical pulse pressure variation but remain purely mechanical predictors, blind to whether the tissue actually needed the flow.
The venous-to-arterial CO₂ gradient is the missing physiological objective. By the Fick principle applied to CO₂, the gradient is approximately proportional to metabolic CO₂ production divided by cardiac output, so it functions as a washout sensor: narrow gradient means brisk flow, widened gradient means stagnant venous effluent. Three findings motivate and constrain its use. It flags occult hypoperfusion even when global oxygen saturation looks normalized (Use of CO2-derived variables in critically ill patients; Veno-arterial carbon dioxide gradient at the early stage of septic shock). Its change after a bolus tracks the cardiac index response with useful accuracy — a decrease of at least 2.1 mmHg gave a 90% positive predictive value with AUC 0.831 in a 205-patient multicenter cohort (Changes in central venous-to-arterial PCO2 difference and central venous oxygen saturation as markers to define fluid responsiveness in critically ill patients: a pot-hoc analysis of a multi-center prospective study) — but its baseline value has no predictive power at all, which is the single most important constraint on our design. And it is confounded by ventilatory state: acute hyperventilation widened the gradient from 4.7 to 7.0 mmHg with no hemodynamic change whatsoever (Acute hyperventilation increases the central venous-to-arterial PCO2 difference in stable septic shock patients), and static gradients failed to predict adverse outcomes after cardiac surgery (Central venous-to-arterial CO2 difference is a poor tool to predict adverse outcomes after cardiac surgery: a retrospective study). Any proposal that feeds the raw gradient into a controller must therefore handle these confounders explicitly, which is precisely where prioritization machinery earns its place.
Dynamic agent prioritization is the methodological lens, imported from autonomous navigation. Three results matter. Soft attention dilutes critical threats when benign background entities dominate — the long-tail and magnification pathologies identified in large-scale multi-agent settings (Concentration Network for Reinforcement Learning of Large-Scale Multi-Agent Systems) — which maps directly onto the bedside problem where dozens of streaming variables bury the two or three that matter. Hard rank-and-prune attention (scoring all entities, keeping the top few, pruning the rest) fixed this and scaled zero-shot in that work, but requires a surrogate supervisory signal to backpropagate through the non-differentiable pruning step. Physics-informed priority — ranking peers by a pairwise safety value rather than raw distance, as in the HJ-reachability neighbor selection of MAD-PINN: A Decentralized Physics-Informed Machine Learning Framework for Safe and Optimal Multi-Agent Control — is the analogue of ranking physiological inputs by threat (a widened, worsening CO₂ gap) rather than by signal magnitude or proximity to an alarm threshold. And behavior-aware attention with dynamic geometric graphs showed that priority weighting driven by interaction structure, not just geometry, yields large gains in dense scenes (BAT: Behavior-Aware Human-Like Trajectory Prediction for Autonomous Driving).
Core Claim
The existing two-way intersections stop short in a consistent way. The CO₂-marker studies are post-hoc evaluators: they tell you whether a bolus did work after it was infused, with modest negative predictive values (58–64%) and a gray zone covering 22% of patients (Changes in central venous-to-arterial PCO2 difference and central venous oxygen saturation as markers to define fluid responsiveness in critically ill patients: a pot-hoc analysis of a multi-center prospective study). The automated-controller line — the rule-based ReFit closed-loop algorithm achieving 100% survival in a swine hemorrhagic model (Autonomous precision resuscitation during ground and air transport of an animal hemorrhagic shock model) and the MIMIC-trained reinforcement-learning policies for septic shock fluid and vasopressor dosing (Raising High-Risk Awareness in Hemodynamic Treatment with Reinforcement Learning for Septic Shock Patients; Learning to Treat Hypotensive Episodes in Sepsis Patients Using a Counterfactual Reasoning Framework) — drive decisions on macro-hemodynamic surrogates and hourly or four-hourly EHR aggregates, blind to microcirculatory washout and to beat-to-beat dynamics alike. No work in the file treats the CO₂ gradient as an active input to a dynamic prioritization policy, and no work imports rank-and-prune or threat-ranked attention into physiological control.
Our claim is that the correct architecture is a two-tier controller. A low-frequency tier tracks the CO₂ gradient (and its ventilatory confounders) to estimate a continuously updated perfusion debt, which sets the operating mode. A high-frequency tier runs a prioritized input layer over streaming hemodynamic signals — arterial waveform features, heart-lung interaction indices, vasopressor state, ventilator settings — where the number of retained inputs and their weights are modulated by the current perfusion-debt estimate, using the rank-and-prune mechanism with a surrogate supervisory head trained to predict the eventual fluid-response outcome. The intuition is physiological: when perfusion debt is low, the controller should trust mechanical predictors of fluid responsiveness and be conservative; when debt is high and narrowing after interventions, the policy is on a productive trajectory; when debt is high and static, mechanical responsiveness is beside the point — the patient needs flow by another route (vasopressor, inotrope, or source control), and the controller should re-prioritize accordingly. The CO₂ gradient is not another feature among fifty; it is the signal that decides which features matter.
Research Program
The program has three components. First, a retrospective study on MIMIC-IV septic shock patients (the 7,792-patient cohort used by Raising High-Risk Awareness in Hemodynamic Treatment with Reinforcement Learning for Septic Shock Patients) asks whether paired venous-arterial blood gases, sampled at their native (intermittent) frequency, carry decision-relevant information about subsequent fluid-response outcomes that macro-hemodynamic features do not — specifically whether the trend of the gradient across sequential draws, rather than any single value, predicts response, and whether this survives adjustment for ventilator changes of the kind shown to confound the gradient (Acute hyperventilation increases the central venous-to-arterial PCO2 difference in stable septic shock patients). This is the feasibility gate: if gradient trends add nothing beyond blood pressure and lactate, the rest of the program does not deserve to run.
Second, we build the prioritized controller itself. The architecture follows ConcNet's rank-and-prune structure (Concentration Network for Reinforcement Learning of Large-Scale Multi-Agent Systems): a scoring network ranks all candidate physiological input streams, the top $d_c$ are retained and passed to a fixed-dimensional policy network, and the rest are pruned. To make the pruning trainable, we attach a surrogate supervisory head predicting the observed outcome of each historical fluid challenge (response vs. non-response, plus safety events), following ConcNet's motivational-subnetwork pattern. The scoring function is physics-informed in the spirit of MAD-PINN's reachability-ranked neighbor selection (MAD-PINN: A Decentralized Physics-Informed Machine Learning Framework for Safe and Optimal Multi-Agent Control): inputs are ranked by an estimated threat value — roughly, the expected harm of acting on this input wrongly given the current perfusion-debt estimate — rather than by signal magnitude. Training data comes from MIMIC-III hourly measurements via the extraction pipeline of MIMIC-Extract: A Data Extraction, Preprocessing, and Representation Pipeline for MIMIC-III and from MIMIC-IV at higher temporal resolution, with boluses labeled by post-bolus hemodynamic change where waveform-level cardiac output is unavailable.
Third, a closed-loop evaluation in a swine hemorrhagic-plus-septic model, designed to be directly comparable to ReFit's published protocol (Autonomous precision resuscitation during ground and air transport of an animal hemorrhagic shock model): same injury model, same transport conditions, same 3-hour survival endpoint, with continuous arterial and central venous access enabling high-frequency paired CO₂ sampling. The decisive comparison is our prioritized CO₂-gated controller versus ReFit under conditions designed to break ReFit's assumptions — low tidal volume ventilation and spontaneous breathing effort — where pulse pressure variation, ReFit's fluid trigger, is known to be unreliable.
Baselines
The do-nothing reference point is standard clinical management as reflected in the MIMIC cohorts: clinician-administered fluid and vasopressor titration, which in the counterfactual study of Learning to Treat Hypotensive Episodes in Sepsis Patients Using a Counterfactual Reasoning Framework meant a first fluid bolus of 283 mL at a mean of 4 hours after hypotension onset. Any learned policy must at minimum act earlier and no worse on survival-linked outcomes than this reference to justify its existence.
The strongest existing automated competitor is ReFit, the rule-based closed-loop resuscitation controller of Autonomous precision resuscitation during ground and air transport of an animal hemorrhagic shock model, which achieved 100% survival over 3 hours in 12 swine with stability within 42 minutes, using pulse pressure variation for fluid decisions and dynamic arterial elastance for vasopressor titration on fixed 15-minute cycles. We will run it on the identical animal model and endpoint, with the same sensor suite, and — critically — under both standard ventilation and low-tidal-volume/spontaneous-effort conditions, since ReFit's fluid trigger requires the heart-lung interaction that low tidal volume invalidates. Beating ReFit means matching its survival under its favored conditions and exceeding it under degraded-ventilation conditions, with lower cumulative fluid administered (a proxy for fluid toxicity).
On the retrospective side, the strongest machine-learning competitor is the MLFRA Random Forest of Improving the precision of shock resuscitation by predicting fluid responsiveness with machine learning and arterial blood pressure waveform data, which reported an AUROC of 0.82 for predicting fluid responsiveness from arterial waveform features (against 0.73 for classical pulse pressure variation). We will reimplement it on our MIMIC-derived bolus labels, give it the same feature budget, tuning budget, and compute as our model, and count ourselves as beating it only if the AUROC improvement exceeds the confidence interval of the original report and — more importantly — if the improvement concentrates in the patient subgroups where MLFRA is weakest (low tidal volume, arrhythmia, gray-zone pulse pressure variation).
The key ablation isolates the prioritization ingredient itself: the identical policy network trained with all inputs always active and uniformly weighted (no pruning, no threat ranking) versus with random or magnitude-based input selection versus with the full rank-and-prune, threat-scored mechanism. A fourth ablation replaces the CO₂-trend gating with lactate gating. If the full mechanism does not beat uniform weighting by a margin that survives multiple-comparison correction, the prioritization story is wrong even if raw performance is good.
Why not just…?
Why not just add the CO₂ gradient as one more feature in an existing classifier? This is the most serious alternative, and we take it as the primary ablation rather than arguing around it. The evidence for it is real: gradient changes after a bolus carry real signal (AUC 0.831 in the multicenter cohort of Changes in central venous-to-arterial PCO2 difference and central venous oxygen saturation as markers to define fluid responsiveness in critically ill patients: a pot-hoc analysis of a multi-center prospective study). But two facts argue that a flat feature union underuses it. Baseline gradient values carry no predictive information in that same study, so a naive classifier trained on mixed baseline and trend features will dilute the trend signal; and the gradient is severely confounded by ventilation (Acute hyperventilation increases the central venous-to-arterial PCO2 difference in stable septic shock patients) and failed outright as a static predictor after cardiac surgery (Central venous-to-arterial CO2 difference is a poor tool to predict adverse outcomes after cardiac surgery: a retrospective study). If the flat-feature model matches our prioritized model in the MIMIC experiments, we will report that honestly and the architectural claim dies — but the degraded-ventilation animal arm remains, because there the question is not classification accuracy but whether the controller correctly stops trusting an invalidated input, which no flat model addresses.
Why not just improve the pulse pressure variation thresholds, or use the passive leg raising test? Refining heart-lung indices attacks the right-heart/preload-slope question with ever more precision, and the passive leg raising literature shows change-in-PPV during leg raising remains robust even at low tidal volumes (Changes in pulse pressure variation induced by passive leg raising test to predict preload responsiveness in mechanically ventilated patients with low tidal volume in ICU: a systematic review and meta-analysis). But all of these remain purely mechanical predictors: they say whether the heart can increase output, never whether the tissue needs it. The right-heart literature makes the safety side explicit — fluid loading past right ventricular limits risks acute cor pulmonale and venous congestion (Fluid therapy in critically ill patients: perspectives from the right heart) — and a mechanically responsive patient with an already-adequate microcirculation is exactly the patient these tests will wrongly flag for more fluid. Our animal protocol includes responsive-but-non-hypoperfused animals specifically to test whether the CO₂-gated policy avoids the fluid load that a pure-PPV controller delivers.
Why not just use a bigger off-the-shelf RL policy on more EHR features? The MIMIC RL line already exists (Raising High-Risk Awareness in Hemodynamic Treatment with Reinforcement Learning for Septic Shock Patients; Learning to Treat Hypotensive Episodes in Sepsis Patients Using a Counterfactual Reasoning Framework), and scaling it is the obvious response. The limitation is structural, not a matter of capacity: those models consume 4-hour or hourly aggregations that destroy the beat-to-beat cardiopulmonary dynamics every dynamic index depends on, and they treat all features as a flat vector with no mechanism for deciding which sensor to trust when they disagree. The multi-agent literature diagnosed exactly this failure mode for soft attention over many benign entities (Concentration Network for Reinforcement Learning of Large-Scale Multi-Agent Systems) — the long-tail effect — and the fix there was architectural, not more parameters. Our retrospective experiments will include a scaled flat-feature RL policy as a comparator; if it matches the prioritized architecture on MIMIC, that will be evidence the prioritization matters only in the real-time, high-frequency regime the EHR data cannot test, which is itself a finding worth reporting.
Proposed Experiments and Decisive Checks
The retrospective study uses the MIMIC-IV septic shock cohort (7,792 patients) to test whether sequential CO₂ gradient trends improve prediction of fluid-response outcomes over macro-hemodynamic features alone; the primary metric is AUROC for post-bolus response, reported with the strong comparator being MLFRA's published 0.82 (Improving the precision of shock resuscitation by predicting fluid responsiveness with machine learning and arterial blood pressure waveform data). For the single-center septic shock cohort of 49 patients, the file reports a Δ-ΔPCO₂ AUC of 0.76 overall and 0.83 in the subgroup without tissue hypoxia (Central venous-to-arterial PCO2 difference as a marker to identify fluid responsiveness in septic shock); we use these as calibration targets — our trend extractor should recover signal of comparable strength before any controller work begins. No reported result found on Lacuna for CO₂ gradient trends in MIMIC cohorts, so the retrospective study is genuinely uncharted and the feasibility gate is honest.
The architecture experiments run on MIMIC-III via the MIMIC-Extract pipeline (MIMIC-Extract: A Data Extraction, Preprocessing, and Representation Pipeline for MIMIC-III) and MIMIC-IV at maximum available temporal resolution, with the ablation battery described above: uniform weighting, random selection, magnitude selection, lactate gating, and full threat-ranked rank-and-prune, all with matched compute and tuning budget. The decisive check is an interaction effect: the prioritized model's advantage over uniform weighting should be concentrated precisely in the patients whose ventilation state or rhythm invalidates the standard indices — the subpopulation where deciding which input to trust is the binding constraint.
The closed-loop animal study compares our controller against ReFit on the hemorrhagic shock model in which ReFit achieved 100% survival across 12 swine (Autonomous precision resuscitation during ground and air transport of an animal hemorrhagic shock model), with the primary endpoints being survival under degraded ventilation (low tidal volume, spontaneous effort), time to hemodynamic stability, and cumulative fluid administered as a fluid-toxicity proxy. The decisive, and falsifiable, prediction is this: under standard ventilation the two controllers will perform comparably, because pulse pressure variation works and prioritization has nothing to fix; under degraded ventilation the prioritized controller will maintain performance while ReFit degrades, because it can re-rank away from the invalidated input. If that interaction does not appear, the proposal's central claim is wrong, and the honest conclusion will be that perfusion-debt gating adds measurement burden without decision value.
References
- Diretrizes para tratamento da sepse grave/choque séptico: ressuscitação hemodinâmica
- Ability of respiratory pulse pressure variation to predict fluid responsiveness in ARDS: still an unanswered question?
- Improving the precision of shock resuscitation by predicting fluid responsiveness with machine learning and arterial blood pressure waveform data
- Use of CO2-derived variables in critically ill patients
- Veno-arterial carbon dioxide gradient at the early stage of septic shock
- Changes in central venous-to-arterial PCO2 difference and central venous oxygen saturation as markers to define fluid responsiveness in critically ill patients: a pot-hoc analysis of a multi-center prospective study
- Acute hyperventilation increases the central venous-to-arterial PCO2 difference in stable septic shock patients
- Central venous-to-arterial CO2 difference is a poor tool to predict adverse outcomes after cardiac surgery: a retrospective study
- Concentration Network for Reinforcement Learning of Large-Scale Multi-Agent Systems
- MAD-PINN: A Decentralized Physics-Informed Machine Learning Framework for Safe and Optimal Multi-Agent Control
- BAT: Behavior-Aware Human-Like Trajectory Prediction for Autonomous Driving
- Autonomous precision resuscitation during ground and air transport of an animal hemorrhagic shock model
- Raising High-Risk Awareness in Hemodynamic Treatment with Reinforcement Learning for Septic Shock Patients
- Learning to Treat Hypotensive Episodes in Sepsis Patients Using a Counterfactual Reasoning Framework
- MIMIC-Extract: a data extraction, preprocessing, and representation pipeline for MIMIC-III
- Changes in pulse pressure variation induced by passive leg raising test to predict preload responsiveness in mechanically ventilated patients with low tidal volume in ICU: a systematic review and meta-analysis
- Fluid therapy in critically ill patients: perspectives from the right heart
- Central venous-to-arterial PCO2 difference as a marker to identify fluid responsiveness in septic shock