Intensive care patients are tracked through many biomarkers at once, such as blood pressure, lactate and creatinine, and they also receive several treatments at once, such as a mechanical ventilator and vasopressors (drugs that raise blood pressure). Clinical risk models describe the patient with summary statistics of the biomarkers and treat each treatment as one more on/off feature. This proposal tracks a different quantity over time: how tightly the patient's state and the combination of treatments are coupled. The coupling has two parts. The first is how informative the current physiology is about which combination of therapies clinicians choose. The second is how informative that combination is about where the physiology goes next. Our hypothesis is that coupling falls, through therapy saturating or physiology no longer responding, before death and before conventional biomarkers saturate. If so, coupling is a dynamic marker of refractoriness that snapshot and rolling-window features do not capture directly.
The measurement comes from cooperative multi-agent reinforcement learning. There, PMIC defined the mutual information between a global state and the joint action vector of all agents, $I(s;\mathbf{u})$. Mutual information is the reduction in uncertainty about one variable from knowing another. PMIC showed that high coupling is not automatically good, because agents can lock into synchronized but poor behaviour, which the paper calls "coordination traps." We do not train a treatment policy. We take the same state–joint-action quantity and turn it into a per-patient, per-hour signal that a hospital can compute from its own unlabeled data, and we then test it as a longitudinal biomarker.
Core Claim
We define two pointwise coupling signals on hourly ICU grids. Let $s_t$ be the patient's biomarker state at hour $t$ and $h_t$ the recent history. Let $\mathbf{u}_t=(a^{\text{vent}}_t,a^{\text{vaso}}_t)$ be the joint intervention vector, which is four discrete combinations in the base version. Policy coupling is
$$c^{\pi}_t=\log p(\mathbf{u}_t\mid s_t,h_t)-\log p(\mathbf{u}_t\mid h_t),$$
the pointwise version of $I(s;\mathbf{u})$ conditioned on history. Response coupling is
$$c^{r}_t=\log p(s_{t+1}\mid s_t,h_t,\mathbf{u}_t)-\log p(s_{t+1}\mid s_t,h_t).$$
This is the pointwise version of the conditional mutual information $I(s_{t+1};\mathbf{u}_t\mid s_t,h_t)$. Response coupling measures how much knowing the treatment combination sharpens the forecast of the next physiological state beyond what the current state already implies. Each density is fit on the population of a single hospital. No outcome labels are used, because these are models of clinician behaviour and of physiological dynamics. Rolling sums of $c^\pi$ and $c^r$, together with their slopes, form the biomarker trajectory.
The claim is falsifiable in three ways. First, coupling trajectories should add prognostic value beyond a strong rolling-window gradient-boosted model that already sees the treatment flags. Second, they should degrade less across hospitals, because each site's coupling is computed relative to its own fitted policy and dynamics models. Third, they should keep their signal in the septic subgroup, where static ratios such as the neutrophil-to-lymphocyte ratio stop discriminating.
Relation to Prior Work
Three papers supply the method.
PMIC supplies the state–joint-action quantity and the warning that coupling is not monotonically good. PMIC maximizes that quantity on good trajectories and minimizes it on bad ones with neural bounds (MINE and CLUB) inside a learning loop. We only estimate it, and we never optimize clinicians' behaviour. Because the action space is small and discrete, we replace those neural estimators with an ordinary calibrated classifier and a probabilistic forecaster. This sidesteps the estimator overhead and hyperparameter sensitivity that PMIC reports.
The response-coupling term reuses the logic of Identifying Driver Interactions via Conditional Behavior Prediction. That work scores how interactive two road users are by the divergence between conditional and marginal trajectory forecasts. It also reports that conditioning improves target-agent prediction by roughly 10% on the Waymo Open Motion Dataset. We carry the conditional-versus-marginal construction over from vehicle paths to physiological trajectories.
Conditioning on the current state follows Information-Theoretic Approach to Detect Collusion in Multi-Agent Games. There, conditioning on the shared environment removes spurious dependence between agents. Here, conditioning on $s_t$ removes the trivial dependence created when sicker patients both get more treatment and have worse next states. This de-confounding is partial. Severity that is not recorded remains, and we treat coupling as a predictive marker, not a causal effect estimate.
Two clinical lines motivate the gap. MIMIC-Extract builds hourly grids with ventilation and vasopressor flags, but models each intervention as an independent prediction target. XMI-ICU uses stacked 6–24 h look-back windows with XGBoost and time-resolved attributions, and it ranks ventilation as one feature among many. Neither asks whether the joint therapy and the state are still informative about each other.
The clinical cohort studies constrain the design. The NLR study in MIMIC-II found no significant mortality association in sepsis, and the trajectory studies on albumin and hs-CRP show that direction and velocity carry risk beyond snapshots. Together they define the regime where a new dynamic marker could matter: severe dysregulation, where individual markers sit at floor or ceiling.
Research Program
We build the coupling estimators on MIMIC-Extract grids of MIMIC-III. The pipeline has four parts.
The policy model is a gradient-boosted classifier over the four joint action states, given the current state plus a history encoding. It is paired with a history-only classifier. Both are calibrated with temperature scaling on held-out patients, because pointwise mutual information is only as good as the calibration of these models.
The dynamics model is a heteroscedastic Gaussian forecaster of the next-hour state for a fixed panel of mean arterial pressure, heart rate, lactate, creatinine and SpO2. Lab values are carried forward with a time-since-measured indicator. It is paired with an identical forecaster that omits $\mathbf{u}_t$.
The density models are trained on each hospital's trajectories without mortality labels. An extension adds a coarse vasopressor-dose dimension where dose is available.
Coupling trajectories are summarized over 6, 12, 18 and 24 h windows, matching the windows in XMI-ICU. These summaries are then appended to, or used instead of, standard features in a downstream outcome model.
We will also do a descriptive analysis to ground the "coordination trap" interpretation. We will check whether low policy coupling before death mainly reflects saturation, meaning both therapies on at maximum with low action entropy, or decoupled escalation, where actions change while the state does not. We will also check how often each pattern occurs in survivors.
The main threat to validity is withdrawal of life-sustaining treatment. When care moves to comfort measures, treatments stop independently of physiology, so coupling collapses right before death for a reason unrelated to biology. We will exclude or censor stays at the first documented comfort-care or do-not-resuscitate transition, and we will report results with and without that censoring.
Baselines
The naive reference point is APACHE IV and the static admission snapshot. XMI-ICU reports APACHE IV at AUROC 69.9 at the 24-hour mark on its eICU myocardial-infarction cohort. This anchors what "no longitudinal modelling" achieves.
The strongest existing method is the XMI-ICU pipeline: XGBoost on stacked rolling means and standard deviations over 6/12/18/24 h windows. It reports AUC 92.0 and balanced accuracy 82.3% for 6-hour mortality on internal eICU validation, ahead of TabNet (84.1) and NODE (85.4), and AUC 80.0 on zero-shot transfer to MIMIC-IV. It is the strongest competitor because it already sees the treatment flags and is free to learn their interactions with physiology. We will rerun it on our cohorts with the same windows, the same patient splits and the same Bayesian hyperparameter search budget as our augmented model. We will add the raw hourly ventilation and vasopressor flags and their window means to its features, so that any gain from coupling cannot come from information the baseline was denied. For general ICU mortality on MIMIC-III, the reference is the MIMIC-Extract baseline suite, including the GRU-D recurrent model, which reports AUROCs between 0.85 and 0.90.
Beating the strongest baseline means the following. Appending coupling summaries to the XGBoost features must raise AUROC and AUPRC by a margin whose patient-level bootstrap 95% interval excludes zero. It must also shrink the internal-to-external AUROC gap relative to the 92.0-to-80.0 drop XMI-ICU reports.
Two ablations isolate the key ingredient. The first is marginal-per-intervention coupling: the same estimators fit separately for ventilation and for vasopressors, with the two scores summed. It tests whether modelling the joint action matters. The second is unconditioned coupling, which drops $s_t$ from the response term. It tests whether conditioning on the current state, the collusion-detection idea, is what removes the trivial sicker-patients-get-more-treatment signal.
Why not just…?
Why not just let XGBoost learn treatment–physiology interactions from the raw flags? This is the most serious threat. Trees can in principle learn "on both therapies and blood pressure still low" from raw features. The direct test is the baseline above, which receives every flag. If coupling adds nothing internally, we will report that honestly. The remaining argument would then rest on transfer, because coupling is computed against each hospital's own label-free policy and dynamics models, while tree splits learned on one site's thresholds are what fail at another. The cross-site experiment settles this. If coupling helps neither internally nor externally, the idea fails.
Why not just use treatment intensity, such as the count of concurrent therapies or the vasopressor dose? This is the clinical analogue of the proximity heuristic. Co-occurrence is cheap, but Identifying Driver Interactions via Conditional Behavior Prediction found that distance-based selection failed to tell co-located, non-interacting agents from distant, interacting ones, where information-based scoring succeeded. The clinical analogue is that two patients on identical maximal therapy can differ completely in whether their physiology still responds. We include intensity features in a separate baseline, and we report coupling within strata of equal intensity. A signal that survives that stratification cannot be intensity in disguise.
Why not just fit a per-patient dose–response slope, such as the change in blood pressure per vasopressor change? This is simpler and interpretable, and for vasopressors alone it may capture much of the response-coupling signal. We implement it as the marginal-per-intervention ablation with a linear dynamics model. If it matches the joint estimator, the honest conclusion is that a single-drug responsiveness slope is the useful marker, and the joint-action machinery is unnecessary.
Why not use latent trajectory classes or splines on the biomarkers themselves? LCTM on hs-CRP and spline Cox models on albumin capture shape and velocity well in stable cohorts. However, they describe passive markers and contain no treatment term. The MIMIC-II NLR result shows that such markers can lose discrimination entirely in sepsis. The sepsis-subgroup experiment below compares coupling directly with the rolling biomarker features in that regime.
Proposed Experiments
The first experiment establishes internal value on MIMIC-III via MIMIC-Extract. The task is in-ICU mortality predicted at 6 and 24 h from windowed features, with comfort-care censoring. We compare the XGBoost-plus-flags baseline against the same model with coupling summaries appended, and against coupling alone. We report AUROC, AUPRC (the MIMIC-Extract paper notes AUPRC often falls below 0.50 under class imbalance on its intervention tasks), calibration, and the hour-to-hour risk flip-flopping that XMI-ICU flags as a failure mode.
The second experiment is cross-site transfer. We train the outcome model on eICU, refit only the label-free coupling estimators on MIMIC-IV, and evaluate there. We first replicate the XMI-ICU myocardial-infarction cohort so the 92.0 to 80.0 comparison is like-for-like, then repeat on a general ICU cohort. Whether eICU records vasopressor administration hourly and reliably is not stated in the file. If it does not, the cross-site study falls back to MIMIC-III to MIMIC-IV with ventilation plus vasopressors.
The PhysioNet 2019 sepsis data (Early Prediction of Sepsis) contains only demographics, vitals and labs. We do not use it, because the coupling signals cannot be computed without intervention records.
The third experiment is the sepsis stratum. We test whether coupling trajectories discriminate survivors from non-survivors among septic patients, where the NLR study found no significant association. We compare this against the rolling biomarker features alone.
Two sanity checks guard the measurement itself. In the first, we permute joint actions between patients matched on state bin and hour, which should drive mean coupling to near zero. If it does not, the estimators are leaking history and the measured coupling is an artifact. In the second, we vary calibration quality of the density models deliberately, to show how sensitive the downstream gain is to estimator error.
The decisive outcome is a significant gain over the strongest baseline in the first experiment or a clearly smaller transfer gap in the second, with the joint-versus-marginal ablation favouring the joint estimator. Without that, the claim that joint coupling is a distinct biomarker does not hold.
References
- PMIC: Improving Multi-Agent Reinforcement Learning with Progressive Mutual Information Collaboration
- Identifying Driver Interactions via Conditional Behavior Prediction
- Large Scale Interactive Motion Forecasting for Autonomous Driving : The Waymo Open Motion Dataset
- Information-Theoretic Approach to Detect Collusion in Multi-Agent Games
- MIMIC-Extract: a data extraction, preprocessing, and representation pipeline for MIMIC-III
- XMI-ICU: Explainable Machine Learning Model for Pseudo-Dynamic Prediction of Mortality and Heart Attack in the ICU
- The association between the neutrophil-to-lymphocyte ratio and mortality in critical illness: an observational cohort study
- Association between albumin changes and prognosis in older patients with acute myocardial infarction
- Association Between High-Sensitivity C-Reactive Protein Trajectories and the Incidence of Metabolic Syndrome:A Retrospective Cohort Study
- Early Prediction of Sepsis from Clinical Data: the PhysioNet/Computing in Cardiology Challenge 2019