Continuous renal replacement therapy (CRRT) is a machine that runs a critically ill patient's blood through plastic tubing and a hollow-fiber filter for days at a time. The blood clots on these artificial surfaces, and when the filter clots, treatment stops. Blood trapped in the circuit is lost, and nurses have to rebuild the setup. Bedside staff already watch the filter pressures rise as clotting sets in. What they lack is an answer to the question that decides the next action. Is this circuit failing for mechanical reasons that a change in pump settings or catheter would fix, or because the patient's blood is prothrombotic, in which case only a change in anticoagulation will help? We propose a clotting-hazard model that is built to answer that question. Its hour-by-hour risk splits exactly into a mechanical part and a biological part. We would judge the explanation by whether that split predicts which intervention actually prolongs circuit life. Whether it merely looks faithful is not the test.
Two sets of findings make this worth doing. On the biology side, standard coagulation tests are poor guides to circuit life. In combined liver and kidney failure, circuits clot at a median of 9 hours despite high INR, and heparin did not lengthen them (Circuit lifespan during continuous renal replacement therapy for combined liver and kidney failure). On the mechanical side, simply raising blood flow from 200 to 250 ml/min lengthened median circuit life from 13.0 to 33.5 hours in anticoagulation-free CRRT, and falling transmembrane pressure mediated up to 60% of that effect (Higher blood flow rates in anticoagulation-free CRRT improve circuit survival and clinical outcome). Both mechanisms are real. Neither applies to every patient. No existing tool tells a clinician which one is operating in the circuit in front of them.
Core Claim
The model predicts the hazard of circuit clotting, meaning the per-hour risk that the filter clots given it has survived so far. Its log-hazard is constrained to be a sum of separately learned pieces, one per group of causes. Let $h(t)$ be the clotting hazard at hour $t$ of a circuit's life, and $\alpha(t)$ a baseline that depends only on circuit age. The model is
$$\log h(t) = \alpha(t) + f_M(m_t) + f_B(b_t) + f_S(s_t),$$
where the three input groups are defined as follows. $m_t$ holds mechanical inputs that clinicians can change: blood flow rate, filtration fraction, dose, and access and return pressures, which reflect catheter function. $b_t$ holds biological inputs: platelet count, fibrinogen, INR, inflammatory markers, liver-failure status, and the current anticoagulation mode and dose. $s_t$ holds symptom readings of existing fouling, namely transmembrane pressure and pre-filter pressure along with their recent slopes. Each $f$ is itself a sum of one-dimensional shape functions over lagged features.
Because the model is additive, the change in risk between two windows splits exactly into $\Delta f_M + \Delta f_B + \Delta f_S$. No integration path or baseline is needed.
The symptom channel carries the key design choice. Rising transmembrane pressure mostly signals that the filter is already clotting, whatever the cause. If it were allowed to compete with the other groups, it would absorb most of the attribution and tell clinicians nothing they did not already see. So we keep $f_S$ as a separate nuisance term that sharpens timing. The interpretable quantity is the mechanical share of the non-symptom risk increase, $\rho_t = \Delta f_M / (|\Delta f_M| + |\Delta f_B|)$, computed over the last few hours.
The central hypothesis follows from this. Circuits with high $\rho_t$ should gain more from raising blood flow or lowering filtration fraction. Circuits with low $\rho_t$ should gain more from starting or intensifying anticoagulation.
Relation to Prior Work
The closest work falls into three groups, and none of them links explanations to circuit-level interventions.
The first group predicts outcomes of extracorporeal therapy from static data. Data-driven prediction of continuous renal replacement therapy survival predicts patient survival at CRRT start from pre-initiation records and explains it with Tree SHAP. It does not model the circuit or use telemetry. Machine learning to predict hemorrhage and thrombosis during extracorporeal membrane oxygenation classifies thrombosis once per stay in 44 ECMO patients with static feature rankings. We model time-to-clot per circuit, with inputs that change during treatment.
The second group explains risk as it changes over time. Delta-XAI explains why a monitoring model's prediction changed between windows. Its SWING method integrates along the patient's own history to avoid unrealistic all-zero references, and it has been applied to mortality and sepsis. It supplies our main black-box comparison and its evaluation style, but it is passive and blind to which inputs are levers. MIMIC-IF is a constraint on our design. It shows that attention weights and some post-hoc attributions fail remove-and-retrain checks on MIMIC-IV, which is one reason we build the decomposition into the model instead of extracting it afterwards.
The third group builds decision support that is safe or interpretable. HRRT learns when to start renal replacement, which modality to use, and how much fluid to remove, via offline reinforcement learning on MIMIC-IV and AmsterdamUMCdb. It treats the circuit as always working. A reluctant additive model framework supplies our fitting recipe: start linear and let spline terms enter only when a concordance criterion earns them. Safe and Interpretable Estimation of Optimal Treatment Regimes motivates the choice of transparent structure over unconstrained reinforcement learning, since deep RL failed often under missing data in its simulations. We do not propose a treatment policy. We propose a diagnostic split whose value is tested against observed intervention effects.
Research Program
Cohort and labels. We will build a circuit-level cohort from MIMIC-IV's charted CRRT records. We expect these to contain hourly settings, pressures, anticoagulation entries and filter-change events. The first deliverable is an audit of how reliably a circuit end can be labelled as clotting rather than elective change, access failure, transport or death. The outcome is a competing-risks label; non-clotting ends are treated as censoring or as competing events, and we will report both.
External data. AmsterdamUMCdb (AmsterdamUMCdb) serves as the external cohort. Circuit-level data of the kind collected at West China Hospital and Austin Health would be sought through collaboration, because those cohorts contain the anticoagulation-free and liver-failure regimes that stress the method most.
Fitting. The additive hazard is fit as a discrete-time survival model on hourly windows. Within each group we use the reluctant spline procedure, with a small number of hand-specified interactions allowed only inside a group, never across groups.
Validation of the split. We validate the decomposition in three ways. First, a double-dissociation test on natural within-circuit changes in blood flow and anticoagulation. Second, pathophysiological sanity checks against published findings. Third, a subgroup audit by sex and ethnicity. The audit is motivated by the ROTEM mortality model whose AUROC fell from 0.84 in men to 0.59 in women (Prognostic value of admission ROTEM in trauma).
Baselines
Naive reference. The do-nothing reference is the rule nurses already use: raise an alarm when transmembrane pressure or its slope crosses a threshold. The threshold is tuned on the training split to match each model's alarm rate. This reference matters because much of the achievable warning is probably in the pressure trend alone. Any learned model must beat it at matched alarm burden, measured by hours of warning before clotting at 80% sensitivity.
Classical interpretable model. A Cox model with time-varying covariates uses the same inputs, lag structure and training data. Its hazard ratios are already an interpretable explanation, so it is the most honest competitor for both prediction and interpretability.
Strongest black box. Two models fill this slot. The first is a gradient-boosted discrete-time hazard model, explained by differencing Tree SHAP across windows. The second is a recurrent or transformer hazard model explained with SWING through the Delta-XAI code. Grouped attributions for both are summed over the same mechanical, biological and symptom partition. That lets us compute $\rho_t$ for them too and run them through the identical intervention test.
Fair comparison and targets. All models receive the same splits, equal hyperparameter search budgets, and the same competing-risk labels. There is no reported result found on Lacuna for circuit-clotting prediction on MIMIC-IV or any CRRT telemetry benchmark. The nearest reported numbers are 64% accuracy for random-forest thrombosis prediction in ECMO (Abbasi et al.) and an AUROC of 0.848 for patient survival, not clotting, at CRRT start (Zamanzadeh et al.). Neither is directly comparable. Our target for prediction is therefore relative. The additive model should come within 0.02 time-dependent concordance of the best black box and beat the pressure alarm on warning time. It should also produce a significantly stronger interaction in the intervention test than the black-box-derived $\rho_t$.
Ablations. Two ablations isolate the key ingredient. In the first, the same additive model lets symptom features compete within the attribution instead of sitting in a nuisance channel. In the second, one unpartitioned additive model is fit and its terms are grouped only afterwards. If either ablation matches the full model on the intervention test, then the structural partition is not what matters.
Why not just…?
Why not just alarm on rising transmembrane pressure? This will probably predict imminent clotting nearly as well as anything, and we expect it to be hard to beat on short-horizon discrimination. The point is that it cannot say what to do. Rising pressure is the common end of both mechanical and biological failure, so a pressure alarm by construction assigns everything to the symptom channel. The proposal succeeds only if $\rho_t$ adds information the alarm lacks. The intervention test measures this directly: we check whether $\rho_t$ modifies the benefit of blood-flow or anticoagulation changes after adjusting for current pressure level and slope.
Why not just fit a time-varying Cox model? This is a real threat. Our model is a nonlinear, grouped generalisation of it. If a linear Cox model with grouped covariates produces an equally strong dissociation, we will adopt it and report that flexibility was unnecessary. The contribution we are claiming is the partition into symptom, mechanical and biological causes plus the interventional validation, not the spline machinery. We do expect nonlinearity to matter in one place. The liver-failure data show platelet counts relating non-monotonically to circuit life and bleeding (Circuit lifespan…), which a log-linear term would misattribute.
Why not just raise blood flow for everyone? Even at 250 ml/min, only 13.76% of anticoagulation-free circuits reached 72 hours (Higher blood flow rates…). The same authors note that access and hemodynamic limits prevent universal use of high flow. A uniform rule leaves most failures unexplained, and its benefit is precisely what we predict will vary with $\rho_t$.
Why not just a deep model explained with SWING? SWING avoids unrealistic reference inputs and was the most faithful method in the Delta-XAI evaluations. However, faithfulness to the model is not the same as correctly naming the lever. Grouped SWING attributions will likely load heavily on pressure features that are consequences rather than causes. We run it through the same intervention test instead of arguing the point.
Decisive Checks
Double dissociation. The make-or-break experiment uses hourly windows from circuits that are still running and emulates a target trial. We compare windows in which blood flow was raised or filtration fraction lowered in the next hour against matched windows where it was not. Separately, we compare windows in which anticoagulation was started or intensified against matched windows without that change. We then estimate how the reduction in clotting hazard varies with $\rho_t$.
The claim holds if two things are true together. The benefit of the mechanical change rises with $\rho_t$, and the benefit of the anticoagulation change falls with it, a crossover that confounding along a single axis is unlikely to produce. The main confounder is that clinicians raise flow when the catheter already works well. We handle this by adjusting for access pressure and prior flow. We also use a negative control: the model should assign no meaningful share to the circuit's hour-of-day.
Pathophysiological checks. Three findings from the literature should be reproduced. In liver-failure circuits the model should assign little protective weight to high INR. In anticoagulation-free cohorts, blood flow should be a major mechanical term. The effect of flow should run largely through the symptom channel, consistent with the reported mediation by transmembrane pressure.
Subgroup audit. We report discrimination and dissociation strength separately by sex and ethnicity. Failure in any subgroup is treated as a finding, not averaged away.
Success on the crossover test would give clinicians a checked answer to "which lever" at the moment pressures begin to climb. A null result would show that published circuit-life findings do not yet translate into bedside-actionable attributions.
References
- Circuit lifespan during continuous renal replacement therapy for combined liver and kidney failure
- Higher blood flow rates in anticoagulation-free CRRT improve circuit survival and clinical outcome
- Data-driven prediction of continuous renal replacement therapy survival
- Machine learning to predict hemorrhage and thrombosis during extracorporeal membrane oxygenation
- Delta-XAI: A Unified Framework for Explaining Prediction Changes in Online Time Series Monitoring
- MIMIC-IF: Interpretability and Fairness Evaluation of Deep Learning Models on MIMIC-IV Dataset
- HRRT: A Holistic Renal Replacement Therapy Decision-Making Support System Using Hierarchical Reinforcement Learning
- A reluctant additive model framework for interpretable nonlinear individualized treatment rules
- Safe and Interpretable Estimation of Optimal Treatment Regimes
- Sharing ICU Patient Data Responsibly Under the Society of Critical Care Medicine/European Society of Intensive Care Medicine Joint Data Science Collaboration: The Amsterdam University Medical Centers Database (AmsterdamUMCdb) Example*
- Prognostic value of admission ROTEM in trauma: enhancing 30-day all-cause mortality prediction using machine learning