Bedside risk models for intensive care patients learn from data that are mostly not measurements. A lab value drawn six hours ago gets carried forward. A blood pressure that spikes because the patient moved is read as physiology. A variable that was never ordered is filled with the population mean. Clinical models that are interpretable by design, which expose exactly how much each variable contributes to a risk score, still report that contribution as if every number were a fresh, trustworthy observation. A clinician reading "lactate contributes +0.8 log-odds" cannot tell whether that comes from a lactate drawn ten minutes ago, a decayed estimate from yesterday, or the fact that nobody ordered lactate.
We propose an additive recurrent risk model whose output splits exactly into three labeled parts. The first is risk carried by real observations, each discounted by how old it is. The second is risk carried by the measurement process itself: when and how often clinicians chose to measure. The third is a record of observations that a per-variable cleaning gate judged implausible and down-weighted. The central empirical bet is that this split does more than bookkeeping. Across hospitals, the measurement-process term should absorb most of the site-to-site shift, because ordering protocols differ between sites. The physiology risk curves should stay comparatively stable. That split is what a clinician needs to trust a transferred model.
Core Claim
Two strong prior models take opposite sides on missing data. GRU-D adds a missingness mask and time-since-last-observation to a recurrent network with learned decay, and it shows that missingness is predictive ("informative missingness"). Its hidden state mixes all variables, so none of this is auditable. I-RNN isolates each variable in its own recurrent "lane" and sums lane outputs into the risk. It deliberately removes observation frequency to avoid encoding clinician concern, and it still matches a black-box GRU on eICU. Neither model reports provenance, meaning where each input value came from. I-RNN also passes sensor spikes straight into a lane's contribution.
Our claim is that the tension between the two dissolves once the model keeps the measurement-process signal in its own named additive term. The model can then use informative missingness, as GRU-D does, while its physiology risk curves stay free of sampling bias, as I-RNN intends. Because every value contribution is anchored to a specific observation time and cleaning decision, the explanation becomes a provenance audit rather than only a feature ranking.
Model
Each variable $k$ has a value lane. This is a small gated recurrent unit that updates only at time steps where $k$ is actually observed. Between observations its state $s_k$ is frozen, so the lane's value always traces back to a specific last-updated observation. Before an observation enters the lane, a plausibility gate $a_{k,t} \in [0,1]$ scales the update. The gate is computed from that variable's own recent history: jump size relative to recent values and elapsed time. It works like a learned, soft version of the median filtering used for wearable and bed-sensor spikes (activity recognition, smart beds). The gate is kept within the lane so that additivity survives. Each variable also has a measurement lane that sees only the mask history and the time-since-observation $\delta_{k,t}$. The prediction is
$$\text{logit}(\hat y_t) = \beta_0 + \sum_{k} \gamma_k(\delta_{k,t})\, f_k(s_{k,t}) + \sum_{k} q_k(\delta_{k,t}, c_{k,t}),$$
The terms are defined as follows. $f_k$ maps the lane state to a log-odds contribution and is centered so that $f_k$ of the population-baseline state equals zero. $\gamma_k(\delta) = \exp(-\max(0, w_k \delta + b_k))$ is a GRU-D-style learned decay. $q_k$ is a small network over time-since-observation and $c_{k,t}$, the count of observations in a trailing window.
The centering matters. As an observation ages, its contribution shrinks toward zero rather than being replaced by an imputed value that looks physiological. So no risk is ever attributed to a number that was not measured. Every prediction therefore decomposes exactly into observed physiology, measurement process, and, from the logged gate values, the risk that rejected observations would have contributed had they been accepted.
Research Program
The build is modest: I-RNN's lane structure, GRU-D's decay, a per-lane gate, and a parallel set of measurement lanes. The substance is in three analyses.
The first asks whether the split is real. The measurement lanes could simply learn everything, leaving the physiology lanes decorative. We test this by ablating each block at test time and by training physiology-only and measurement-only variants.
The second is robustness to sensor artifacts. We inject synthetic spikes, dropouts, and slow drift into vital-sign channels with known ground truth. We then measure how often an artifact produces a large attribution spike, a false alarm in the explanation, compared with I-RNN.
The third is cross-hospital stability, the question the 2019 sepsis challenge raised by showing large drops on a held-out hospital. We train on one site and compare learned risk curves $f_k$ and measurement terms $q_k$ across sites. The prediction is that $q_k$ diverges much more than $f_k$. If it does, practitioners get a principled transfer recipe: keep the physiology lanes and recalibrate only the measurement lanes on the new site's data.
Baselines
The naive reference points are logistic regression and XGBoost on windowed summary statistics after median filtering and forward-fill. They show what fixed cleaning plus an off-the-shelf interpretable model achieves. On eICU hemodynamic-intervention prediction, I-RNN reports logistic regression at AUC 0.797 and XGBoost at 0.848. We rerun both on our splits with the same features available to the recurrent models, each given the same random-search budget.
The strongest interpretable competitor is I-RNN itself, reported at AUC 0.862 on eICU, matching the black-box GRU-Simple (the file gives no exact GRU-Simple decimal). It is the strongest because it already has additive lanes and decay. Our model must match its AUC within noise across five seeds while adding provenance. Beating it on explanation quality means fewer artifact-driven attribution spikes and more stable cross-site curves at equal AUC.
The strongest black-box accuracy reference is GRU-D, reported at AUC 0.8527 for MIMIC-III in-hospital mortality, against 0.8192 for GRU with mean imputation and 0.8294 for random forest (GRU-D). Since GRU-D uses informative missingness without the additive constraint, the gap to GRU-D is the price of auditability.
For PhysioNet 2019, the reported top challenge utility scores are 0.33 to 0.36 (sepsis challenge). Those scores are on the hidden Hospital C, which we cannot access, so we compare against GRU-D and XGBoost rerun on public cross-site splits. For PhysioNet 2012 mortality we found no reported result on Lacuna for either I-RNN or GRU-D, so all numbers there come from our reruns.
Two ablations isolate the key ingredients. The first is I-RNN with mask and time-since-observation appended as ordinary extra lanes, with no provenance-anchored value lanes and no centering. The second is our model with the plausibility gate removed. All models share splits, input variables, hyperparameter budget (the same number of trials), and GPU hours.
Why not just…?
Why not just run I-RNN and shade the explanation wherever the input was imputed? An overlay of the mask shows that a value was missing at time $t$. It does not show how much of the lane's current contribution comes from the last real observation versus the decay target, because I-RNN's decayed state is still scored as if it were physiology. We will measure this directly by comparing overlay-based "imputed share" estimates against our exact decomposition on the same patients. If the two agree closely, the architectural change is unnecessary, and we will say so.
Why not just add missingness indicators as extra additive features in I-RNN? This is the most serious threat, since it is most of our measurement lane. It is exactly our first ablation. It does not by itself stop the value lanes from assigning risk to carried-forward values, and it gives no rejected-observation record. If it matches our model on artifact false alarms and cross-site stability, the honest conclusion is that centering and gating add nothing. The experiments are set up so that this result would be visible.
Why not just use GRU-D with post-hoc attribution? TSInsight reports that gradient-based methods such as Integrated Gradients and GradCAM give noisy time-series attributions that fail suppression tests. The RETAIN authors argue that post-hoc readings of recurrent networks approximate rather than compute contributions, and they report no accuracy loss from exact decomposition (0.8705 vs 0.8706 AUC). We still include GRU-D with integrated gradients in the artifact test, to check whether its attributions track injected spikes.
Why not just use an LLM agent to explain the vitals? Vivaldi reports a 14-point drop in explanation relevance for reasoning models under multi-agent orchestration. Its approach also depends on iterative code execution, which does not fit continuous monitoring. It is not a competitor for the risk computation itself.
Proposed Experiments and Decisive Checks
We use MIMIC-III mortality, eICU 6-hour hemodynamic intervention, PhysioNet 2012 mortality, and PhysioNet 2019 sepsis with cross-site splits (train A, test B and the reverse). eICU also gets a multi-center hospital-held-out split. Metrics are AUC and AUPRC, plus challenge utility on sepsis, reported with confidence intervals over five seeds.
Three checks decide the proposal. The first is whether AUC stays within 0.01 of I-RNN on eICU. The second is the artifact test: whether our model reduces the rate at which injected spikes produce top-ranked attributions compared with I-RNN, and whether gate rejections line up with injected spike positions (precision and recall against ground truth). The third is cross-site: whether the change in learned curves between sites is significantly larger for the measurement terms $q_k$ than for the physiology terms $f_k$, and whether recalibrating only the measurement lanes on a small labeled slice of the target site recovers more AUC than recalibrating only the physiology lanes.
If the measurement-only variant matches the full model, or the cross-site shift falls mainly on the physiology curves, the core claim fails. We would report that as evidence that ICU risk models largely encode clinician behavior.
References
- Recurrent Neural Networks for Multivariate Time Series with Missing Values
- Interpretable Additive Recurrent Neural Networks For Multivariate Clinical Time Series
- Applying Multivariate Segmentation Methods to Human Activity Recognition From Wearable Sensors’ Data
- Identity and Posture Recognition in Smart Beds with Deep Multitask Learning
- Early Prediction of Sepsis from Clinical Data: the PhysioNet/Computing in Cardiology Challenge 2019
- TSInsight: A local-global attribution framework for interpretability in time-series data
- RETAIN: An Interpretable Predictive Model for Healthcare using Reverse Time Attention Mechanism
- A Multi-Agent Framework for Interpreting Multivariate Physiological Time Series