Two Kinds of Patient Neighbors: Separating Chronic and Acute Similarity for Rolling ICU Risk

Hospital risk scores such as APACHE-II and SAPS-II are computed once, from the first day of a stay, and then trusted for days. They lose accuracy quickly as the patient's course unfolds. A natural fix is to ask "what happened to patients like this one?" and to keep asking as the patient changes. Existing attempts to do this define "like this one" in a single way, usually by shared admission diagnoses, so every link between two patients counts the same. This proposal argues that two different relations should be kept apart. One is chronic similarity: two patients carry the same long-standing conditions. The other is acute similarity: two patients are in the same physiological state right now. The second relation changes every few hours. The first never does.

The core claim is narrow and testable. Every six hours, a model can retrieve past patients who were in a similar acute state and check what happened to them over the next day. Doing this separately from retrieving chronically similar patients adds predictive information beyond what the patient's own trajectory and static score supply. We expect the added information to sit mostly in patients whom static scores call low-risk. We also expect that the separation of relation types, not the use of a graph neural network, is what produces the gain. The proposal is built so that this last point can fail cleanly.

Where This Sits

The research file supports two motivating findings about static scores. On the UK CCHIC database, Optimal intensive care outcome prediction over time using machine learning shows that cumulative daily data beats a fixed APACHE-II score: AUC 0.895 on day 2 against 0.836 for APACHE-II. On minute-level intraoperative vitals, Beyond Thresholds finds that latent state transitions predict outcomes even when vitals stay in normal ranges. Its largest gain, +0.223 AUROC, comes in the tertile that a static score rates lowest-risk. That stratified result is the template for our main analysis.

The relational side borrows from social-network work that separates kinds of ties. TIMME models persistent follow relations separately from transient interactions such as retweets and replies. Embedding Identity and Interest for Social Networks splits the cause of a link into stable identity and shifting interest. NetDP shows that mixing transaction, transfer and social ties into one graph hides credit-risk signal. That paper also feeds neighbor-derived features into gradient-boosted trees, a design we reuse as our cheapest implementation. Our analogy is that chronic comorbidity plays the role of identity or the follow tie, and the current acute state plays the role of the transient interaction.

Two clinical papers come closest to this idea.

Hypergraph Convolutional Networks for ICU Patient Similarity links eICU patients through shared ICD-9 codes. It reaches AUROC 0.7739 at 24 hours and 0.8141 at 48 hours, against 0.7925 for its best non-graph baseline. Its links come from admission codes, so they never change, and all diagnoses are treated alike. We differ in using a time-varying acute relation, keeping it separate from the chronic one, and predicting on a rolling basis rather than at fixed checkpoints.

A Controlled Study of Fairness Interventions for Temporal Graph Transformers does separate edge types: time-adjacent steps versus clinically similar steps. However, every edge connects steps within a single patient's own record. That study also reports a sobering result: plain LightGBM had the best AUROC (0.837) on MIMIC-IV 48-hour mortality, ahead of its graph transformers. We treat that as a constraint. No graph method gets credit unless it beats strong trees.

Does Temporal Encoding Matter? finds that explicit time encodings add nothing over event order in GRU clinical models, and that some encodings hurt pretraining. Our design therefore avoids timestamp encodings altogether. Time enters only through the six-hour binning of state.

Core Mechanism

At each prediction time $t$, a test patient $i$ is described by two vectors. The first, $c_i$, is a chronic profile built from comorbidity codes; it is fixed for the stay. The second, $a_i(t)$, is an acute-state embedding of the last six hours of vitals, labs, vasopressor and ventilation status, plus missingness indicators, following the observation in P-BiLSTM that missingness reflects clinical protocol.

A memory bank holds training-set patient-slices. Each slice carries its known outcome over the following 24 hours, $y_j$. For each relation $r \in \{\text{chronic}, \text{acute}\}$, we retrieve the $k$ nearest bank slices $N_r(i,t)$ and form a neighbor risk:

$$ \hat{p}_r(i,t) = \frac{1}{k}\sum_{j \in N_r(i,t)} y_j . $$

Here $\hat{p}_r$ is the fraction of retrieved neighbors who died or deteriorated within 24 hours. Acute neighbors are restricted to slices at a comparable time since admission.

We also record neighborhood turnover: the overlap between $N_{\text{acute}}(i,t)$ and $N_{\text{acute}}(i,t-6\text{h})$. This tests a clinical version of the observation by Jurgens et al. that ties reorganize after a shock. A patient whose acute neighbors suddenly change is moving into a different state.

The bank contains only training-period admissions. Splits follow MIMIC-IV admission-year groups, so no future outcome can leak into a prediction.

The method is implemented at two levels. The cheap version adds $\hat{p}_{\text{chronic}}$, $\hat{p}_{\text{acute}}$ and turnover as features to a rolling LightGBM. The rich version is a typed message-passing model that attends over the two neighbor sets with separate weights for each relation type, in the spirit of TIMME. Retrieval uses approximate nearest-neighbor search, so cost grows with the number of neighbors retrieved rather than quadratically with stay length. This sidesteps the memory limit that the graph-transformer study reports.

Research Program

We first build a rolling prediction task on MIMIC-IV, using the cohort pipeline of An Extensive Data Processing Pipeline for MIMIC-IV. Every six hours after ICU admission, the model predicts mortality or a deterioration event (new vasopressor or mechanical ventilation) within the next 24 hours, until discharge. Six-hour blocks follow Monitoring ICU Mortality Risk with an LSTM.

We then learn the acute embedding. It is trained only on next-window physiology, never on outcomes, to keep retrieval separate from the label. We fit the two retrieval relations, train both implementations, and run the stratified and ablation analyses below.

Next we test transfer. Models trained on MIMIC-IV are applied to eICU with an eICU-only memory bank. This checks whether the relation types carry over when the graph statistics differ, which the relational literature flags as a known weakness.

Finally, we reconstruct the fixed 24-hour and 48-hour eICU checkpoints so the results can be read against HGCN.

Baselines

The naive reference is the static score held fixed. SAPS-II and APACHE-II are computed from the first 24 hours and carried forward unchanged to every later prediction time. This is the practice the proposal criticizes. The CCHIC result gives its expected weakness: APACHE-II AUC 0.836 against 0.895 for day-2 learned models. Beating it is necessary but tells us little.

The strongest competitor is a rolling LightGBM trained on the patient's own trajectory: summary statistics, variability features of the kind shown predictive for sodium and blood pressure, missingness rates, and comorbidity indicators. Its reported benchmark is AUROC 0.837 for 48-hour MIMIC-IV mortality in the temporal graph transformer fairness study. That number came from a fixed-window task, so we re-run the model on our rolling task. It gets the same splits, the same features, and the same hyperparameter-search budget as our models. Beating it means a statistically reliable gain in AUPRC under a patient-level bootstrap. We pre-specify that the gain must hold in the lowest static-score tertile. Overall AUROC parity with a gain in that tertile would still support the claim. Parity in the tertile would not.

The sequence baseline is a GRU over six-hour blocks with comorbidity embeddings concatenated, following P-BiLSTM. This tests whether a patient's own sequence already contains whatever neighbors add. The static-relation baseline is a reimplementation of the ICD hypergraph model. We check our reimplementation against the reported 0.7739 and 0.8141 AUROC on eICU at 24 and 48 hours before using it.

Three ablations isolate the key ingredient. The first, untyped retrieval, uses one neighbor set from concatenated chronic and acute features, so the gain from separating relation types can be measured directly. The second and third keep only the acute relation or only the chronic relation.

Why not just…?

Why not just train LightGBM on the patient's own trajectory? This is the most serious threat. In principle, a flexible model fit on a patient's own features already knows the outcome rate of similar patients, so retrieval could be redundant. The file shows trees are hard to beat: they reached 0.837 and outperformed graph transformers. Our answer is empirical. The cheap implementation adds retrieval features to exactly this LightGBM. If the neighbor features add nothing, the main claim fails, and we will report that outcome. We expect retrieval to help where data are thin: rare acute states in patients who look healthy on day one. That is why the low-risk tertile is the pre-specified test.

Why not just retrieve similar patients with one similarity measure? If untyped retrieval matches typed retrieval, then the relational-heterogeneity half of the proposal adds nothing. The NetDP finding that mixing relation types hides risk signal, and HGCN's gain from similarity alone, both make this a live comparison rather than a straw man. The untyped ablation settles it directly.

Why not just use a static diagnosis hypergraph, as HGCN does? Its links are fixed at admission, so they cannot follow a patient into an acute state that the codes do not predict. The chronic-only ablation stands in for this design, and the reimplemented HGCN tests it head on.

Why not just add better time encodings to a sequence model? The evidence argues against this. Does Temporal Encoding Matter? found no downstream benefit, and our design deliberately puts no extra weight on time.

Decisive Checks

Four results decide whether the idea holds. First, acute-neighbor risk must improve AUPRC over the rolling LightGBM in the lowest static tertile, with a bootstrap interval that excludes zero. Second, typed retrieval must beat untyped retrieval. Third, the cheap tree implementation and the typed message-passing model must be compared; if the trees match the neural model, we will recommend the trees.

Fourth, we run a secondary fairness check against the race AUROC gaps reported in the graph-transformer study: 0.129 for the full graph transformer and 0.135 for the LSTM. That study found that similarity edges within a patient's record narrowed the race gap. We test whether acute-similarity edges between patients behave the same way, without any per-subgroup thresholds.

Turnover is a separate probe. We check whether a drop in neighbor overlap precedes deterioration events, which would give clinicians a readable signal rather than a bare score.