A person's inherited risk for a disease such as heart attack can be summarized by a polygenic risk score: a weighted sum of thousands of small-effect genetic variants. Once a person's blood pressure, cholesterol and weight have been measured for years, that score adds little, because the genetic risk has already shown up in those measurements. Much of the population has no such record, though. Real health records are sparse and irregular. Many people have one lab panel a decade, and many young adults have nothing. We propose that the value of genetic information depends on how much of a patient's own clinical trajectory has been observed. Age matters mainly as a proxy for observation. We propose a risk model in which genetic evidence enters the prediction in proportion to how little the record shows. That evidence comes in two forms: the patient's own pathway-level risk scores, and the observed outcomes of genetically similar past patients.
The model is additive by construction. Each prediction splits exactly into three parts: what the patient's own measurements say, what their genes say, and what happened to their genetic neighbors. A weight shows how much the model trusted each part at that moment in the record. This gives clinicians an explanation that is faithful to the computation, rather than a post-hoc attribution laid over a black box. The work also tests directly whether a patient-similarity graph earns its place against a well-tuned tabular model, a test that most published population-graph work has failed.
Core Claim
Highly elevated polygenic risk scores are better predictors of myocardial infarction risk early in life than later motivates this proposal. That study used logistic regression on about 440,000 UK Biobank participants. It found that adding a polygenic score to sex and ancestry raised the area under the ROC curve (AUC) from 0.70 to 0.79 for heart attack at or before age 50, but only from 0.67 to 0.73 for heart attack at any age. It attributes the decline to genetic risk "leaking" into measured clinical factors. Their analysis stratifies by age and cannot separate getting older from having been measured more.
Our first claim is empirical and falsifiable. Within a fixed age band, the extra predictive value of genetics should be larger for patients with fewer prior clinical observations. Our second claim is methodological. If this interaction holds, a model that routes genetic evidence through a gate driven by sparsity should beat a model that treats the polygenic score as one more static column, especially for patients with thin records. This advantage should hold under strict leakage controls.
The Model
Each patient $i$ has an irregular history of labs and vitals up to a prediction time $t$. A time-aware recurrent encoder reads this history. In the spirit of Temporal Belief Memory, its hidden state drifts back toward a population baseline as the time since the last measurement grows. The encoder produces a summary $h_i(t)$ and an observation-density signal $s_i(t)$. That signal is built from measurement counts and recency, and optionally from the encoder's own predictive variance.
Genetic input is a vector $r_i$ of pathway-level polygenic scores rather than one genome-wide number. Each score is computed with Bayesian shrinkage (PRS-CS-auto) restricted to a biological gene set. This follows the pathway and module scores used for depression (NETRIN1 pathway PRS) and Alzheimer's disease (module-level PRS NetWAS). Those studies show that splitting the score by pathway yields mechanistically interpretable components.
The population graph links each patient to the $k$ training-set patients nearest in pathway-score space, with similarity also matched on sex and age band. Edges never come from the clinical features themselves. This is deliberate. Are Population Graphs Really as Powerful as Believed? showed that graphs built from the same features a tabular model already sees mostly add noise. Genetic neighbors instead carry something a single record lacks: the realized outcomes of people with similar inherited architecture. We do not run a deep graph network. We use a single attention-weighted average of the neighbors' outcome labels. Each neighbor's weight is a learned function of the genetic similarity between the two patients. The prediction is
$$\text{logit}\, p_i(t) = a\big(h_i(t)\big) + w\big(s_i(t)\big)\Big[\beta^\top r_i + \gamma \sum_{j \in N(i)} \alpha_{ij}\, y_j\Big],$$
where the symbols are defined as follows. $a$ is the own-trajectory risk. $w$ is a gate constrained to fall monotonically as observation density rises. $\beta$ are linear pathway weights. $N(i)$ is the genetic neighborhood. $\alpha_{ij}$ are normalized neighbor weights. $y_j$ is neighbor $j$'s observed outcome. $\gamma$ scales the neighbor term.
We borrow the "reluctant" discipline of RAITR, which defaults to a sparse linear model and admits nonlinear terms only if they pass a penalized concordance criterion. Under that discipline, the neighbor term and any nonlinearity in $a$ are retained only if they improve held-out concordance over the linear-genetics model. The explanation shown to a clinician is the three additive terms, the gate value, the named pathways, and the neighbors with their outcomes. These quantities are the computation itself, not an approximation of it.
Relation to Prior Work
The closest relational work is An Interpretable Population Graph Network to Identify Rapid Progression of Alzheimer's Disease Using UK Biobank. It builds a clinical-similarity graph with a hybrid Jaccard and Chebyshev distance, binarizes time into rapid versus non-rapid progressors, and explains predictions after the fact with GNN Explainer and SHAP. We differ in three ways: our edges are genetic, our model respects irregular time, and our explanations are additive by design.
That paper reports GraphSAGE at an AUROC of 0.976 against a Random Forest at 0.532 for MCI-to-AD progression. The Müller et al. benchmark found the opposite pattern, with a Random Forest at 0.9474 accuracy against the best GNN at 0.9404 on TADPOLE. A gap this large usually signals transductive leakage or an undertuned baseline. We therefore treat it as a claim to re-test under inductive splits, not as support for graphs.
EEPRS embeds ICD-10 histories and runs genome-wide association studies on the embedding dimensions. It reports average gains of 24.06% in AUC and 92.48% in $R^2$ over single-trait scores. It shows that clinical histories carry heritable signal. We supply an EEPRS-style multi-trait score as an alternative genetic input. EEPRS, however, combines information at the level of summary statistics and has no notion of record sparsity.
The continuous-time imputation literature, including Continuous-time Autoencoders and Diff-MN, models each series in isolation. It also evaluates mostly on non-clinical data. We use a continuous-time encoder only as an ablation. The core mechanism is the gate, not the solver.
Research Program
The study runs in UK Biobank on participants of European ancestry, the setting of the Isgut et al. comparison. The primary task is incident heart attack within ten years of a prediction time. Inputs are linked primary-care and hospital labs and vitals, which are naturally sparse. A second task is MCI-to-AD progression, chosen to re-examine the graph claim of Meng et al.
Splits are by patient and by time. Neighbors are drawn only from training patients whose outcomes were observed before the test period. No test patient's label or features touch the graph.
Phase one is the stratified analysis alone. Within five-year age bands, we fit a logistic model with and without pathway scores and estimate how the AUC gain changes with the number of prior measurements. Phase two fits the gated model and its ablations.
Phase three is an intervention on sparsity. We take data-rich test patients and delete their measurements at random, down to the density of sparse patients. We then check two things. The first is whether the gate rises and the genetic share of the logit grows as predicted. The second is whether accuracy on the thinned records stays closer to the full-record accuracy than a tabular baseline's does.
Ancestry transfer is a known weakness of polygenic scores (portability across diverse populations). We report results on non-European UK Biobank participants as a stress test, not as a claim.
Baselines
The naive reference point uses birth-established factors plus a genome-wide polygenic score in logistic regression, reproducing Isgut et al. They report an AUC of 0.73 for general heart attack and 0.79 for early heart attack. A second naive reference uses the same model plus the most recent value of each clinical measurement, with missing values flagged. Any proposed model that fails to beat this second reference has no reason to exist.
The strongest competitor is gradient-boosted trees on concatenated features. The feature set includes the last values, counts and time-since-last for every measurement, the full pathway-score vector, age, and explicit polygenic-score-by-density interaction features. It also includes a "neighbor outcome" feature computed by the same leakage-safe genetic $k$-nearest-neighbor procedure. This baseline can in principle learn everything our model does: the gate, the interaction, and the neighbor evidence. The Müller et al. benchmark shows that a well-tuned tabular model of this kind matches or beats population-graph networks on TADPOLE, ABIDE (0.7046 Random Forest against 0.695 GNN accuracy), and UK Biobank brain age (Ridge MAE 3.41 against 3.35, not significantly different). No reported result was found on Lacuna for a boosted-tree model on longitudinal UK Biobank heart-attack prediction, so we produce the number ourselves.
Both models receive identical splits, identical inputs, and the same hyperparameter search budget, set by trials and GPU-hours and tuned on the same validation fold. The proposal beats this baseline only under two conditions. The first is a statistically significant AUC gain in the lowest-density third of patients, with confidence intervals from a bootstrap over patients. The second is no loss in calibration or AUC overall. Matching it everywhere while offering exact additive explanations would be a weaker but still reportable outcome.
Three ablations isolate the key ingredient. In the first, the gate is fixed to a constant, which turns genetics into a static feature. In the second, the graph edges are shuffled while each patient keeps the same number of neighbors. In the third, the pathway scores are replaced by a single genome-wide score. A fourth variant swaps in a controlled-differential-equation encoder to test whether continuous time matters.
Why not just…?
Why not just give a gradient-boosted tree all the features, including the polygenic score? This is a real threat. Trees can learn polygenic-score-by-density interactions, especially when we hand them those interactions explicitly, as we do in the baseline. The proposal's defensible advantages are two. One is a monotone, data-efficient gate that should generalize better where sparse, low-density patients are scarce in training. The other is an explanation that is exact rather than post-hoc. If the tree matches us on the sparse third and on the thinning experiment, we will report that the sparsity effect is real but needs no special architecture. The contribution would then shrink to the empirical finding and an interpretable equivalent.
Why not just use the genome-wide score as one static feature? The constant-gate and genome-wide-score ablations answer this directly. Isgut et al.'s own numbers show the static score's value is not constant across life stages. EEPRS and the module-level scores both found that a single genome-wide number leaves signal unused. If the static feature does as well, the gate is unnecessary.
Why not just forward-fill measurements into an LSTM or a time-embedded Transformer? Forward-filling erases the information carried by a missing test, and the gate is built from exactly that information. Our default encoder is already a cheap recurrent model with time decay, not an ODE solver. The forward-filled LSTM is included as an encoder ablation. We expect the encoder choice to matter less than the gate. If it matters more, that is informative too.
Why not just explain a black-box model with SHAP? SHAP treats patients as independent rows. It cannot faithfully separate a patient's own signal from evidence borrowed from neighbors. Meng et al. needed a separate graph explainer for this and still produced only post-hoc summaries. Our additive form makes each contribution an exact term in the computation. We will test whether that matters in practice by deleting the top-attributed inputs and measuring how much the prediction changes, under both approaches.
Decisive Checks
The program stops or changes course at four points. First, if the within-age analysis shows no interaction between genetic gain and record density, the gate hypothesis is false, and we would publish that as a negative refinement of the leakage account. Second, the gated model's genetic share must rise as records are artificially thinned. If it does not, the gate is fitting something other than sparsity. Third, if shuffling the edges costs nothing, the neighbor term is dropped and the graph claim is abandoned. Under the reluctant-inclusion rule, the model should then retain only pathway genetics, consistent with Müller et al. Fourth, the Meng et al. Alzheimer's comparison is rerun under inductive splits with a tuned tree. If the reported graph advantage disappears, we state this plainly, because that finding constrains every graph-based claim in this area, including ours.
References
- Highly elevated polygenic risk scores are better predictors of myocardial infarction risk early in life than later
- Temporal Belief Memory: Imputing Missing Data during RNN Training
- Association of Whole-Genome and NETRIN1 Signaling Pathway–Derived Polygenic Risk Scores for Major Depressive Disorder and White Matter Microstructure in the UK Biobank
- A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study
- Are Population Graphs Really as Powerful as Believed?
- A reluctant additive model framework for interpretable nonlinear individualized treatment rules
- An Interpretable Population Graph Network to Identify Rapid Progression of Alzheimer’s Disease Using UK Biobank
- Improving polygenic risk prediction performance through integrating electronic health records by phenotype embedding
- Continuous-time Autoencoders for Regular and Irregular Time Series Imputation
- Diff-MN: Diffusion Parameterized MoE-NCDE for Continuous Time Series Generation with Irregular Observations
- Polygenic risk score portability for common diseases across genetically diverse populations