Many important clinical questions concern rare combinations: two drugs that are seldom co-prescribed, or a diagnosis that rarely follows another. For these, the binding constraint is statistical power, meaning the chance of detecting an effect that is really there. A health system cannot tell from today's counts how many more years of data, or how many partner hospitals, it would need before such a combination becomes testable. This proposal treats that as a forecasting problem.
The approach draws on network densification, the empirical finding that as real networks grow, links accumulate faster than nodes. In symbols, edge count $E$ scales as node count $N$ to a power $\alpha$ between 1 and 2. We read a clinical co-occurrence graph's densification curve as a record of how fast previously unseen or rare pairs of clinical events fill in as patients accumulate. Fitted correctly, that curve gives a population-level prior for how quickly a specific rare pair's event count will grow, and therefore when it will cross a power threshold. The catch, and the main scientific content, is that the curve is fitted from data that are both fragmented across institutions and bounded in time. We argue that these two limitations bias the fitted exponent in ways that can be derived and corrected.
Core Claim
Densification is already documented in a biomedical setting. Jeong et al. (Revealing the dynamic landscape of drug-drug interactions through network analysis) extracted 21,262 drug pairs from 160,114 PubMed articles and showed that documented interactions grow super-linearly in the number of drugs. They used densification only to describe the literature. They did not use it to predict which pairs will be documented next, and they did not study patient records.
Our claim has two parts. First, a rare pair's expected count can be forecast by borrowing strength from the whole graph's densification. Second, two features of real data distort the observed exponent:
- Fragmentation. A single site sees only part of each patient's care.
- Temporal truncation. The observation window is short, or crosses a coding change such as the move from ICD-9 to ICD-10.
Correcting both distortions should produce calibrated time-to-power forecasts precisely where per-pair trends cannot be estimated, which is for pairs with a handful of events or none.
The fragmentation bias follows from a short argument. Suppose a site captures each of a patient's clinical events independently with probability $q$. Lasko's work on laboratory testing (Nonstationary Gaussian Process Regression for Evaluating Repeated Clinical Laboratory Tests) shows that external care leaves exactly such voids. Under this assumption, a single event is recorded with probability $q$, but a co-occurring pair is recorded only with probability about $q^2$. Pair counts are therefore deflated more than node counts. At sites with low capture, the graph should look less dense, $\alpha$ should look smaller, and rare pairs should look rarer than they are, so per-site power forecasts should be systematically pessimistic. Pooling sites raises the effective $q$ but also duplicates patients who visited several of them. That is an empirical prediction, not an assumption.
The Forecasting Model
We work with a graph whose nodes are clinical codes: diagnoses, drug ingredients, and binned lab abnormalities. The weight on an edge is the number of patients in whom both codes co-occur within a fixed window. Let $P(t)$ be cumulative patients at time $t$. Let $\lambda_{uv}(t)$ be the expected co-occurrence count for codes $u$ and $v$. We model
$$\lambda_{uv}(t) = c_{uv}\, P(t)^{\beta_{uv}},$$
where $c_{uv}$ is a pair-specific scale and $\beta_{uv}$ is a pair-specific growth exponent.
The exponent is pooled hierarchically toward a global value implied by the graph's densification. The pooling also depends on the degree class of the two endpoints, because Jeong et al. find growth concentrated within drug classes and leave cross-class growth unexplored. The fragmentation correction rescales observed counts by estimated capture rates $q_s$, one per site $s$. The coding-epoch correction maps codes to a common granularity before fitting, or models the ICD transition as a change-point in node count.
For a target relative risk $\rho$, a standard Poisson test needs some count $n^{*}(\rho)$ to reach 80% power. The forecast is the posterior distribution of the crossing time $t^{*}$ at which $\lambda_{uv}(t^{*}) = n^{*}(\rho)$. The same machinery gives a "sites-to-power" answer: the number of additional hospitals, with their estimated capture rates, needed to cross the same threshold.
Research Program
The work has three stages.
- Synthetic validation. We generate graphs from the Multiplicative Attribute Graph model (Multiplicative Attribute Graph Model of Real-World Networks). This is a generator in which link probability is a product of attribute affinities and which provably densifies. We use it here only as a ground-truth simulator, not as the predictive model. On these graphs we impose site-level capture masks and truncated windows, and check that the $q$-versus-$q^2$ bias and its correction behave as derived.
- Measurement on real records. We fit densification on real electronic health records and measure how the fitted exponent varies with site, window length, and the ICD transition. This also answers the open question of whether patient-record graphs densify at all. Thomas et al. (Predicting onset of complications from diabetes: a graph based approach) built a heterogeneous clinical graph from the Indiana Network for Patient Care, but only within a single fixed window on a centralized exchange.
- Prospective-style evaluation. We train on early periods and evaluate forecasts on later periods, never on random splits. DDI-Ben (Benchmarking Graph Learning for Drug-Drug Interaction Prediction) shows that random splits inflate performance badly: EmerGNN's F1 falls from 56.9% to 34.0% under distribution shift on DrugBank.
Baselines
Naive reference. For each pair, we take its current event rate per patient, extrapolate the count linearly in cumulative patients, and apply the same Poisson power formula. This is what a power calculation done today implicitly assumes. All methods receive identical training periods, code mappings and thresholds, so any difference comes from the forecasting model alone.
Per-pair trend regression. This is the strongest structure-free competitor. It fits a Poisson regression of each pair's count on log cumulative patients, using only that pair's history, with empirical-Bayes shrinkage of the rate across pairs. Shrinkage of this kind is the standard tool for sparse co-occurrence signals. It is the main threat because, for pairs with enough history, it should match any network-informed method. We give it the same hyperparameter search budget as our model, tuned on the same validation period.
Distinct-pair accumulation curve. This extrapolates how many new pairs appear per added patient and assigns each unseen pair an equal share. It isolates whether the pair-level structure in our model matters beyond the global curve.
Ablations. We run the full model three ways: without the capture-rate correction, without the coding-epoch correction, and with a single global exponent instead of degree-class-dependent exponents. The capture-rate ablation is the key one, since it tests the proposal's central ingredient.
Prior-number caveat. We found no reported result on Lacuna for forecasting time-to-power or the emergence of rare clinical co-occurrences, so there is no published baseline number to beat. The nearest reported figures describe how hard the rare tail is, not this task: PubMedBERT reaches 86.86% AUROC but only 17.24% PR-AUC on MIMIC-IV emergency diagnosis prediction (Revisiting Clinical Outcome Prediction for MIMIC-IV).
What counts as beating the baselines. On held-out periods, among pairs with fewer than five events at forecast time, our model must lower the Brier score for "crosses the power threshold within the horizon" relative to per-pair shrinkage. The improvement must exclude zero under a bootstrap over pairs. Calibration slope must stay near one. Matching shrinkage on frequent pairs is acceptable.
Why not just…?
Why not just run a standard sample-size calculation on current rates? This is the naive baseline. It assumes counts grow in proportion to patients, which densification contradicts at the graph level. Whether the contradiction matters for individual rare pairs is exactly what the first comparison measures. If linear extrapolation turns out to be calibrated, the densification idea adds nothing for planning, and we will report that.
Why not just fit each pair's own trend with empirical-Bayes shrinkage? This is a real threat. For pairs with a usable history it should do as well as anything. Our bet is narrower: pairs with zero to four events have no estimable trend, and shrinking toward a single global rate ignores that growth rates differ by endpoint degree class. If the degree-class ablation shows no gain over shrinkage, the network component is unnecessary and the contribution reduces to the fragmentation correction.
Why not just pool all sites or federate, and skip the forecast? Pooling raises effective capture. Decentralized training (Learn Electronic Health Records by Fully Decentralized Federated Learning) makes pooled modeling feasible across 20 hospitals without moving records. But neither tells a consortium how many partners or years a given question needs, which is the planning decision this proposal targets. Moreover, a naive pooled exponent inherits capture-rate heterogeneity across sites. The site-addition experiment below shows whether that heterogeneity matters.
Why not just use species-accumulation estimators for unseen pairs? These are the classical estimators of how many new kinds appear as sampling continues. They predict how many new pairs appear, not which ones. The accumulation-curve baseline tests whether that pair-level resolution is worth anything.
Proposed Experiments
MIMIC-IV. We build yearly-group code graphs using the coarse calendar-year groups released with the database (An Extensive Data Processing Pipeline for MIMIC-IV). We fit on early groups and evaluate crossing forecasts on later ones. We deliberately straddle the ICD-9 to ICD-10 transition to test the epoch correction. The prediction is that an uncorrected fit shows an artificial jump in node count and a fall in $\alpha$ at the transition.
eICU. We use its many hospitals for the fragmentation test, with two analyses:
- Per-site versus pooled exponents. We compare per-site against pooled exponents and check whether per-site estimates are biased downward in proportion to proxy capture rates.
- Site addition. We add hospitals in random orders and score sites-to-power forecasts against the realized crossings.
We also thin patients' records synthetically to known capture rates. The derived $q$-versus-$q^2$ deflation should then appear quantitatively.
PubMed interaction graph. For a long-horizon external check, we rebuild the interaction graph following Jeong et al.'s pipeline and forecast, from data through an earlier cutoff, which drug pairs are first reported in later years. The specific split years are to be fixed after the 1987–2022 timeline is reconstructed. In this setting temporal scope is decades rather than years.
Decisive check. The proposal stands or falls on two results in the rare-pair stratum, judged on held-out periods only:
- The full model must beat per-pair shrinkage in calibrated crossing forecasts.
- Removing the capture correction must measurably worsen per-site forecasts.
If either fails, we will treat densification as descriptive rather than predictive for clinical power planning.
References
- Revealing the dynamic landscape of drug-drug interactions through network analysis
- Nonstationary Gaussian Process Regression for Evaluating Repeated Clinical Laboratory Tests
- Multiplicative Attribute Graph Model of Real-World Networks
- Predicting onset of complications from diabetes: a graph based approach
- Benchmarking drug–drug interaction prediction methods: a perspective of distribution changes
- Revisiting Clinical Outcome Prediction for MIMIC-IV
- Learn Electronic Health Records by Fully Decentralized Federated Learning
- An Extensive Data Processing Pipeline for MIMIC-IV