Longitudinal cohorts such as Alzheimer's studies measure every enrolled patient on a fixed calendar, for example once a year. Yet the information that matters is concentrated. Most patients change slowly, while a minority pass through short windows in which risk rises sharply. In MCI, plasma p-Tau217 between 0.08 and 0.56 pg/mL corresponds to four-year conversion risk rising from 10% to 45% (Chen et al.). Chen et al. worked from a single baseline draw and point to longitudinal tracking as the next step. The practical question for a study with a fixed assay budget is which patients should be brought in this quarter.
We propose to answer it with a determinantal point process (DPP). A DPP is a random way of picking a subset that favours items that are individually valuable but penalises picking items that are similar to one another. In our setting, each quarter's batch of patients would be drawn from a DPP. A patient's value is their current probability of being inside a clinically critical window. The similarity between two patients is their posterior correlation under a population trajectory model.
The part that matters most is that the batch is sampled at random rather than chosen by greedy maximisation. For a DPP, each patient's probability of being selected, and each pair's probability of being selected together, has a closed form. So the same adaptive schedule that concentrates visits on patients near the window also supports design-unbiased population estimates with valid confidence intervals. Examples are the cohort's mean rate of decline and the fraction of fast progressors. A greedy uncertainty-driven schedule cannot provide these guarantees when the trajectory model is wrong.
Core Claim
Adaptive follow-up schedules usually trade two goals against each other. Patient-level detection wants visits aimed at the people most likely to be crossing a threshold. Population-level science wants something like a representative sample. Randomised selection with unequal probabilities reconciles the two, provided the probabilities are known. A DPP adds a further benefit. Because its selections are negatively correlated, it spreads the budget across dissimilar trajectories, which should give lower-variance weighted estimates than independent (Poisson) sampling at the same expected probabilities.
The claim has three testable parts. First, DPP-sampled schedules detect critical-window entry nearly as quickly as greedy variance-driven schedules. Second, they keep confidence-interval coverage near nominal for cohort estimands even when the trajectory model is misspecified, which greedy schedules do not. Third, they achieve smaller estimator variance than independent randomised sampling with identical marginal inclusion probabilities.
What Exists and What Does Not
Three prior papers supply the machinery we reuse.
- Kathuria et al. samples batches for parallel Bayesian optimisation from a DPP whose kernel is the Gaussian-process posterior covariance. They show that greedy DPP maximisation is equivalent to the "hallucinate observations" heuristic UCB-PE, and that true sampling explores better. Their goal is regret. Ours is estimation and detection, and we use the randomness for a different reason: to obtain known inclusion probabilities, not to explore.
- Biyik et al. uses the fact that a DPP conditioned on already-acquired items is still a DPP to avoid re-querying explored regions in batch active learning. We reuse this conditioning across successive visit waves.
- Fast Dynamic Sampling for DPPs provides low-rank samplers for repeated draws. This matters for the dense MIMIC-IV setting.
The clinical DPP work so far is static. BAREB puts a DPP prior on cluster locations for cross-sectional periodontal data, and nothing in it decides when to measure. Yuan and Kitani use a DPP loss to diversify forecasted trajectories, but in kinematics, and the loss shapes a generator rather than a measurement design.
The longitudinal-modelling papers motivate the problem without solving it. IMMFM, the Longitudinal Siamese Network and the INSPIRE hidden Markov model of Beyond Thresholds all model trajectories from whatever visits happened to be collected. None chooses the visits. We are not aware of work that treats follow-up scheduling as DPP sampling, or that uses DPP inclusion probabilities for design-based inference in any domain covered here.
Two papers constrain the design.
- Does Temporal Encoding Matter? found that explicit time encodings added nothing to GRUs on dense MIMIC-IV event streams. We therefore do not assume continuous-time structure helps in the dense regime. We test the method separately in sparse outpatient data (ADNI) and dense inpatient data (MIMIC-IV), and we expect the gains to be larger in the sparse case.
- IMMFM reports fidelity decay at long horizons. This argues for a simple, well-calibrated trajectory model as the kernel source rather than a heavy generative model.
Research Program
Time is divided into visit slots, for example quarters in ADNI and hours in MIMIC-IV. At slot $w$, the ground set is the set of patients still under follow-up. A trajectory model supplies a posterior predictive for each patient's biomarker at that slot. We use a hierarchical Gaussian-process mixed-effects model with patient-level random slopes and curvature, fitted to all visits observed so far.
Quality. Patient $i$'s quality $q_i$ is the posterior probability that their biomarker lies in, or will cross into, a prespecified critical window before the next slot. We add a floor $\epsilon$ so that every patient keeps a positive selection probability.
Similarity. $S_{ij}$ is the posterior correlation between patients $i$ and $j$'s current-slot values under the population model. Two patients whose trajectories the model explains through the same shared components are redundant: measuring one tells you much of what measuring the other would.
Kernel and sampling. The kernel is $L_{ij} = q_i S_{ij} q_j$. A batch of fixed size $k$ (the slot's assay budget) is drawn from the corresponding k-DPP, conditioned on the previous slot's selections in the manner of Biyik et al.
Inclusion probabilities. For an ordinary DPP, the marginal kernel is
$$K = L(L+I)^{-1}.$$
It gives each patient's inclusion probability $\pi_i = K_{ii}$ and each pair's joint inclusion probability $\pi_{ij} = K_{ii}K_{jj} - K_{ij}^2$. The joint probability is always at most $\pi_i \pi_j$, which is the negative dependence we rely on. For the fixed-size k-DPP, the same quantities follow from elementary symmetric polynomials of the eigenvalues of $L$.
Estimation. For any cohort quantity written as a sum over patient-slots, $\tau = \sum_{w,i} y_{wi}$, we use the sequential Horvitz–Thompson estimator
$$\hat\tau = \sum_{w}\sum_{i \in Y_w} \frac{y_{wi}}{\pi_{wi}}.$$
Here $y_{wi}$ is patient $i$'s biomarker-derived quantity at slot $w$, $Y_w$ is the sampled batch, and $\pi_{wi}$ is the inclusion probability conditional on the history up to slot $w$. Because each slot's selection is randomised given the past, $\hat\tau$ is unbiased whether or not the trajectory model is correct. Its variance estimator uses the pairwise probabilities $\pi_{ij}$.
Slope-type estimands, such as mean annual decline and the fraction of patients in a fast-decline class, are fitted by inverse-probability-weighted estimating equations using the same weights.
Tuning. The floor $\epsilon$ and a power on $q$ control how aggressively the schedule chases risk. We will map the trade-off between detection delay and estimator variance as these two settings vary.
Baselines
Every baseline receives the same per-slot assay budget $k$, the same retrospective ground set, and the same trajectory model refitted after each slot. Tuning parameters are chosen on a held-out ADNI split with an equal grid for every method.
No reported result found on Lacuna for adaptive follow-up scheduling on ADNI or MIMIC-IV. The external reference points are therefore full-data numbers:
- Chen et al.'s four-year conversion AUROC of 0.787 from baseline plasma p-Tau217, versus 0.758 for amyloid PET (Chen et al.).
- The Longitudinal Siamese Network's MMSE trajectory AUC of 0.99 on ADNI and 0.883 on AIBL.
We report what fraction of the full-data performance each budgeted schedule recovers.
Naive references. The first is the fixed-calendar schedule: each patient is visited every $n/k$ slots on a staggered grid. It mirrors ADNI's annual and AIBL's 18-month cadences. The second is uniform random selection of $k$ patients per slot. It is the simplest design with known inclusion probabilities, and it sets the coverage standard.
Greedy conditional variance. This is the strongest competitor on patient-level detection. At each slot it selects the $k$ patients one by one, choosing the largest risk-weighted posterior variance and conditioning after each pick. By Kathuria et al.'s equivalence this is DPP maximisation (UCB-PE style). It gets model-based likelihood estimates, which are valid under the missing-at-random assumption when the model is correct.
- On detection, we count a draw if the DPP is within 10% of this baseline's median detection delay.
- On estimation, the DPP beats it if its coverage stays within two points of nominal under misspecification while greedy's does not.
Independent Poisson sampling. This is the key ablation for the "determinantal" ingredient. Each patient is included independently with exactly the DPP's marginal $\pi_i$, and the batch is resized to $k$ by conditional Poisson sampling. Identical marginals isolate the effect of the negative pairwise dependence. The DPP wins if its Horvitz–Thompson variance is lower at equal coverage.
Quality-only ablation. Setting $S = I$ turns the DPP into probability-proportional-to-risk sampling. This ablation isolates the role of the population-covariance similarity.
Why not just…?
Why not just greedily pick the most uncertain, highest-risk patients each quarter? This is the real threat for detection. We expect greedy selection to detect window entry equal to or slightly faster than the DPP, since it spends nothing on low-risk patients. Its weakness is inference. It assigns zero probability to some patients, so design-based estimation is impossible. Its validity then rests entirely on the trajectory model being right, and adaptive designs make misspecification bias worse because the observed patients differ systematically from the unobserved ones. Kathuria et al. also show that greedy batches over-concentrate around variance peaks. If greedy's coverage survives our misspecification scenarios, the inference argument for the DPP fails, and the proposal reduces to a detection comparison that it may lose.
Why not just randomise independently with risk-proportional probabilities? Independent risk-proportional sampling already yields known inclusion probabilities and unbiased Horvitz–Thompson estimates, so the DPP must earn its complexity through lower variance and less redundancy within a batch. The matched-marginal Poisson ablation measures exactly this. If the variance gap is negligible, which is plausible when patients' posteriors are nearly uncorrelated, we will recommend the simpler scheme and report that negative dependence did not help. The DPP should have more room to help in the dense MIMIC-IV setting, where posterior correlations between patients are larger.
Why not just keep the fixed calendar and model better? Fixed grids are easy to run and are unbiased by construction. However, a critical window as narrow as p-Tau217's can be crossed between annual visits. The static-versus-trajectory gap on INSPIRE also shows the cost of missing dynamics: area under the precision–recall curve rose from 0.244 to 0.339 when trajectory features were added (Beyond Thresholds). The detection-delay comparison against the calendar baseline settles this question directly.
Proposed Experiments
Retrospective replay. In ADNI, the full observed visit history serves as ground truth. We use MMSE and ADAS-13 across up to nine timepoints, MRI ventricular volume, and plasma p-Tau217 where repeated draws exist. Schedules may only reveal the visits they select. Budgets are 25%, 40% and 60% of the observed visits.
Metrics. We record three:
- median delay from true window entry to detection, where the window is p-Tau217 in 0.08–0.56 pg/mL or assignment to the fast-decline ADAS-13 class;
- bias, root mean squared error and 95% interval coverage for mean annual decline and for the fast-decliner fraction;
- four-year conversion AUROC of a survival model refit on the selected visits, expressed as a fraction of the full-data AUROC.
Misspecification check. Coverage is the decisive test. We fit the scheduling model with a deliberately wrong structure: linear slopes when the true trajectories are S-shaped, and the APOE4 interaction omitted. We then check whether DPP-weighted estimates keep near-nominal coverage while greedy model-based estimates lose it.
External replication. AIBL tests whether the tuning chosen on ADNI transfers across cohorts.
Dense regime. In MIMIC-IV, the ground set is hourly slots of patients eligible for a sodium or creatinine draw. Here we test whether the advantage shrinks where, per Does Temporal Encoding Matter?, event order already carries most of the temporal information. At this scale we use the low-rank samplers of Fast Dynamic Sampling and report wall-clock cost per slot.
What would sink the idea. The proposal fails if greedy selection matches the DPP on coverage under misspecification, or if matched-marginal Poisson sampling matches the DPP on variance.
References
- Predicting Progression from Mild Cognitive Impairment to Alzheimer’s Disease using Survival Analysis
- Batched Gaussian Process Bandit Optimization via Determinantal Point Processes
- Batch Active Learning Using Determinantal Point Processes
- Fast Dynamic Sampling for Determinantal Point Processes
- BAREB: A Bayesian repulsive biclustering model for periodontal data
- Diverse Trajectory Forecasting with Determinantal Point Processes
- Longitudinal Flow Matching for Trajectory Modeling
- Modeling and prediction of clinical symptom trajectories in Alzheimer's disease using longitudinal data
- Beyond Thresholds: Learning Perioperative Risk from Intraoperative Physiological Trajectories
- Does Temporal Encoding Matter? Evidence from GRU-based Clinical Foundation Models