Robots that build maps from video face a choice about which camera frames to keep. Keeping every frame wastes computation, and keeping frames on a fixed clock misses the moments that matter. The best systems keep a new "keyframe" only when doing so would sharply reduce uncertainty about the robot's position. Long-term studies of Alzheimer's disease face the same choice with a much higher cost per frame. Each amyloid PET scan, which images protein buildup in the brain, is expensive and adds radiation dose. Cohorts still schedule scans on a fixed calendar, roughly every two years for everyone. Meanwhile the most powerful genetic studies of the disease now test genes against continuous brain measurements rather than yes/no diagnoses. These measurements are called quantitative endophenotypes. The rate at which a person accumulates amyloid is among the most direct of them.
We propose to borrow the logic of keyframe selection to decide whom to scan and when, under a fixed scan budget. The goal is to estimate as precisely as possible how a genetic variant changes the rate of amyloid accumulation. The bet is that a fixed calendar wastes most scans on people whose trajectories are currently uninformative. Two groups fall into this category. Amyloid-negative people sit on a flat floor, and late-stage people sit on a ceiling. Meanwhile the scheduler would starve people in the steep middle of the curve. An adaptive, information-targeted schedule could therefore match the genetic power of a full cohort with a fraction of the scans. Alternatively it could deliver more power from the same scans.
Core Claim
The quantity a scan should reduce uncertainty about is the cohort-level genetic effect on progression rate, not the individual's own disease state. We build on the progression model of Bilgel et al., which assigns each subject $i$ at visit $j$ a progression score $s_{ij} = \alpha_i t_{ij} + \beta_i$. Here $t_{ij}$ is age, $\alpha_i$ is the subject's progression rate, and $\beta_i$ is the subject's baseline offset. We extend it so that the rate depends on genotype: $\alpha_i = a + \gamma g_i + u_i$. In this equation $g_i$ is the allele count at a SNP, $\gamma$ is the genetic effect we want to estimate, and $u_i$ is subject-level random variation. Regional amyloid is modeled as $y_{ijv} = f_v(s_{ij}) + \epsilon_{ij}$. The function $f_v$ is a saturating curve for region $v$, which captures the plateau effects that Bilgel et al. flag as a limitation of their linear version. The noise term $\epsilon_{ij}$ is spatially correlated with covariance $C(\rho)$.
Given the data seen so far, the value of scanning subject $i$ at a candidate time $t$ is the expected gain in Fisher information about $\gamma$:
$$\Delta_i(t) = \mathbb{E}\left[\, I_\gamma(\mathcal{D} \cup \{y_i(t)\}) - I_\gamma(\mathcal{D}) \,\right],$$
Here $\mathcal{D}$ is the data acquired so far and $I_\gamma$ is the marginal Fisher information about $\gamma$ after integrating out each subject's $(\alpha_i, \beta_i)$. Fisher information measures how sharply the data pin down a parameter, and its inverse bounds the achievable variance. A greedy scheduler spends the budget on the highest-$\Delta$ (subject, time) pairs. This mirrors the greedy entropy-reduction point selection in Information-Driven Direct RGB-D Odometry.
Because $f_v$ saturates, $\Delta_i(t)$ is near zero for subjects on the floor or ceiling and large in the steep middle. It also naturally balances genotype groups. This is the endophenotype analogue of keyframe density adapting to scene difficulty in Online 3D Reconstruction and 6-DoF Pose Estimation for RGB-D Sensors.
Relation to Prior Work
The SLAM papers supply the method but not the target. Lim et al. trigger keyframes when the Fisher-information determinant of the pose drops, and they keep 1.4% of frames on one TUM RGB-D sequence. Redesigning SLAM for Arbitrary Multi-Camera Systems thresholds a running pose-entropy filter. Wudenka et al.'s monocular system (paper) retrospectively promotes the previous frame when entropy spikes. All of these target the platform's own state, which here corresponds to subject $i$'s $(\alpha_i, \beta_i)$. Our proposal differs in targeting a parameter shared across the cohort. We test explicitly whether that difference matters.
The longitudinal imaging papers supply the model and motivate the gap. Bilgel et al. fit progression scores to all acquired visits with no genotype term and no visit selection. They also show that ignoring spatial noise correlation drops correctly estimated slopes from 98% to 24% in simulation. That result constrains us, because an information score computed under voxel independence would badly overstate what each scan is worth. LSOR learns longitudinal brain representations from all scan pairs and leaves genetics to future work.
The imaging-genetics papers set the application. Kim et al. correlate SNP and diagnosis effect maps across 116 regions on cross-sectional AV-45 PET and recover 41 of 54 known loci versus 2 for binary GWAS. Li Shen et al. report 4–10x power gains from quantitative traits. EndoPRS uses endophenotypes as priors for risk scores. All three treat the available measurements as fixed. None asks which measurements to acquire.
Research Program
We will first implement the genotype-extended progression model with a sigmoid $f_v$ and rational-quadratic spatial covariance. It will be fit by EM following Bilgel et al., on regional rather than voxel data to keep information computations tractable.
We will then derive $\Delta_i(t)$ using a Laplace approximation to each subject's posterior. Under that approximation, the per-candidate cost is a small matrix update, which mirrors the low-rank information updates in the SLAM front ends.
Third, we will run the greedy scheduler in two modes. In the prospective-simulation mode, any candidate time is allowed. In the retrospective-subsampling mode, the scheduler may only choose among visits that actually occurred. Retrospective subsampling is the only honest test on real data, but it restricts the choice set. ADNI amyloid PET intervals are coarse, so real-data gains will understate prospective gains. We report both modes.
Because scheduling decisions depend only on observed data and genotype, the missingness is ignorable for likelihood inference. Estimates of $\gamma$ therefore stay valid under the adaptive design. We will nonetheless check calibration of the $\gamma$ estimates under the null by permuting genotype.
Baselines
The naive reference is the full cohort: every acquired scan is used, as in Bilgel et al.. The full cohort sets the ceiling. We report what fraction of its genetic information each budgeted method retains.
The default practice is fixed-interval subsampling at matched budget, meaning every $k$-th visit per subject. Its spread-maximizing variant keeps each subject's first and last available scans. For a linear trajectory model, this endpoint design is the textbook optimum for estimating slopes, so it is the strongest simple competitor. All budgeted methods will use the same subjects, the same total scan count, the same fitted model, and the same EM settings. They will differ only in which scans enter the fit.
The strongest existing selection method is the direct SLAM transfer. It schedules a scan when the subject's own progression-score entropy exceeds a running threshold, following Kuo et al.. It also includes the retrospective trigger of Wudenka et al.. Its threshold is tuned on the training split to hit the same budget. This baseline is the ablation that isolates our key ingredient: it uses information theory but targets the individual rather than $\gamma$.
A second ablation computes $\Delta$ under voxel independence ($C = I$). It tests whether the spatial-covariance result of Bilgel et al. carries over to scheduling.
There is no reported result on Lacuna for longitudinal scan scheduling or visit selection in imaging genetics. The closest reported numbers address a different question. One is the cross-sectional 41/54 known-loci recovery of Kim et al.. The other is the 98% correct-slope rate under spatial-covariance modeling in Bilgel et al.. We use them only as context.
Beating the baselines means a smaller standard error for $\gamma$ at equal budget, with confidence intervals from subject-level bootstrap. Equivalently, it means reaching the endpoint design's precision with at least 25% fewer scans. Both must hold in simulation and on the APOE ε4 effect in real data.
Why not just…?
Why not just scan everyone at the first and last visit? For linear progression this design is optimal, and our method should not beat it. If real amyloid trajectories in ADNI and BLSA are effectively linear over the observed window, the idea fails, and we will say so. The adaptive advantage can only come from saturation and from heterogeneity in disease stage. Our simulations therefore sweep curve shape from linear to strongly sigmoid. We also report gains stratified by fitted stage. The predicted signature is no gain at the linear end, with gains concentrated in subjects on the steep part of the curve.
Why not just enrich the cohort by screening, for example scanning only amyloid-positive or APOE ε4 carriers longitudinally? This simple rule captures much of the "skip the flat floor" benefit and is a genuine threat. We include it as a baseline at matched budget. Our scheduler differs in two ways: it keeps genotype groups balanced, and it drops subjects who have reached the ceiling. Screening on genotype also distorts the genotype distribution in ways that hurt estimation of $\gamma$. If screening matches our scheduler, the honest conclusion is that a one-line rule suffices.
Why not just use the subject-level uncertainty trigger from SLAM directly? That rule spends scans wherever an individual is most uncertain. The most uncertain individuals are often those whose rate tells us little about $\gamma$, such as subjects in a genotype group already well covered or subjects whose uncertainty lies mainly in the offset $\beta_i$. The ablation in the Baselines section settles this directly. If subject-level targeting matches cohort-level targeting, the cross-domain transfer is still useful, but our core claim is wrong.
Why not just use all the data and skip selection? On existing cohorts, one should. Selection matters for designing the next cohort or extension. The full-data fit is reported as the ceiling, so the question is how close a given budget can get.
Proposed Experiments
The first experiment is simulation, extending the synthetic design of Bilgel et al.: 100 to 1,000 subjects with up to 7 visits, spatially correlated noise, and a planted genetic effect on $\alpha$. We sweep effect size, curve shape, and budget from 20% to 80% of full data. Metrics are the standard error of $\gamma$, empirical power at $5 \times 10^{-8}$ over replicates, and type-I error under permuted genotype.
The second experiment is retrospective subsampling on BLSA PiB-PET (104 subjects, 300 visits) and longitudinal ADNI AV-45 PET with genotypes. The test effect is APOE ε4 on amyloid accumulation rate, a large effect whose existence is not in doubt. The metric is the fraction of full-data information retained at each budget.
The third experiment is exploratory. We will test the rate-based association across the 54 known AD loci used by Kim et al. and count nominal recoveries at each budget for each scheduler. We will not claim comparability to their cross-sectional 41/54.
The decisive check is the interaction between curve shape and gain in simulation, together with the subject-level ablation on real data. If our scheduler does not beat the endpoint design, the screening rule, and the subject-level trigger on the APOE ε4 rate effect at matched budget, the proposal has failed.
References
- Temporal Trajectory and Progression Score Estimation from Voxelwise Longitudinal Imaging Measures: Application to Amyloid Imaging
- Information-Driven Direct RGB-D Odometry
- Online 3D Reconstruction and 6-DoF Pose Estimation for RGB-D Sensors
- Redesigning SLAM for Arbitrary Multi-Camera Systems
- Towards Robust Monocular Visual Odometry for Flying Robots on Planetary Missions
- LSOR: Longitudinally-Consistent Self-Organized Representation Learning
- Identifying genetic markers enriched by brain imaging endophenotypes in Alzheimer’s disease
- Genetic analysis of quantitative phenotypes in AD and MCI: imaging, cognition and biomarkers
- EndoPRS: Incorporating Endophenotype Information to Improve Polygenic Risk Scores for Clinical Endpoints