PaperlayerView exact revision on arXiv

Source-linked AI summary

Patient, Place, Prior (P$^3$): What Counts as Personalization in Medical World Models?

Xingrui Gu, Hanxue Gu, Yuxiang Zhang, Yang Yang

arXiv:2610.09194v1cs.LGcs.AI

TL;DR

Forecast accuracy alone does not show whether a medical prediction benefits from an individual patient's history or only reflects supplied spatial support and cohort patterns. The paper proposes P3 and Cancer JEPA, then finds that history and patient-matched lesion maps benefit forecasts, while evidence of value beyond a matched population average remains unresolved.

  • Problem

    Aggregate accuracy cannot establish whether a forecast uses a patient's longitudinal history, supplied lesion location, or a cohort-shared trajectory, although these sources have different scientific meanings.

    Source: §1 · p. 1

  • Method

    P3 uses matched input comparisons to audit a frozen Cancer JEPA model, which adds a lesion-constrained neural correction to a patient-conditioned reduced-rank regression baseline.

    Source: §1 · p. 2

  • Results

    Patient contrast favors own over substituted history (1.48 × 10−3) and Place favors matched over substituted support (0.99 × 10−3), while the Prior contrast interval includes zero.

    Source: §5.5 · pp. 9–10

  • Takeaways & Limitations

    P3 separates beneficial use of patient history and supplied spatial support from evidence of predictive value beyond cohort-level dynamics.

    Source: §5.5 · p. 10, §6 · p. 10

    Validation analyses are post-hoc and descriptive rather than independent or confirmatory because the validation cohort was previously used during model development.

    Source: §5.3 · p. 9

Abstract

from arXiv · show

Longitudinal models forecast how a patient's imaging state evolves, but accuracy does not show whether the patient's observed trajectory drives the prediction. A population-average forecast may be useful but cannot establish a patient-specific world-model claim. We introduce Patient, Place, Prior (P$^3$), an audit asking whether a forecast benefits from the patient's longitudinal imaging history (Patient), benefits from patient-matched externally supplied spatial support (Place), and gains predictive value beyond a population-average prediction under matched support and context (Prior). We also propose Cancer JEPA, a one-step model that forecasts frozen representations of future breast dynamic contrast-enhanced MRI examinations during neoadjuvant therapy. It adds a lesion-constrained neural correction, trained with an occlusion-based latent objective, to a patient-conditioned low-complexity reduced-rank regression baseline. This factorization permits a post-hoc P$^3$ audit of the frozen model. In a validation cohort previously used in development, forecast error is lower when the neural correction receives the patient's history rather than another patient's and patient-matched lesion occupancy maps rather than substituted maps. However, the descriptive 95% interval comparing the correction computed from patient history with the population-average neural correction includes zero. P$^3$ thus separates input use from evidence of patient-specific predictive value beyond a population-level pattern.

1 Introduction

Forecast accuracy alone cannot establish that a prediction uses a patient's trajectory, because history, supplied lesion location, and cohort-level patterns can all support an accurate forecast. The paper introduces P3 to separate these sources and applies it to Cancer JEPA, where input substitutions favor patient-matched history and lesion maps but do not establish value beyond a population pattern.

1 Introduction

  • Aggregate forecast error cannot reveal whether predictions rely on patient history, supplied lesion location, or cohort-shared trajectories.Saliency maps show local sensitivity, while replacing history and lesion masks together confounds their contributions.Source: §1 · pp. 1–2
  • P3 compares patient versus donor history, patient-matched versus substituted lesion masks, and patient history versus a population-average prediction under matched conditions.Each comparison retains the patient's future target and evaluation procedure, distinguishing beneficial history use, spatial-support dependence, and value beyond cohort dynamics.Source: §1 · p. 2
  • Cancer JEPA forecasts frozen representations of future breast DCE-MRI examinations with a patient-conditioned reduced-rank regression anchor and an occlusion-trained neural correction.History and core–ring lesion occupancy enter the correction separately, and supplied core occupancy constrains where it can alter the anchor.Source: §1 · p. 2
  • Substituting either the patient's history or lesion mask increases forecast error, but the advantage over a matched population-average correction remains uncertain.The paper therefore distinguishes use of patient information from evidence of predictive value beyond cohort-level dynamics.Source: §1 · p. 2

2 Preliminaries

Longitudinal latent MRI forecasting predicts a future encoded examination from imaging history, supplied spatial support, and observed context; it is an observational forecast, not an identified causal response. The preliminaries distinguish receiving patient history from beneficial use of it and from predictive value beyond a matched population-average prediction.

2.1 Longitudinal latent MRI forecasting

  • Longitudinal forecasting predicts a future latent MRI state from encoded history, supplied spatial support, forecast interval, and observed context rather than reconstructing the complete image.The encoded examination is not treated as a direct biological-state measurement because it may also represent scanner hardware and acquisition protocols.Source: §2.1 · pp. 2–3
  • The forecast is observational and does not identify a causal treatment response without assumptions governing treatment assignment, confounding, and identifiability.Agreement with follow-up imaging may reflect temporal continuity as well as learned change.Source: §2.1 · p. 3

2.2 Patient conditioning and predictive personalization

  • Patient-conditioned means receiving a patient's imaging history; beneficial use requires more accurate prediction from that history than from another patient's under matched conditions.Patient-specific predictive value is the stronger case: the patient's history also improves on a population-average prediction under the same spatial support and observed context.Source: §2.2 · p. 3
  • Patient-matched externally supplied spatial support tests dependence on the provided location, whereas localization requires the model to infer that location from the image.Patient conditioning, beneficial history use, patient-specific predictive value, and dependence on supplied support are distinct properties.Source: §2.2 · p. 3

3 Methodology

P3 audits whether forecasts benefit from a patient’s own imaging history and matched spatial support, and whether that history adds value beyond a population-average prediction under matched conditions. Its interpretation distinguishes these forms of predictive attribution and limits conclusions to the observed data distribution.

3 Methodology

  • P3 varies imaging history and spatial support independently at inference to test Patient, Place, and Prior while holding the target and remaining observed context fixed.The audit applies when history and support can be recombined without changing the prediction target or remaining context.Source: §3 · p. 3

3.1 Fixed-patient evaluation

  • For each patient, model parameters, forecast context, future target, and evaluation procedure remain fixed while only the source of history or support changes.Any recipient-specific model component excluded from substitution is also held fixed and reported.Source: §3.1 · p. 3

3.2 Patient and Place: a factorial substitution design

  • Patient and Place contrasts compare matched inputs with substituted history or support; joint and interaction contrasts describe replacing both sources and how support effects depend on history.Positive contrasts indicate lower error with the patient’s own input; Place reflects dependence on externally supplied support, not autonomous localization.Source: §3.2 · p. 4

3.3 Population-average prediction

  • Prior compares the patient-history prediction with a cohort-level prediction averaged in prediction space, preserving the patient’s support and observed context.The audit should specify the reference cohort, preserved variables, and averaging procedure.Source: §3.3 · p. 5

3.4 Evidence interpretation

  • Patient-specific predictive value requires both positive Patient and Prior contrasts; Place is orthogonal, while causal treatment effects require a separate identification design.These attribution claims concern forecasting within the observed data distribution.Source: §3.4 · p. 5

4 Cancer JEPA: Lesion-Constrained Latent Forecasting

Cancer JEPA forecasts future latent MRI representations by adding a lesion-constrained neural correction to a patient-conditioned reduced-rank regression baseline. Its design separates longitudinal history, supplied lesion occupancy, and observed context for the P3 audit.

4 Cancer JEPA: Lesion-Constrained Latent Forecasting

  • Cancer JEPA adds a lesion-constrained neural correction to a low-complexity reduced-rank regression baseline to forecast future latent MRI representations.The model is structured so the history and supplied lesion occupancy maps enter the correction separately.Source: §4 · p. 5

4.1 Longitudinal latent-state forecasting

  • A frozen volumetric encoder maps serial examinations into a common spatial representation, and the model predicts a future representation rather than reconstructing pixels.The image encoder remains fixed, with no exponential-moving-average target encoder.Source: §4.1 · pp. 5–6

4.2 Low-complexity baseline prediction

  • The patient-conditioned RRR baseline shrinks toward a training-partition mean and uses a low-rank linear map to predict coarse spatial change.Its fitted components are restricted to the corresponding training partition, and the baseline is fixed during neural-correction training.Source: §4.2 · p. 6

4.3 Lesion-constrained neural correction

  • The neural branch predicts a correction from masked imaging history, lesion occupancy, spatial coordinates, and observed context, but the supplied core occupancy gates where it can alter the baseline.Ring occupancy informs the correction as surrounding context; the model relies on externally supplied lesion localization rather than localizing lesions autonomously.Source: §4.3 · p. 7

4.4 Occlusion-based latent prediction objective

  • Training optimizes the neural correction only at occluded history tokens with nonzero supplied core occupancy, normalizing error within patients before averaging across patients.Inference averages predictions over the same prespecified family of occlusion masks used in training.Source: §4.4 · p. 7

4.5 Patient, Place, and Prior in Cancer JEPA

  • The P3 audit substitutes history, lesion occupancy maps, or cohort histories for the neural correction while holding the future target, baseline, context, and evaluation fixed.The population-average correction is averaged before applying the patient's core gate; comparison with the RRR baseline separately tests whether the neural correction improves on that baseline.Source: §4.5 · pp. 7–8

5 Experiments

The experiments evaluate Cancer JEPA on a post-hoc I-SPY2 validation audit. Patient and Place contrasts favor the patient's history and matched spatial support, while the Prior comparison does not establish an advantage over the population-average correction.

5.1 I-SPY2 cohort and forecasting task

  • The patient-disjoint I-SPY2 partitions contain 409 training, 88 validation, and 90 test patients; the task predicts frozen T3 SER latents from T0/T1, not response or a future image.The validation cohort had already been used during development, so its analyses are post-hoc and descriptive; the test partition is not reported.Source: §5.1 · p. 8

5.2 Imaging and model instantiation

  • T0 is the fixed spatial reference, and fractional core and 20-mm ring occupancies are projected from T0 annotations onto the latent grid.The RRR baseline and Cancer JEPA are fitted on training patients and then frozen for validation and P3 analyses.Source: §5.2 · p. 8

5.3 Evaluation and audit protocol

  • The primary endpoint is patient-level latent MAE within the supplied T0 lesion core (N = 87), with the target, RRR baseline, horizon, context, and occlusion pattern fixed across comparisons.Four history–map conditions use 32 deterministic patient permutations and ten fixed occlusion patterns; confidence intervals use 10,000 paired patient-level bootstrap replicates.Source: §5.3 · pp. 8–9

5.4 Incremental value of the neural correction

  • 2.15 × 10−3 is the largest estimated primary-core improvement over fixed RRR, but its 95% interval includes zero; smaller gains in other reported regions have intervals above zero.Table 1 compares validation MAE for Cancer JEPA and its fixed RRR anchor.Source: Table 1 · p. 9, §5.4 · p. 9

5.5 Predictive attribution with Patient–Place–Prior

  • 1.48 × 10−3 and 0.99 × 10−3 MAE contrasts favor own versus substituted history (Patient) and matched versus substituted support (Place); Prior's 95% interval includes zero.The Patient–Place interaction interval also includes zero.Source: Table 2 · p. 9, §5.5 · p. 9
  • Place reflects dependence on externally supplied lesion occupancy maps, not autonomous lesion localization; Patient alone does not establish predictive value beyond cohort-level dynamics.The Prior contrast compares patient history with a population-average correction under matched support and context.Source: §5.5 · pp. 9–10

6 Conclusion

The conclusion presents P3 as a predictive-attribution framework and applies it to Cancer JEPA. In this validation case study, input-specific benefits are observed, but predictive value beyond cohort dynamics remains unresolved.

6 Conclusion

  • P3 separates the value of a patient's imaging history, dependence on matched external spatial support, and improvement over a population-average prediction under matched conditions.Source: §6 · p. 10
  • In the validation case study, Cancer JEPA's neural correction is more accurate with the patient's longitudinal history and lesion occupancy maps than with substituted inputs.Source: §6 · p. 10
  • The correction's advantage over the matched population-average correction and the complete model's primary-core improvement over RRR remain unresolved.The conclusion presents these results as separating patient conditioning, useful patient information, and value beyond cohort dynamics.Source: §6 · p. 10
Loading 2610.09194v1…