Source-linked AI summary
SCOPE-AD: Sequential cost-aware ordinal-belief planning with energy-based models for diagnostic agents
Ziwen Yu, Ivan Koychev, Elizabeth Coulthard, Ting Zhou, Bolin Chen, Dian Hong, Zinuo You, Yujiao Wang, Anthony Mulholland, Qiang Liu
TL;DR
AD diagnosis involves sequential evidence acquisition, but existing approaches do not jointly represent ordered diagnostic uncertainty and decide when to acquire or stop. SCOPE-AD uses ordinal beliefs, an energy-based planner, and a distilled Qwen policy; on ADNI it achieves 77.70% Macro-F1 at $50.46 average cost, while full-modality acquisition adds only 1.89 points at 116.7 times the cost.
Problem
AD diagnosis requires decisions from partially observed evidence, but existing approaches do not jointly represent uncertainty along the CN–MCI–AD continuum and choose acquisitions or stopping points.
Source: §1 · p. 1Method
SCOPE-AD uses ordinal beliefs to plan cost- and feasibility-constrained acquisitions or terminal diagnoses, distilling an offline EBM teacher’s policy into a Qwen agent.
Source: §1 · pp. 1–2Results
77.70% Macro-F1 at an average cost of $50.46; full-modality acquisition adds 1.89 points at 116.7 times SCOPE-AD’s acquisition cost.
Source: §4 · p. 4, §3.4 · p. 4Takeaways & Limitations
The ADNI results support selective acquisition, since tests with limited marginal diagnostic benefit can be avoided at substantially lower cost than full-modality acquisition.
Source: §4 · p. 4, §3.4 · p. 4
Abstract
from arXiv · showhide
Alzheimer's disease (AD) diagnosis requires sequential evidence acquisition under heterogeneous test costs and patient burden. Fixed-modality predictors do not jointly decide which test to acquire or when the available evidence is sufficient for diagnosis. We propose SCOPE-AD (Sequential Cost-Aware Ordinal-Belief Planning with Energy-Based Models for Diagnostic Agents) for cost-aware classification of cognitively normal (CN), mild cognitive impairment (MCI), and AD cases. A mask-aware ordinal model represents uncertainty along the ordered CN--MCI--AD continuum. Retrospective training records provide sampled Bellman targets for an energy-based teacher, whose action distributions are distilled into a Qwen policy. At deployment, the agent selects acquisition or diagnosis actions under availability and budget constraints without access to unacquired values. After each acquisition, the evidence and ordinal belief are updated before the next decision. On ADNI, SCOPE-AD achieves 77.70\% Macro-F1 at an average acquisition cost of \$50.46, exceeding the strongest evaluated baseline by 9.34 percentage points. Full-modality evaluation raises Macro-F1 by only 1.89 points while increasing acquisition cost by 116.7 times. These results support selective acquisition for cost-effective diagnosis.
1. INTRODUCTION
AD diagnosis requires repeated decisions about whether partial evidence is sufficient and which additional modality is worth acquiring. SCOPE-AD combines ordinal-belief planning with sequential acquisition and diagnosis, achieving 77.70% Macro-F1 at an average acquisition cost of $50.46.
1. INTRODUCTION
- AD diagnosis requires deciding whether partially observed evidence suffices or whether another test is worth its cost, invasiveness, and patient burden.Candidate modalities can differ substantially on all three dimensions.Source: §1 · p. 1
- Prior AD approaches often use binary tasks or assume all modalities are available, while acquisition-aware methods have limited ordinal-continuum coverage.The paper identifies conversion prediction, binary classification, restricted modality sequences, and categorical patient states as limitations of existing acquisition-aware approaches.Source: §1 · p. 1
- SCOPE-AD formulates CN–MCI–AD diagnosis as cost-aware sequential decisions over modality acquisitions and terminal diagnoses, using ordinal beliefs.Its contributions combine cross-fitted ordinal beliefs, an offline-trained EBM teacher, and policy distillation into a deployed Qwen agent.Source: §1 · pp. 1–2, Figure 1 · p. 2
- 77.70% Macro-F1 at an average acquisition cost of $50.46 in the sequential ADNI evaluation.The introduction reports this result for SCOPE-AD.Source: §1 · p. 2
2. PROPOSED METHOD
SCOPE-AD builds ordinal beliefs from observed evidence and uses them to guide sequential acquisition or diagnosis. An offline energy-based teacher is distilled into a Qwen policy that acts under availability and budget constraints.
2.1. Overview
- SCOPE-AD constructs ordinal beliefs from observed evidence, trains an offline energy-based teacher on retrospective Bellman targets, and distills its policy into Qwen.At deployment, Qwen acts under a hard valid-action mask without EBM calls or access to unacquired values.Source: §2.1 · p. 2, §2.4 · p. 3
2.2. Ordinal Belief Model
- Masked modality encoders and an ordinal head use ordered thresholds and temperature calibration to produce a CN–MCI–AD posterior from observed features.Patient-level cross-fitting provides out-of-fold training beliefs, while the planner state includes the posterior, availability mask, accumulated cost, and remaining budget.Source: §2.2 · p. 2
2.3. Offline Energy-Based Planning
- The planner compares terminal-diagnosis risk with cost-penalized acquisition followed by a sampled Bellman continuation value, using only valid actions.The hard mask excludes unavailable, already observed, over-budget, or beyond-horizon requests; at zero remaining depth, only terminal diagnoses are permitted.Source: §2.3 · pp. 2–3
2.4. Planner-to-Agent Policy Distillation
- EBM rollouts reveal recorded outcomes only after requests and provide states, valid-action masks, and soft policies for training the Qwen student.At deployment, each acquisition updates evidence, cost, budget, and the observation mask, then recomputes the ordinal belief and planner state.Source: §2.4 · p. 3
3. EXPERIMENTS AND ANALYSIS
The ADNI experiments compare sequential acquisition with language-model and active-acquisition baselines, then assess component and modality controls. SCOPE-AD reports the highest Macro-F1 at low acquisition cost, with cognitive assessments dominating requests.
3.2. Experiment Setup
- Evaluation compares SCOPE-AD with LA-CDM, ReAct, ACTMED [13], SM-DDPO [15], MedGemma-27B, and Qwen2.5-32B.Results average three runs and report class-wise recall, Macro-F1, and average test cost per patient.Source: §3 · p. 3, §3.2 · p. 3
3.1. Dataset
- The retrospective ADNI benchmark retains 2,218 labeled participants (1,552/333/333 train/validation/test); demographics are initially visible, with modality availability known but values hidden until requested.The benchmark uses one baseline record per participant and assigns prespecified acquisition costs to requestable modalities.Source: §3.1 · p. 3
3.3. Comparison with State-of-the-Art Methods
- 77.70% Macro-F1 is highest, 9.34 percentage points above ACTMED; average acquisition cost is $50.46, 63.6% and 98.4% below SM-DDPO and ACTMED.The comparison also reports the highest MCI recall (63.91%) and a tie with ReAct for highest AD recall (75.18%); CN recall is 92.67%, second to MedGemma.Source: Table 2 · p. 3, §3.3 · pp. 3–4
- The policy makes approximately one modality request per case, with cognitive assessments accounting for 99.8% of requests.The reported request pattern usually acquires cognitive evidence and stops without requesting expensive modalities.Source: Figure 2 · p. 4, §3.3 · p. 4
3.4. Ablation and Cost–Utility Analysis
- Removing ordinal belief or the EBM planner lowers Macro-F1 to 67.20% or 61.10%, respectively, versus 77.70% for SCOPE-AD.The table reports component ablations and fixed-policy modality controls.Source: Table 3 · p. 4, §3.4 · p. 4
- Cognitive Only achieves 74.34% Macro-F1 at $50.00; disabling cognitive requests yields 64.60% at $1,839.49, as the policy redirects acquisition to other modalities.These controls reuse the same trained Qwen checkpoint without further fine-tuning.Source: §3.4 · p. 4
- Full-Modality adds only 1.89 percentage points over SCOPE-AD while costing 116.7 times as much, supporting selective acquisition over revealing every available modality.Full-Modality reaches 79.59% Macro-F1 at $5,889.49; the comparison suggests cognitive assessments provide most attainable diagnostic benefit at low cost.Source: §3.4 · p. 4
4. CONCLUSION
SCOPE-AD combines ordinal belief estimation, energy-based planning, and policy distillation for sequential CN–MCI–AD diagnosis from partial multimodal evidence. On the ADNI-based benchmark, it achieves 77.70% Macro-F1 at $50.46, while full-modality acquisition adds only 1.89 points at substantially higher cost.
4. CONCLUSION
- SCOPE-AD integrates ordinal belief estimation, EBM-based planning, and planner-to-agent policy distillation to choose between evidence acquisition and terminal diagnosis.The framework operates on partially observed multimodal evidence for sequential CN–MCI–AD diagnosis.Source: §4 · p. 4
- On the ADNI-based benchmark, SCOPE-AD achieves 77.70% Macro-F1 at an average cost of $50.46; full-modality acquisition adds only 1.89 points at substantially higher cost.The authors conclude that the framework balances diagnostic performance and acquisition cost by avoiding tests with limited marginal diagnostic benefit.Source: §4 · p. 4
- Future work will extend the evaluation to longitudinal patient records and dynamic disease progression.Source: §4 · p. 4