Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Timely Clinical Diagnosis through Active Test Selection

Silas Ruhrberg Estévez, Nicolás Astorga, Mihaela van der Schaar

TL;DR

This work addresses the challenge of timely, resource-aware clinical diagnosis by introducing ACTMED, a framework that combines Bayesian Experimental Design with large language model surrogates to adaptively select informative diagnostic tests. ACTMED maintains clinician-in-the-loop reasoning, generating intermediate outputs and updating beliefs via posterior probability shifts computed through information gain measured by KL divergence. The approach is evaluated on CKD, Hepatitis, Diabetes, and OSCE-style datasets, demonstrating improved diagnostic accuracy, reduced testing burden, and enhanced interpretability compared with baselines. The results suggest a path toward transparent, adaptive, and generalizable diagnostic systems that operate efficiently under real-world constraints, while acknowledging limitations in data scale, modality, and computational costs. Future work will extend to multimodal, longitudinal data and broader clinical settings to further validate generalizability and impact.

Abstract

There is growing interest in using machine learning (ML) to support clinical diagnosis, but most approaches rely on static, fully observed datasets and fail to reflect the sequential, resource-aware reasoning clinicians use in practice. Diagnosis remains complex and error prone, especially in high-pressure or resource-limited settings, underscoring the need for frameworks that help clinicians make timely and cost-effective decisions. We propose ACTMED (Adaptive Clinical Test selection via Model-based Experimental Design), a diagnostic framework that integrates Bayesian Experimental Design (BED) with large language models (LLMs) to better emulate real-world diagnostic reasoning. At each step, ACTMED selects the test expected to yield the greatest reduction in diagnostic uncertainty for a given patient. LLMs act as flexible simulators, generating plausible patient state distributions and supporting belief updates without requiring structured, task-specific training data. Clinicians can remain in the loop; reviewing test suggestions, interpreting intermediate outputs, and applying clinical judgment throughout. We evaluate ACTMED on real-world datasets and show it can optimize test selection to improve diagnostic accuracy, interpretability, and resource use. This represents a step toward transparent, adaptive, and clinician-aligned diagnostic systems that generalize across settings with reduced reliance on domain-specific data.

Timely Clinical Diagnosis through Active Test Selection

TL;DR

This work addresses the challenge of timely, resource-aware clinical diagnosis by introducing ACTMED, a framework that combines Bayesian Experimental Design with large language model surrogates to adaptively select informative diagnostic tests. ACTMED maintains clinician-in-the-loop reasoning, generating intermediate outputs and updating beliefs via posterior probability shifts computed through information gain measured by KL divergence. The approach is evaluated on CKD, Hepatitis, Diabetes, and OSCE-style datasets, demonstrating improved diagnostic accuracy, reduced testing burden, and enhanced interpretability compared with baselines. The results suggest a path toward transparent, adaptive, and generalizable diagnostic systems that operate efficiently under real-world constraints, while acknowledging limitations in data scale, modality, and computational costs. Future work will extend to multimodal, longitudinal data and broader clinical settings to further validate generalizability and impact.

Abstract

There is growing interest in using machine learning (ML) to support clinical diagnosis, but most approaches rely on static, fully observed datasets and fail to reflect the sequential, resource-aware reasoning clinicians use in practice. Diagnosis remains complex and error prone, especially in high-pressure or resource-limited settings, underscoring the need for frameworks that help clinicians make timely and cost-effective decisions. We propose ACTMED (Adaptive Clinical Test selection via Model-based Experimental Design), a diagnostic framework that integrates Bayesian Experimental Design (BED) with large language models (LLMs) to better emulate real-world diagnostic reasoning. At each step, ACTMED selects the test expected to yield the greatest reduction in diagnostic uncertainty for a given patient. LLMs act as flexible simulators, generating plausible patient state distributions and supporting belief updates without requiring structured, task-specific training data. Clinicians can remain in the loop; reviewing test suggestions, interpreting intermediate outputs, and applying clinical judgment throughout. We evaluate ACTMED on real-world datasets and show it can optimize test selection to improve diagnostic accuracy, interpretability, and resource use. This represents a step toward transparent, adaptive, and clinician-aligned diagnostic systems that generalize across settings with reduced reliance on domain-specific data.
Paper Structure (52 sections, 20 equations, 15 figures, 16 tables, 1 algorithm)

This paper contains 52 sections, 20 equations, 15 figures, 16 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Clinical diagnosis.
  3. Current models for diagnosis.
  4. LLMs for clinical diagnosis.
  5. Contributions.
  6. Problem formalism
  7. Agent-based diagnosis model.
  8. Optimal diagnostic test selection.
  9. ACTMED: Probabilistic reasoning for clinical diagnosis
  10. Surrogate model sampling.
  11. LLM-driven sampling for ACTMED.
  12. Deciding when to stop information acquisition.
  13. Mitigating LLM hallucinations.
  14. Encouraging diverse test result sampling.
  15. Experiments
  16. ...and 37 more sections

Figures (15)

  • Figure 1: Illustration of KL divergence-based diagnostic test selection.(A) The agent begins with an initial prior belief about the likelihood of each diagnosis (blue), which represents current uncertainty, and seeks to approximate the unobserved true posterior (black dashed). (B) For each candidate diagnostic test, the agent uses a surrogate model to simulate possible posterior distributions (gray) conditioned on hypothetical outcomes, and computes the Kullback–Leibler (KL) divergence between each simulated posterior and the prior. (C) The test whose expected KL divergence is greatest (green) is selected, as it is expected to yield the largest information gain. After the test result is observed, the agent updates its belief to a new posterior distribution (orange).
  • Figure 2: Overview of ACTMED. A clinician queries the system (Step 1), prompting the agent to estimate prior disease risk. The agent simulates test outcomes (Step 2), updates beliefs (Step 3), and computes expected KL divergence to select the most informative test (Step 4). The clinician reviews and conducts the test (Steps 5–6), updates the context (Step 7), and the process repeats iteratively.
  • Figure 3: Feature sampling performance. Heatmap showing the best MAE percentage for each test averaged across 5 seeds.
  • Figure 4: Model accuracy across methods and datasets. Bars show mean accuracy over five seeds; error bars indicate standard deviation.
  • Figure 5: Sequential diagnosis refinement. Accuracy improves as additional tests are acquired; error bars show standard deviation across datasets.
  • ...and 10 more figures