COTCAgent: Preventive Consultation via Probabilistic Chain-of-Thought Completion
COTCAgent: Preventive Consultation via Probabilistic Chain-of-Thought Completion
Authors: Zihan Deng, Xiaozhen Zhong, Chuanzhi Xu Date: 2026-05-14 Paper ID: openalex:2605.15016
Summary
COTCAgent is a hierarchical framework designed to improve clinical reasoning on longitudinal electronic health records by addressing issues with statistical hallucination and long-range temporal dependency modeling. The system utilizes a Temporal-Statistics Adapter to perform quantitative trend analysis via executable code, alongside a Chain-of-Thought Completion layer that enforces structured evidentiary reasoning. By decoupling these tasks, the framework achieves high-accuracy diagnostic inference with lower computational overhead compared to multi-modal alternatives. Evaluation on internal datasets and HealthBench confirms the model’s superior performance in reliable clinical decision support.
Key Contributions
- Proposes COTCAgent, a hierarchical reasoning framework that decouples statistical computation, feature matching, and language generation for longitudinal EHR analysis.
- Introduces the Temporal-Statistics Adapter (TSA) to convert analytical plans into executable code for standardized clinical trend output.
- Achieves 90.47% Top-1 accuracy on a self-built EHR dataset and 70.41% on HealthBench, outperforming standard LLM-based medical agents.
Open Questions & Future Work
Key Concepts
- probabilistic-chain-of-thought-completion: A hierarchical reasoning layer that uses a weighted symptom-trend-disease knowledge base to guide clinical risk assessment in longitudinal EHR.
- temporal-statistics-adapter: A module that translates clinical reasoning plans into executable code to derive standardized, accurate trend metrics from raw longitudinal data.
Archivist Review
I have approved the core hierarchical reasoning framework’s components (the adapter and the completion layer) as they provide a reusable pattern for integrating statistical code-generation into EHR foundation models. I have also approved the ‘clinical-utility-vs-benchmark-performance’ open question, as it highlights a persistent, significant bottleneck in the field of clinical AI. HealthBench is included as a critical, named benchmark for evaluating such models.
Approved Concepts
- Probabilistic Chain-of-Thought Completion: Addresses the limitation of LLMs in reasoning over long-range, non-uniform temporal EHR data by decoupling statistical computation and reasoning.
- Temporal-Statistics Adapter: Provides a reusable mechanism for LLMs to generate reliable trend analysis by offloading quantitative computation to structured code rather than relying on inherent language generation.
Approved Open Questions
- Clinical Utility vs Benchmarking: The distinction between benchmark performance and real-world clinical reliability is a critical, unresolved obstacle for the adoption of medical foundation models.
Rejected Candidates
- [concept] Probabilistic Chain-of-Thought Completion (
probabilistic-chain-of-thought-completion) - other: This candidate was retained, but I am formalizing the TSA as the second concept to capture the core decoupling mechanism.
Datasets
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.15016
- paper_id
- 2605.15016
- paper_source
- openalex
- domain
- nlp
- tags
- nlp
- architectures
-
- datasets
- healthbench
- concept_slugs
- probabilistic-chain-of-thought-completiontemporal-statistics-adapter
- dataset_slugs
- healthbench
- skill
- TimeSeriesSkill
- created_at
- 2026-05-17T06:08:52Z