Rethinking Explainable Disease Prediction: Synergizing Accuracy and Reliability via Reflective Cognitive Architecture

Zijian Shao, Haiyang Shen, Mugeng Liu, Gecheng Fu, Yaoqi Guo, Yanfeng Wang, Yun Ma

TL;DR

The Reflective Cognitive Architecture (RCA) framework shatters the traditional trade-off between predictive accuracy and explainability in clinical AI. By combining iterative, error-driven rule optimization with global statistical validation, RCA enables Large Language Models (LLMs) to achieve state-of-the-art disease prediction performance while generating highly reliable, clinician-approved narrative explanations.

What Problem Does This Paper Address?

In clinical decision-making, AI models typically force a difficult choice: adopt highly accurate but opaque "black boxes" (like deep learning or tree-based ensembles), or use interpretable models that lack predictive precision.

While LLMs offer a promising path toward generating natural language explanations, they suffer from a critical flaw when applied directly to tabular clinical data: statistically de-grounded reasoning. They frequently generate explanations that sound medically plausible but are completely unsupported by the underlying data distribution. This paper addresses this gap by proving that high predictive accuracy and high-quality, evidence-based explanations can be synergistic outcomes of a deep, experience-based understanding of data.

Methods

RCA bypasses external code interpreters and APIs, engaging with tabular data directly through a novel dual-process cognitive architecture. The pipeline consists of four main steps:

  1. Data Narrative: Converts structured tabular patient features into unstructured text sentences that an LLM can naturally read and reason about.
  2. Distribution Extraction: Extracts global statistical properties (means, quantiles, frequencies) from the training set to provide a global context.
  3. Iterative Rules Optimization (Bottom-Up): A prediction LLM ($M_{pred}$) forecasts outcomes guided by a dynamic rule base. When it makes an error, a reflection LLM ($M_{ref}$) analyzes the misclassified sample to refine the rule base, turning short-term mistakes into long-term memory.
  4. Distribution-Aware Rules Check (Top-Down): To prevent the model from overfitting to specific errors, a checking LLM ($M_{chk}$) acts as a statistical validator. It audits the updated rules against the global data distribution, filtering out spurious or overly specific logic.
RCA Pipeline Overview

Datasets

The authors evaluated RCA across several distinct clinical tasks:

  • CRT: A proprietary, high-stakes real-world dataset comprising 315 cancer patients evaluated for Catheter-Related Thrombosis, curated in collaboration with Feitian Hospital.
  • Diabetes & Heart Disease: Two public benchmark datasets used to validate performance on well-understood clinical problems.
  • Scalability Benchmarks: To test real-world viability, the framework was scaled to CD (a massive Cardiovascular Disease dataset with 70,000 patient records) and CRT_ex (an expanded version of the CRT dataset with 1,891 patients).

Key Insights and Ideas

  • Synergy Over Trade-Off: Accuracy and explainability are not competing goals. Forcing the model to optimize for correct predictions creates a robust internal representation of clinical patterns, which naturally yields superior, highly articulate explanations.
  • The Danger of General Knowledge: Advanced reasoning models (like DeepSeek-R1) can fail in specialized domains because they rely on general medical text rather than dataset-specific statistical realities. For instance, a baseline might flag a normal biomarker level as "high risk" simply because it knows the biomarker is generally associated with the disease, ignoring the specific quantitative thresholds of the cohort.
  • Robust Anomaly Handling: By anchoring logic in global statistics, RCA naturally identifies extreme data-entry errors or sensor malfunctions, lowering their predictive weight and flagging them for clinical verification rather than letting them corrupt the model.

Key Results

  • State-of-the-Art Performance: RCA significantly outperformed over 22 baselines, including traditional machine learning (CatBoost), neural-symbolic networks, and advanced LLM agents. On the proprietary CRT dataset, RCA achieved an accuracy of up to 0.8730.
  • Superior Noise Resilience: When subjected to data stress (10% missing values or 10% abnormal outliers), RCA demonstrated remarkable stability, whereas traditional baselines suffered severe performance drops.
Figure 2. Results of the main experiment and the robustness experiment on CRT dataset. Hollow dots represent the main experiment, solid dots represent the robustness experiment, and dots of the same shape represent the same approach. The dashed line measures performance variation. RCA demonstrates not only the best results (4.2) but also gets little performance fluctuations(4.3, showing the resilience to data noise.
  • Clinician Approved: In a blind head-to-head study, expert clinicians overwhelmingly preferred RCA's explanations over the strongest baseline (Qwen3-235B) by rates exceeding 70% to 80% across metrics like Cognitive Load, Logical Argumentation, and Evidence-Based Medicine.

Limitations and Future Work

The authors do not include a dedicated limitations section, but a few constraints are apparent:

  • Computational Overhead: Operating a multi-LLM loop (Prediction, Reflection, and Checking LLMs) over multiple training epochs introduces significant API latency and computational costs.
  • Tabular Focus: The current architecture is designed strictly for structured tabular data.

Future work could explore optimizing the framework to run efficiently on smaller, open-source LLMs or extending the architecture to handle multimodal clinical data, such as combining tabular metrics with medical imaging or unstructured clinical notes.

Summary updated

Abstract

In clinical decision-making, predictive models face a persistent trade-off: accurate models are often opaque "black boxes," while interpretable methods frequently lack predictive precision or statistical grounding. In this paper, we challenge this dichotomy, positing that high predictive accuracy and high-quality descriptive explanations are not competing goals but synergistic outcomes of a deep, first-hand understanding of data. We propose the Reflective Cognitive Architecture (RCA), a novel framework designed to enable Large Language Models (LLMs) to learn directly from tabular data through experience and reflection. RCA integrates two core mechanisms: an iterative rules optimization process that refines logical argumentation by learning from prediction errors, and a distribution-aware rules check that grounds this logic in global statistical evidence to ensure robustness. We evaluated RCA against over 20 baselines - ranging from traditional machine learning to advanced reasoning LLMs and agents - across diverse medical datasets, including a proprietary real-world Catheter-Related Thrombosis (CRT) cohort. Crucially, to demonstrate real-world scalability, we extended our evaluation to two large-scale datasets. The results confirm that RCA achieves state-of-the-art predictive performance and superior robustness to data noise while simultaneously generating clear, logical, and evidence-based explanatory statements, maintaining its efficacy even at scale. The code is available at https://github.com/ssssszj/RCA.
BibTeX
@misc{shao2025rethinking,
  author = {Zijian Shao and Haiyang Shen and Mugeng Liu and Gecheng Fu and Yaoqi Guo and Yanfeng Wang and Yun Ma},
  title = {{Rethinking Explainable Disease Prediction: Synergizing Accuracy and Reliability via Reflective Cognitive Architecture}},
  year = {2025},
  eprint = {2509.21266},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.21266},
}
Figure