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

TL;DR

This work tackles the long-standing trade-off between predictive accuracy and explainability in clinical disease prediction on tabular data. It introduces Reflective Cognitive Architecture (RCA), an LLM-based framework that learns from data through experience and reflection, coupling Iterative Rules Optimization with Distribution-aware Rules Check to produce both accurate predictions and grounded, narrative explanations. Across CRT, Diabetes, and Heart Disease datasets—including large-scale CRT_ex and a cardiovascular dataset—RCA achieves state-of-the-art accuracy and robust, evidence-based explanations, while remaining resilient to data noise and scalable to real-world settings. The approach offers a path toward trustworthy clinical AI by unifying predictive strength with high-quality, clinically usable reasoning, and code is publicly available for reproducibility.

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.

Figures (4)

• Figure 1: RCA pipeline. $S^{error}$ is a collection of misclassified samples (errors) from $M_{pred}$. $R^k$ is the rule base for the current iteration $k$, and $\hat{R^k}$ denotes the checked and refined rule base from iteration $k$. RCA uses reflective cycles and additional checks to directly analyze data, building a deep understanding that enhances both prediction accuracy and the generation of detailed, grounded explanatory statements.
• 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 (\ref{['sec:MainExperiments']}) but also gets little performance fluctuations(\ref{['sec:RobustExperiments']}, showing the resilience to data noise.
• Figure 3: Comparison of explanations from 'DeepSeek-R1' and RCA for the same patient. RCA demonstrates superior reasoning by integrating quantitative thresholds and providing a balanced, evidence-based argument, a direct result of its deep data understanding. 'DeepSeek-R1' ś explanation, while fluent, is statistically ungrounded and leads to an incorrect prediction.
• Figure 4: Two samples of explanation generated are provided to help better understand the criteria used in explanation experiment. Protective factors in the text are highlighted in blue, while risk factors are highlighted in red.