Iterative Refinement Funnel: an architectural pattern for interactive diagnostic systems with asymmetric response processing

DOI: 10.31673/2412-9070.2026.318109

Abstract

The article focuses on the development and formalization of a novel architectural pattern, *Iterative Refinement Funnel* (IRF), intended for the design of interactive diagnostic systems. The pattern addresses the task of efficiently identifying the correct solution among a large set of alternatives under conditions of incomplete and incrementally acquired information. The proposed approach is based on the concept of asymmetric response processing, which accounts for the observed difference in informational value between confirmation (YES) and negation (NO). A confirmed feature triggers a full recomputation of the machine learning model with an updated projection onto a self-organizing map and subsequent re-ranking of hypotheses by a neural network, whereas a negative response allows continuation of the current diagnostic phase without incurring computationally expensive updates.
The paper proposes two interchangeable question selection strategies—*Cluster Heuristic* and *Differential Heuristic*—implemented via the Strategy design pattern, ensuring adaptability to diverse diagnostic scenarios and input data characteristics. A set of robustness mechanisms is introduced, including *Critical NO* for handling negation of pathognomonic features, *Exploration Question* for overcoming local minima effects, and *Soft Differential* for mitigating the impact of noisy or imprecise user responses.
Experimental validation on a medical dataset comprising 844 diseases and 460 unique symptoms demonstrates a Top-1 accuracy of 83.3–84.1%, alongside a 150-fold speedup compared to the classical *Expected Information Gain* approach. Furthermore, it is shown that asymmetric response processing significantly reduces computational costs without degrading diagnostic quality, which is important for real-time systems. The proposed pattern is domain-independent and can be integrated into various intelligent systems, particularly in decision support tasks requiring efficient interactive narrowing of the hypothesis space. Additionally, the formalization of the pattern enables its future automated optimization and extension within hybrid artificial intelligence architectures.

Keywords: architectural pattern, interactive diagnostics, machine learning, Self-Organizing Map, asymmetric processing, Expected Information Gain, system robustness, clinical decision support.