A Modality-Adaptive Software Architecture for Explainable Clinical Decision Support

Danila Lissovoy, Assylbek Khalyk, Akylbek Oralov, Rakhat Mustafin, Alisher Kenzhegaliyev, Hashim Ali

2026

Full text Open access: Hybrid Metadata (CSL-JSON)
DOI 10.66395/globeis.25
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Artificial intelligence (AI)-enabled clinical decision support systems (CDSSs) increasingly combine structured clinical data and medical images, but their explanations differ in cost, format, storage, and failure behavior. This study presents a modality-adaptive software architecture for a browser-based CDSS supporting 10-year coronary heart disease (CHD) prediction, 30-day hospital readmission prediction, and chest-radiograph classification with 14 independent output scores. The architecture treats explanation selection as a request-time service policy: Shapley additive explanations (SHAP) accompany tabular predictions; local interpretable model-agnostic explanations (LIME) are requested separately; gradient-weighted class activation mapping (Grad-CAM) and its generalized variant (Grad-CAM++) are computed only for threshold-positive classes within the five highest-ranked scores; and large language model (LLM) narratives are attempted when configured. React, FastAPI, PostgreSQL, and a seven-day Redis cache implement the design. Seven synthetic fixture families satisfied the specified selection conditions. In a backend experiment involving 7,818 radiographs, heatmap generation averaged 140.2 ms per radiograph, or 78.2% of the 179.3-ms mean total. A separate in-process service-layer test completed all 3,600 requests successfully using deterministic prediction and explanation substitutes and SQLite persistence. The architecture specifies bounded explanation work and separate retention paths. Full-stack performance, clinical-model calibration, explanation quality, and clinician benefit remain subjects for future evaluation.