A systematic review of 35 peer-reviewed studies published across five major databases has found that while artificial intelligence software systems for cardiovascular disease detection report an average accuracy of 93.7 percent, critical shortcomings in validation and implementation impede their translation into clinical practice. Conducted using PRISMA guidelines across Scopus, Web of Science, PubMed, Embase, and IEEE Xplore, the review evaluated quality and risk of bias using an adapted QUADAS-2 framework to assess the state of machine learning tools trained on physiological data.
The research addresses an urgent global healthcare crisis, as cardiovascular disease claims more than 17.9 million lives each year, accounting for nearly one-third of all global deaths. To reduce diagnostic delays, particularly in resource-limited environments, researchers have increasingly investigated scalable, real-time assessment tools. Integrating artificial intelligence with physiological measurements—such as electrocardiograms, photoplethysmograms, and phonocardiograms—presents potential diagnostic pathways, yet the underlying evidence base has remained fragmented and difficult to translate into practical healthcare settings.
Within the 35 identified investigations, electrocardiogram data served as the dominant signal source, utilized by 71.4 percent of the examined systems. Developers predominantly applied deep learning architectures, relying heavily on convolutional neural networks and recurrent neural networks to process signal features. Although the 21 studies reporting diagnostic performance documented accuracies spanning from 70.5 to 99.9 percent, the authors observed that inconsistent reporting of metrics and a pronounced absence of rigorous evaluation severely limit confidence in these high figures.
The review quantified several substantial translational gaps across the existing literature, noting that only 23 percent of the analyzed studies carried out external validation and only 14 percent reached clinical-scale deployment. Furthermore, none of the 35 studies incorporated a principled multi-signal fusion strategy combining electrocardiogram, photoplethysmogram, and phonocardiogram signals, and only a single system, or 2.9 percent, featured a conversational interface designed for patients. To address these deficits, the authors outlined a three-directional framework prioritizing multi-signal models for generalisability, large language model clinical narratives for interpretability, and edge-cloud deployment for accessibility.
Subject of Research: Artificial intelligence software systems for cardiovascular disease detection using physiological signals
Article Title: Artificial intelligence driven software systems for cardiovascular disease detection using physiological signals: a systematic review
Article References: Alam, M. T., Bulbul, A. A.-M., Hossen, M. I., Azad, A. K. M., Khan, A., & Moni, M. A. (2026). Artificial intelligence driven software systems for cardiovascular disease detection using physiological signals: a systematic review. Artificial Intelligence Review. https://doi.org/10.1007/s10462-026-11756-7
Image Credits: AI Generated
DOI: 10.1007/s10462-026-11756-7
Keywords: cardiovascular disease, artificial intelligence, physiological signals, systematic review, electrocardiogram, photoplethysmogram, phonocardiogram, deep learning, convolutional neural networks, recurrent neural networks, external validation, clinical deployment, QUADAS-2, PRISMA, multi-signal fusion, edge-cloud deployment, conversational interface
News Source: Blake Davidson. (November 14, 2026). Most Cardiovascular AI Systems Lack External Validation, Systematic Review Finds. Scienmag.
