Moderate evidence
More
Review of 35 AI heart-screening studies finds only 23% tested on outside data
Study title: Artificial intelligence driven software systems for cardiovascular disease detection using physiological signals: a systematic review
- Evidence type
- Systematic review or meta-analysis
- Sample
- n = 35
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-2955cite or check it
Study title: Artificial intelligence driven software systems for cardiovascular disease detection using physiological signals: a systematic review
A systematic review of 35 peer-reviewed studies on AI software for detecting cardiovascular disease from heart signals found strong reported accuracy but thin real-world validation. Reported accuracies ranged from 70.5% to 99.9%, with an unweighted mean of 93.7% across the 21 studies that reported it. Only 23% of studies tested their systems on external data, and only 14% reached clinical-scale deployment.
Read the study abstract
Abstract Cardiovascular disease exacts an extraordinary global toll, with annual mortality exceeding 17.9 million lives and representing nearly one-third of all deaths worldwide. Reducing diagnostic delays, particularly in resource-limited settings, requires innovative, scalable, and cost-effective solutions. Artificial intelligence integrated with physiological signals, such as electrocardiogram (ECG), photoplethysmogram (PPG) and phonocardiogram (PCG), offers promising pathways for scalable, real-time cardiova…

The problem
A systematic review of 35 studies of AI heart-disease detection from ECG, PPG and PCG signals found only 23% used external validation and only 14% reached clinical-scale deployment.
In brief
The gaps matter because cardiovascular disease kills more than 17.9 million people a year, nearly a third of all deaths worldwide, and faster diagnosis in resource-limited settings is a stated aim of these tools. The authors say inconsistent metric reporting and limited external validation weaken confidence that benchmark results will carry over to clinics.
Most systems relied on a single signal: ECG was used in 71.4% of systems, mainly with deep learning models. None combined ECG, PPG and PCG through principled multi-signal fusion, and just one of the 35 systems offered a patient-facing conversational interface. The authors propose a framework built on multi-signal models, LLM-generated clinical narratives and edge-cloud deployment, but it is a proposal and has not been tested.
Main points
- The review searched five databases (Scopus, Web of Science, PubMed, Embase and IEEE Xplore) under PRISMA guidance and identified 35 eligible peer-reviewed studies.
- Study quality and risk of bias were appraised with a framework adapted from QUADAS-2.
- ECG was the input signal in 71.4% of systems, mostly analysed with convolutional and recurrent neural networks.
- Reported accuracy ranged from 70.5% to 99.9% (21 studies reporting, unweighted mean 93.7%), but metric reporting was inconsistent across studies.
- Only 23% of studies carried out external validation and only 14% achieved clinical-scale deployment.
- No study combined ECG, PPG and PCG through principled multi-signal fusion, and only 1 of 35 systems (2.9%) had a patient-facing conversational interface.
- The authors propose three research directions: multi-signal models for generalisability, LLM-driven clinical narratives for interpretability, and edge-cloud deployment for accessibility.
The full story
The review covered software systems that apply machine learning to ECG, photoplethysmogram and phonocardiogram signals, including apps, wearables and cloud platforms. Deep learning dominated, with convolutional and recurrent neural networks the most common architectures. Accuracy alone was often the headline metric, and the authors say the lack of consistent reporting made cross-study comparison difficult.
Validation was the weakest area. Many systems were evaluated on legacy public benchmark datasets such as MIT-BIH, which the authors say lack demographic diversity and real-world signal noise, and hospital-acquired data were rarely used. Large-scale deployments were rare, with examples such as the Huawei Heart Study of 187,912 participants standing out.
The authors argue single-signal ECG systems miss the mechanical and blood-flow dimensions of heart function, and that explainability tools such as saliency maps do not match how clinicians reason. They also report that regulatory alignment and real-world testing for latency, battery use and network variability were largely absent, so prototypes cannot move toward clinical certification without redesign.
The proposed framework would fuse ECG, PPG and PCG signals, use large language models to turn probability outputs into clinical narratives, and run lightweight screening on edge devices with a patient chatbot. These are recommendations for future research, not tested systems.
Sources
1 on file- Artificial Intelligence Reviewdoi:10.1007/s10462-026-11756-7
Talk it over
Reported accuracies in the 35 studies ran from 70.5% to 99.9%, yet only 23% tested on external data. How much weight would you give a high accuracy figure from a system tested only on benchmarks like MIT-BIH?
Review of 35 AI heart-screening studies finds only 23% tested on outside datahttps://truvace.com/item/artificial-intelligence-driven-software-systems-for-cardiovascular-disease-detec
Reported accuracies in the 35 studies ran from 70.5% to 99.9%, yet only 23% tested on external data. How much weight would you give a high accuracy figure from a system tested only on benchmarks like MIT-BIH?
Answer in the threadNearly all the systems used one signal, with ECG in 71.4%, and none fused ECG, PPG and PCG. Who might be missed when a screening tool listens to only one part of how the heart works?
Answer in the threadIf you were a hospital team choosing among these AI heart-screening systems, which kind of proof would you ask for first: external validation, clinical-scale deployment or consistent metric reporting?
Answer in the thread
If you have more time
Look up one of the legacy datasets the authors name, such as MIT-BIH. Who was included in it, and how might that shape what a tool trained on it learns?
Answer in the threadWhat result from a hospital using hospital-acquired data would make you trust these tools more, and how would you want it measured?
Answer in the thread
ace
What readers say