Smartwatch AI Reaches 91% Accuracy in Heart Rhythm Study, With Important Trade-Offs

The researchers analysed data from the Pulsewatch clinical trial, using 116,313 recordings lasting 30 seconds each from 72 participants aged 50 or older with a history of ischemic stroke.

An irregular heartbeat can come and go without obvious symptoms, leaving a short clinic recording with little to capture. A smartwatch could offer a longer view of everyday heart activity, provided its software can separate genuine rhythm changes from signals distorted by movement or poor skin contact.

In "Development of an efficient deep learning model for arrhythmia detection from consumer smartwatches," published in Frontiers in Artificial Intelligence, Dan-Cristian George and colleagues at Romania's Babeş-Bolyai University compared three approaches to reading smartwatch recordings. The findings reveal a practical choice between overall accuracy, catching more abnormal beats, and the computing resources each approach needs.

A real-world test built around 72 people

The researchers analysed data from the Pulsewatch clinical trial, using 116,313 recordings lasting 30 seconds each from 72 participants aged 50 or older with a history of ischemic stroke. Participants wore Samsung smartwatches alongside chest-mounted electrocardiogram patches, which supplied the reference heart-rhythm labels used to check the algorithms. The watches recorded photoplethysmography, or PPG, a light-based measurement of blood-volume changes, together with movement readings from an accelerometer.

Each recording belonged to one of three categories: normal sinus rhythm, atrial fibrillation, or premature atrial and ventricular contractions grouped as PAC/PVC. Atrial fibrillation is an irregular rhythm associated with increased stroke risk; premature contractions are beats that arrive earlier than expected.

Normal rhythm accounted for roughly 68% of recordings, atrial fibrillation for 21%, and premature contractions for 11%, making the less common patterns harder for a model to learn.

The team split participants into two separate groups, trained and validated models using one group, and tested them on the other before reversing the process. Every main test involved people unseen during training, offering a more meaningful check of performance on new individuals. Recordings containing more than five seconds of movement-related noise had already been excluded, placing a limit on how closely the results represent unrestricted daily use.

The strongest model depends on what needs catching

The Random Forest model combined 500 decision trees with 29 measurements describing beat timing, pulse shape, movement, and patterns of rhythm variability. Its preparation steps cleaned the signals before extracting those measurements.

The two deep-learning models, CNN–GRU and CNN–BiLSTM, learned patterns directly from minimally processed pulse and movement waveforms. Both combined recognition of local signal features with tracking of patterns across the recording, using an attention mechanism to give informative moments greater weight.

Training gave extra emphasis to underrepresented rhythms through class weighting for Random Forest and repeated sampling of minority examples for the neural networks. Test recordings retained their original distribution.

Random Forest delivered the highest overall accuracy at 91.22%, compared with 88.93% for CNN–GRU and 87.82% for CNN–BiLSTM. It led on macro precision, macro F1 and macro ROC-AUC, measures that give each rhythm category equal weight when assessing prediction reliability, detection balance and separation between classes.

Its macro F1 score of 84.17% slightly exceeded the strongest selected earlier comparison model's 83.64%. Published results from earlier Pulsewatch research lacked some overall measurements, limiting the comparisons the authors could make.

CNN–BiLSTM achieved the highest macro recall at 87.37%, meaning it had the best detection rate averaged equally across the three rhythm categories. CNN–GRU reached 86.85%, and Random Forest reached 83.21%.

The difference becomes clearer for premature contractions: CNN–BiLSTM detected 79.97% of these recordings, compared with 58.64% for Random Forest. Its precision for that category was only 57.07%, meaning roughly 57 of every 100 recordings it labelled as premature contractions were correct; Random Forest reached 74.63% precision.

Random Forest retained the strongest atrial fibrillation results among the compared methods, detecting 95.83% of those recordings with 88.87% precision. CNN–BiLSTM detected 94.89%, showing that its clearest advantage concerned premature beats.

Movement helps the software and sometimes misleads it

Explanations of Random Forest's decisions showed that beat regularity and timing carried substantial information, with peak movement intensity emerging as another influential measurement. Its strong contribution to atrial fibrillation predictions raises a question about whether some decisions reflect movement patterns associated with the recordings, rather than heart rhythm itself.

Removing movement measurements weakened performance, including a 5.08 percentage-point drop in atrial fibrillation detection. Removing signal cleaning reduced premature-contraction detection by 11.14 percentage points, demonstrating how much the preparation process contributed to the results.

Attention and saliency maps helped researchers inspect which parts of recordings influenced the neural networks. Their examination of mistakes highlighted the difficulty of separating genuine irregularity from distorted pulse signals. Normal recordings could trigger abnormal classifications, and motion could obscure features needed to recognise an arrhythmia. These errors have consequences for everyday monitoring, where false detections could generate unnecessary alerts and extra clinical work, and missed abnormalities could delay further assessment.

Random Forest produced closely matched accuracy across the two participant groups. CNN–BiLSTM's accuracy varied from 90.08% to 85.56%, suggesting its performance depended more on which people supplied the test recordings.

Fast processing still needs a place on the wrist

On the same single-core laptop benchmark, Random Forest processed a completed 30-second recording in an average of 33.08 milliseconds, including preparation and prediction. CNN–BiLSTM needed 51.87 milliseconds, comfortably shorter than the recording itself. The faster model carried a larger memory burden. Random Forest occupied 107.64 MB in storage, compared with 6.49 MB for CNN–BiLSTM, and loading it added about 283 MiB of memory use versus 9.9 MiB for the neural model.

Researchers still need to test them on actual smartwatches to check battery use, heat and performance during continuous monitoring. The study included only older stroke survivors, so the results may differ for other people or devices. Different methods of preparing the signals made it difficult to separate the AI's contribution from the rest of the process.

Future work will involve more groups of people, better handling of movement, improved detection of less common rhythms and smaller models. Clinical studies are needed before these systems can become reliable tools for everyday heart screening.