Short-Term Arrhythmia Prediction Using AI Based on Daily Data From Implantable Devices: Multicenter Prospective Observational Study.

The AI model achieved a global sensitivity of 66.4% and specificity of 77.4% for predicting short-term arrhythmia changes.

• Global sensitivity was 66.4% (95% CI 64.3%-68.3%)
• Global specificity was 77.4% (95% CI 76.4%-78.4%)
• Performance was calculated from 9711 prediction-observation pairs
• The model predicted whether arrhythmic episodes would increase, decrease, or remain the same in the following 14 days

For patients with baseline arrhythmia, the model demonstrated higher sensitivity but markedly lower specificity compared to global performance.

• Sensitivity for patients with baseline arrhythmia was 76.8% (95% CI 74.6%-78.8%)
• Specificity for patients with baseline arrhythmia was 39.6% (95% CI 35.8%-43.5%)
• This represents substantially lower specificity compared to the global specificity of 77.4%

For patients with no baseline arrhythmia, the model showed lower sensitivity but higher specificity.

• Sensitivity for patients with no baseline arrhythmia was 39% (95% CI 35.1%-43%)
• Specificity for patients with no baseline arrhythmia was 81% (95% CI 80.0%-81.9%)
• This subgroup showed substantially lower sensitivity compared to patients with baseline arrhythmia (39% vs 76.8%)

The model performed best in the subgroup of patients without a history of atrial fibrillation, which comprised the majority of the study population.

• Patients without history of atrial fibrillation comprised 232 of 314 patients (73.9% of the cohort)
• Sensitivity in this subgroup was 69% (95% CI 66.5%-71.5%)
• Specificity in this subgroup was 80% (95% CI 79.3%-81.3%)

The study collected 65,243 data sequences from 314 patients across multiple centers, with the majority used for training the algorithm.

• Total of 314 patients were analyzed in this multicenter prospective observational study
• 65,243 data sequences were collected in total
• 55,532 sequences (85.1%) were used to train the algorithm
• The model used 31-day records as input to predict arrhythmic episode changes in the following 14 days

The authors anticipated that model performance would improve progressively as more data samples become available.

• The authors stated the model's performance 'is expected to improve progressively as more data samples become available'
• Expected improvements include incorporation of demographic data and clinical records
• The current model relied on remote monitoring data extracted from pacemaker devices