Generalizable calibrated machine learning models for real-time atrial fibrillation risk prediction in ICU patients

Background: Atrial Fibrillation (AF) is the most common arrhythmia in the intensive care unit (ICU) and is associated with increased morbidity and mortality. Identification of patients at risk for AF is not routinely performed as AF prediction models are almost solely developed for the general population or for particular ICU populations. However, early AF risk identification could help to take targeted preemptive actions and possibly reduce morbidity and mortality. Predictive models need to be validated across hospitals with different standards of care and convey their predictions in a clinically useful manner. Therefore, we designed AF risk models for ICU patients using uncertainty quantification to provide a risk score and evaluated them on multiple ICU datasets. Methods: Three CatBoost models, utilizing feature windows comprising data 1.5-13.5, 6-18, or 12-24 hours before AF occurrence, were built using 2-repeat-10-fold cross-validation on AmsterdamUMCdb, the first freely available European ICU database. Furthermore, AF Patients were matched with no-AF patients for training. Transferability was validated using a direct and a recalibration evaluation on two independent external datasets, MIMIC-IV and GUH. The calibration of the predicted probability, used as an AF risk score, was measured using the Expected Calibration Error (ECE) and the presented Expected Signed Calibration Error (ESCE). Additionally, all models were evaluated across time during the ICU stay. Results: The model performance reached Areas Under the Curve (AUCs) of 0.81 at internal validation. Direct external validation showed partial generalizability with AUCs reaching 0.77. However, recalibration resulted in performances matching or exceeding that of the internal validation. All models furthermore showed calibration capabilities demonstrating adequate risk prediction competence. Conclusion: Ultimately, recalibrating models reduces the challenge of generalization to unseen datasets. Moreover, utilizing the patient-matching methodology together with the assessment of uncertainty calibration can serve as a step toward the development of clinical AF prediction models.