Real-time prediction of atrial fibrillation in intensive care unit: a meta-learning approach

Figures

• 

  Figure 1.. Flow diagram illustrating inclusion and exclusion criteria and final number of patients included in the study.

• 

  Figure 2.. Overview of the meta-learning pipeline for AF prediction. In the data preparation step, patients data filtered by a minimum and a maximum LOS value. Various variables, including dosage of medications and fluids administered during treatment, were retrieved. Additionally, static variables were retrieved along with other dynamic features such as heart rate, blood pressure, and more. In the Data pre-processing, dynamic and static features are merged together, and then frequency of measurements harmonized so every hour each patient has up to 1 measurement per hour. Next, missing data imputed and data scaled and padded for next steps. Labeling involves assigning an AF label to the 24-hour period leading up to the onset of AF, while excluding data from the last 2 hours as a buffer window. This strategy enables the model to predict the onset of AF at least 2 hours in advance. In the Task selection step, patients were sampled from both AF and non-AF patients and grouped in tasks for training. The Training and Testing phase is structured around a meta-learning approach. It consists of an inner-loop and outer-loop learning process. In the inner-loop, the model is trained on the data (Task…

• 

  Figure 3.. Panel (A) The percentage of patients experiencing at least 1 alarm within the previous 48 hours is displayed, with red dots representing AF patients and blue dots representing non-AF patients. For AF patients, the values correspond to the period preceding the AF event, whereas for non-AF patients they refer to their surrogate AF prediction point. Panel (B) Average risk scores for patients with AF are shown by the blue line, while scores for non-AF patients are represented by the grey line. The shaded areas indicate the standard deviation around the mean for each group. The dashed horizontal line marks the decision threshold of 0.5, used to distinguish between alarm and no-alarm measurements. Panel (C) Individual risk‐score trajectories for a sample of AF and non-AF patients. The red line represents the AF patient, while the gray lines correspond to non-AF patients, with risk scores ranging from 0 to 1 over the last x hours prior to the AF event (for the AF patient) or the surrogate AF prediction point (for non-AF patients). The dashed line indicates the decision threshold used to distinguish between alarm and non-alarm classifications.

• 

  Figure 4.. Precision (green) and recall (blue) as a function of time for AF patient at panel (A) and non-AF patients in panel (B). Right y-axis showing number of patients in their pre-AF window (red).

• 

  Figure 5.. The SHAP values depict the impact of the 15 most important features on the model’s risk score prediction. The bars represent the significance of each feature in predicting AF. SHAP values for each feature were calculated by taking the absolute values and then averaging them to determine their contribution. Features marked with an asterisk (*) denote static variables..

• 

  Figure 6.. SHAP values for different AF patients are shown for those with LOS < 24, LOS between 24 and 100, and LOS > 100 in panels (A), (B), and (C), respectively, while panels (D), (E), and (F) represent non-AF patients for the same LOS categories. f(x) is the aggregated SHAP value for time x.