Dynamic isolation forest for anomaly detection in post-PCI myocardial infarction patients

Area of Science:

• Cardiology
• Artificial Intelligence in Medicine
• Biomedical Data Science

Background:

• Post-percutaneous coronary intervention (PCI) care for myocardial infarction (MI) patients necessitates continuous laboratory monitoring.
• Current assessment of dynamic laboratory changes relies heavily on clinical expertise, lacking objective methods.
• Identifying deviations from typical recovery trajectories is crucial for patient management.

Purpose of the Study:

• To develop and validate an objective method for assessing post-PCI recovery trajectories in MI patients.
• To compare the performance of Dynamic Isolation Forest (DIF) against other unsupervised methods for anomaly detection.
• To evaluate the prognostic value of DIF in an external dataset.

Main Methods:

• Inclusion of 183 MI patients who underwent PCI.
• Construction of fixed three-time-point windows for analysis.
• Comparison of Dynamic Isolation Forest (DIF) with other unsupervised anomaly detection techniques.
• Validation using blinded expert review and an external prognostic analysis in the MIMIC-IV database.

Main Results:

• DIF demonstrated the highest agreement with expert ratings (Spearman's ρ = 0.585, Kendall's τ = 0.452).
• DIF achieved an Area Under the Curve (AUC) of 0.859, outperforming other methods.
• Higher DIF anomaly scores in the MIMIC-IV cohort correlated with an increased risk of mortality.

Conclusions:

• Dynamic Isolation Forest (DIF) can objectively identify abnormal recovery windows post-PCI, deviating from typical trajectories.
• DIF shows potential for risk stratification in myocardial infarction patients.
• Further clinical validation is required to establish the full clinical utility of DIF.