•  
  •  
 

Abstract

Anomaly detection plays a crucial role in healthcare, supporting early disease identification, process improvement, and the reliability of medical equipment. Traditional machine learning (ML) and deep learning (DL) models often fail to generalize across heterogeneous medical datasets due to variations in feature distributions, class imbalance, and high dimensionality. To address this, we extend our earlier ASAD framework and propose ASADH (Auto-Selective Anomaly Detection in Healthcare), designed specifically for medical anomaly detection. ASADH applies meta-learning over a unified set of cross-modal meta-features spanning tabular, time-series, and medical image data to characterize each dataset and automatically recommend the most suitable anomaly detection model. This eliminates the need for manual trial-and-error testing, which in conventional practice requires significant human expertise, time, and repeated evaluations for each new dataset. Instead, ASADH leverages prior knowledge to deliver scalable and adaptive model recommendations. To train the meta-learner, 23 ML/DL models were evaluated across 140 healthcare datasets with seven performance metrics. Experiments on nine external datasets demonstrated that ASADH consistently identified meaningful dataset similarities and successfully transferred high-performing models, achieving a 90.9% success rate. These results confirm the effectiveness of ASADH as a robust and practical framework for automating anomaly detection in healthcare.

Keywords

Anomaly detection, Deep learning, Healthcare, Machine learning, Meta-features, Meta-learning

Article Type

Article

First Page

9

Last Page

23

Publication Date

12-31-2025

Download

Share

COinS