Attention-Guided Autoencoder for Retrospective Physiological Anomaly Detection in ICU Vital Signs Using MIMIC-IV Demo Data
Abstract
This study evaluates threshold-defined physiological anomaly screening in retrospective vital-sign data from the MIMIC-IV Clinical Database Demo v2.2. Six vital signs were temporally aligned; Fahrenheit temperature values were converted to Celsius, physiologically implausible measurements were converted to missing values before within-stay forward/backward filling, and remaining incomplete rows were removed. After cleaning, the ICU stays were separated at stay level into 82 training, 28 validation, and 28 test stays. These partitions produced 9,210, 2,784, and 5,183 windows, respectively. Only normal windows from the training partition were used to fit the neural models, while labeled validation windows were used to choose the operating thresholds. Across five seeds, AGAE achieved accuracy 0.7838, precision 0.3417, recall 0.4954, specificity 0.8335, F1 0.4025, ROC-AUC 0.7709, and PR-AUC 0.3519. The threshold-independent results for AGAE and LSTM-AE were almost the same (ROC-AUC 0.7709 versus 0.7722; PR-AUC 0.3519 versus 0.3530). At the selected operating points, AGAE produced higher accuracy, precision, and specificity and fewer false positives; LSTM-AE produced higher recall and fewer false negatives. Taken together, these results do not show a consistent advantage for AGAE over LSTM-AE, and AGAE recall remained moderate. Prospective streaming and clinical validation are required before any real-time use.
Related
- A cascaded U-nets model based on hybrid attention for video anomaly detection
- Hyperspectral Anomaly Detection via Group Sparse Low-Rank Tensor Factorization With Automatic Anomaly Grouping
- Real-time and adaptive anomaly detection algorithm for cyclostationary models
- Moment-Based Anomaly Detection in High-Dimensional Mixed Data
- Statistical Feature Augmentation for Anomaly Detection in Dynamic Graphs
BibTeX
@article{suprayogi2026attentionguided,
title = {{Attention-Guided Autoencoder for Retrospective Physiological Anomaly Detection in ICU Vital Signs Using MIMIC-IV Demo Data}},
author = {Suprayogi and Mochamad Agung Wibowo and De Rosal Ignatius Moses Setiadi},
journal = {Journal of Information Systems and Informatics},
year = {2026},
month = oct,
doi = {10.63158/journalisi.v8i5.1808},
url = {https://doi.org/10.63158/journalisi.v8i5.1808},
note = {CC BY},
}
RIS
TY - JOUR
TI - Attention-Guided Autoencoder for Retrospective Physiological Anomaly Detection in ICU Vital Signs Using MIMIC-IV Demo Data
AU - Suprayogi
AU - Mochamad Agung Wibowo
AU - De Rosal Ignatius Moses Setiadi
PY - 2026
DA - 2026/10/24
JO - Journal of Information Systems and Informatics
DO - 10.63158/journalisi.v8i5.1808
UR - https://doi.org/10.63158/journalisi.v8i5.1808
AB - This study evaluates threshold-defined physiological anomaly screening in retrospective vital-sign data from the MIMIC-IV Clinical Database Demo v2.2. Six vital signs were temporally aligned; Fahrenheit temperature values were converted to Celsius, physiologically implausible measurements were converted to missing values before within-stay forward/backward filling, and remaining incomplete rows were removed. After cleaning, the ICU stays were separated at stay level into 82 training, 28 validation, and 28 test stays. These partitions produced 9,210, 2,784, and 5,183 windows, respectively. Only normal windows from the training partition were used to fit the neural models, while labeled validation windows were used to choose the operating thresholds. Across five seeds, AGAE achieved accuracy 0.7838, precision 0.3417, recall 0.4954, specificity 0.8335, F1 0.4025, ROC-AUC 0.7709, and PR-AUC 0.3519. The threshold-independent results for AGAE and LSTM-AE were almost the same (ROC-AUC 0.7709 versus 0.7722; PR-AUC 0.3519 versus 0.3530). At the selected operating points, AGAE produced higher accuracy, precision, and specificity and fewer false positives; LSTM-AE produced higher recall and fewer false negatives. Taken together, these results do not show a consistent advantage for AGAE over LSTM-AE, and AGAE recall remained moderate. Prospective streaming and clinical validation are required before any real-time use.
SN - 2656-5935
SN - 2656-4882
ER -