EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Robust Temporal Pattern Mining for Early Detection of Acute Clinical Events in ICU Settings Using LSTM Variants

Ahmed A.F Osman, Sultan Ahmad, Rajit Nair, Ramgopal Kashyap, Mosleh Hmoud Al-Adhaileh, Theyazn H.H Aldhyani, Hikmat A. M. Abdeljaber, Maryam Nasser Almusallam

Seminars in Medical Writing and Education · 01.12.2025 · 4 · 898

DOI: 10.56294/mw2025898 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Seminars in Medical Writing and Education, online erschienen am 01.12.2025. Verfasst von Ahmed A.F Osman und 7 weiteren Autor:innen. Die Arbeit nennt 27 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Introduction

Anomaly detection in the intensive care unit (ICU) is very important for the detection of acute events in the initial stages. However, the sensitivity of the current methods based on time series for anomaly detection in the ICU is low.

Objective

To build an exhaustive anomaly-conscious learning system for the temporal pattern discovery process in the ICU environment to better identify critical incidents within the ICU earlier in time while being highly robust, interpretable, and computationally efficient.

Method

The proposed architecture combines convolutional, attention-based concepts, and recurrence for temporal representation in an integrated approach. It uses attention-based models for the representation of sequences in the ICU data, multi-scale representations for temporal embeddings, and refined representations for anomalous areas in the input ICU data. Such an approach achieves continuous representation of time-series physiology for the detection of anomalous regions.

Results

The proposed framework emerges consistently better than the state-of-the-art solutions such as LSTM networks, GRU-D networks, TCN networks, Informer networks, TFT networks, TimesNet networks, and PatchTST networks on all prominent performance measures. It presents an F1-score of 0.924, an AUPRC of 0.941, an ECE of 1.8%, and an overall accuracy of 93.4%. Even in the context of the domain shift problem, the degradation in performance is minimized with the ΔAUC of 1.5% and ΔECE of 0.4%. However, the proposed framework consumes just 0.9 joules per inference through the processing of 38 sequences per second in 92 milliseconds. Weighted scoring for anomalous cases also helps significantly in their immediate detection.

Conclusion

Through the careful integration of the goals of accuracy, understandability, and computational efficiency, the proposed anomaly-aware model achieves the state-of-the-art performance for the task of early warning in critical care. Its high-level predictive ability, robustness against the problem of domain shift, along with low computational complexity, make it very promising for its real-time implementation within the ICU setting.

Zitieren

APA

Osman, A. A., Ahmad, S., Nair, R., Kashyap, R., Al-Adhaileh, M. H., Aldhyani, T. H., Abdeljaber, H. A. M., & Almusallam, M. N. (2025). Robust Temporal Pattern Mining for Early Detection of Acute Clinical Events in ICU Settings Using LSTM Variants. Seminars in Medical Writing and Education, 4, 898. https://doi.org/10.56294/mw2025898

Vancouver

Osman AA, Ahmad S, Nair R, Kashyap R, Al-Adhaileh MH, Aldhyani TH, et al. Robust Temporal Pattern Mining for Early Detection of Acute Clinical Events in ICU Settings Using LSTM Variants. Seminars in Medical Writing and Education. 2025;4:898. doi:10.56294/mw2025898

BibTeX
@article{Osman2025,
  author = {Osman, Ahmed A.F and Ahmad, Sultan and Nair, Rajit and Kashyap, Ramgopal and Al-Adhaileh, Mosleh Hmoud and Aldhyani, Theyazn H.H and Abdeljaber, Hikmat A. M. and Almusallam, Maryam Nasser},
  title = {Robust Temporal Pattern Mining for Early Detection of Acute Clinical Events in ICU Settings Using LSTM Variants},
  journal = {Seminars in Medical Writing and Education},
  year = {2025},
  volume = {4},
  pages = {898},
  doi = {10.56294/mw2025898}
}
Kostenlos starten →