Abstract:
Massive volumes of multivariate time-series data are produced by the ongoing patient monitoring
in intensive care units (ICUs), which records vital physiological indicators such as heart
rate, blood pressure, oxygen saturation and respiration rate. Finding early indicators of clinical
deterioration and facilitating timely medical intervention depends on the timely and accurate
detection of aberrant patterns in this data. However, threshold-based alerts and conventional
rule-based systems frequently have high false alarm rates and little flexibility to accommodate
the heterogeneity of individual patients. To tackle the crucial problem of automated anomaly
identification in ICU patient monitoring data, this study suggests a deep learning-based system
that provides explainability for clinical transparency in addition to detecting anomalies
in multivariate time-series signals. Using the publicly available eICU Collaborative Research
Database, we introduce an end-to-end anomaly detection pipeline that combines an Encoder-
Decoder-Discriminator (EDD) model, structured as a Variational Autoencoder (VAE). This architecture
learns latent representations of normal physiological behaviour to effectively identify
deviations indicative of abnormal conditions. Both periodic and aperiodic vital sign recordings
were preprocessed using a sliding window approach to extract temporal relationships. Classification
criteria such as confusion matrices, F1-score, and classification reports were used to
thoroughly assess performance. The model achieved a precision of 0.81 and a 0.5 recall in
anomaly identification after being trained on data from 150 patients and validated on 50 unseen
patients. We used SHAP (SHapley Additive exPlanations) to measure feature importance
to meet the crucial need for model interpretability in clinical decision-making. This method
improves physicians’ confidence and comprehension of AI-driven warnings by highlighting the
physiological factors that have the biggest impact on the model’s anomaly predictions. The
results show that our method not only increases the accuracy of anomaly identification in intensive
care units but also opens the door for the incorporation of reliable machine learning tools
into clinical processes in real time. Future research will examine the usage of Temporal SHAP
for deployment in real-time intensive care unit settings, integration with multimodal inputs such
as lab findings and notes, and time-aware interpretability.
