gms | German Medical Science

Introduction: Understanding the phenomenon of unplanned readmission to the intensive care unit (ICU) is crucial, as it has a profound impact on patient outcomes, healthcare resources, and costs [1]. A number of risk factors have been identified, encompassing patient demographics, clinical parameters and premature discharge, which increase the likelihood of this issue occurring [2]. While current discharge protocols rely predominantly on subjective evaluations and occasionally scoring systems, there exists a deficiency in a dedicated scoring mechanism specifically designed for predicting readmission risks. The application of machine learning (ML) and deep learning (DL) methodologies represents a promising approach to enhance the predictive capabilities across the spectrum of ICU patient care.

Methods: In this retrospective study, we analyzed unplanned readmission within 30, 7 and 3 days among ICU patients. Data utilized for this analysis was sourced from electronic health records (EHR) in the Medical Information Mart for Intensive Care IV (MIMIC-IV) [3] dataset, collected within 24 hours prior to the initial ICU discharge. Patient records included demographic data, vital signs, and laboratory measurements. We applied machine learning (ML) techniques, notably XGBoost, and deep learning (DL) models, specifically Long Short-Term Memory (LSTM) networks optimized for sequential data processing, to analyze the patient records and predict the readmission. Addressing the challenge of dataset imbalance, we computed an optimal decision threshold for the Precision-Recall (PR) curve. Subsequently, these methodologies were transferred to an ICU patient dataset obtained from University Hospital Leipzig (UKL) to assess model generalizability and performance in a practical setting.

Results: The findings of this study indicate that XGBoost demonstrates robust predictive performance, achieving an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.74 and an Area Under the Precision-Recall Curve (AUCPR) of 0.76 in forecasting readmission within 30 days post-discharge. However, initial analyses on the UKL dataset revealed challenges concerning data quality and feature availability. Feature importance analysis on the MIMIC-IV dataset revealed that additional patient information, such as total number of ICU stays and discharge location, had a significant impact on the classification process.

Discussion: While XGBoost outperformed alternative models, demonstrating its effectiveness in predicting ICU readmissions, LSTM models showed relatively poorer performance, highlighting the complexities associated with integrating sequential data due to missing data. The application of predictive models to the UKL dataset, which is not explicitly tailored for research, presents a variety of hurdles which affect direct comparisons between datasets. In addition, compliance with German data protection regulations imposes additional constraints on data accessibility, which requires cautious consideration in the use of predictive models in clinical settings.

Conclusion: Despite these challenges, ML techniques show promise in the context of ICU patient prediction for readmission risk assessment, even in the absence of time-resolved data, potentially supporting clinical decision making regarding discharge protocols. However, successful translation of predictive models into clinical practice requires careful evaluation, taking into account the complex interplay between data quality, feature availability and regulatory compliance, highlighting the inherent complexity in generalizing findings from MIMIC-IV to the UKL dataset.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.