Sepsis prediction in ICU patients using deep learning
Loading...
Supplementary material
Other Title
Authors
Rajan, Deepa
Author ORCID Profiles (clickable)
Degree
Master of Applied Technologies (Computing)
Grantor
Unitec, Te Pūkenga – New Zealand Institute of Skills and Technology
Date
2025
Supervisors
Ramirez-Prado, Guillermo
Sharifzadeh, Hamid
Sharifzadeh, Hamid
Type
Masters Thesis
Ngā Upoko Tukutuku (Māori subject headings)
Keyword
sepsis
intensive care units (ICU)
electronic health records
medical informatics
neural networks
deep learning
intensive care units (ICU)
electronic health records
medical informatics
neural networks
deep learning
ANZSRC Field of Research Code (2020)
Citation
Rajan, D. (2025). Sepsis prediction in ICU patients using deep learning (Unpublished document submitted in partial fulfilment of the requirements for the degree of Master of Applied Technologies (Computing)). Unitec, Te Pūkenga - New Zealand Institute of Skills and Technology
https://hdl.handle.net/10652/6945
Abstract
RESEARCH QUESTIONS
• How does the performance of the CNN-LSTM model compare to traditional machine learning models (e.g., Decision Tree, Random Forest, SVM) in terms of accuracy, precision, recall, and F1-score for early sepsis detection?
• What is the minimum subset of clinical features (e.g., heart rate, temperature, lactate levels) required to maintain a high-performing prediction model with reduced computational cost?
• How do SHAP (SHapley Additive exPlanations) values contribute to the interpretability of the CNN-LSTM model and support clinical decision-making in ICU environments?
• What are the limitations of using synthetic oversampling techniques such as SMOTE in the context of highly imbalanced clinical datasets, and how do they affect the model’s generalizability?
ABSTRACT
Sepsis, the most lethal cause of intensive care unit (ICU) admissions, is a life-threatening condition. Early and correct diagnosis is key in improving survival. The conventional tools, like SOFA and SIRS scores, are static and retrospective, thereby resulting into a delayed identifying of the condition, and also to false positive detection. This is a pilot study aimed to solve this set of challenging questions by a novel approach of deep learning based early sepsis prediction using multi-variate time-series lab test, vital sign, and electronic health record (EHR) data.
The proposed approach leverages Convolutional Neural Networks (CNNs) for automatic data feature extraction, combined with Long Short-Term Memory (LSTM) networks to identify temporal progressions in patients’ data. Database preprocessing by normalisation, imputation, and the Synthetic Minority Over-sampling Technique (SMOTE) is utilized to correctly handle common issues, e.g., noise, missing values, and class imbalance. Interpretability is strengthened by explanations generated via Shapley Additive explanations (SHAP), thus promoting clinical use and enabling clear interpretation of model predictions.
Our CNN-LSTM model also performed well in 22,824 test samples, with the training data consisting of almost 58,000 hourly CCU data: accuracy 95.63%, precision 97.6%, recall 93.5%, F1-score 95.5% for sepsis classification. The high NPV for excluding non-sepsis was similarly established. The attention mechanism allowed the model to be significantly more interpretable by focusing on key prediction characteristics.
Together, this work shows that deep learning can predict sepsis with high accuracy well in advance of the manifestation of clinical symptoms and may be used as a practical way to trigger early intervention and improve outcomes in the ICU.
Publisher
Permanent link
Link to ePress publication
DOI
Copyright holder
Author
Copyright notice
All rights reserved