• Login
    View Item 
    •   SUSpace Home
    • Faculty of Science and Engineering
    • Department of Computer Science and Engineering
    • 2026-2030
    • View Item
    •   SUSpace Home
    • Faculty of Science and Engineering
    • Department of Computer Science and Engineering
    • 2026-2030
    • View Item
    JavaScript is disabled for your browser. Some features of this site may not work without it.

    “Explainable Risk Prediction from ICU Vital Signs: SHAP-Based Interpretation and Independent Cohort Evaluation of an LSTM Model”

    Thumbnail
    View/Open
    CSE- 260310.pdf (4.970Mb)
    Date
    2026-09-10
    Author
    Tahmina, Kazi
    Metadata
    Show full item record
    Abstract
    Sepsis is a life-threatening clinical condition associated with a dysregulated response to infection and requires timely recognition to reduce the risk of adverse outcomes. This thesis presents an explainable risk prediction framework using an Improved Long Short-Term Memory (LSTM) model to analyze ICU time-series data from the Prediction of Sepsis dataset associated with the PhysioNet Computing in Cardiology Challenge 2019. The model uses a six-hour observation window comprising 34 clinical features to predict sepsis-related risk within a subsequent one-hour prediction horizon. To enhance model interpretability, SHAP GradientExplainer was applied to investigate global and local feature contributions, temporal importance, and feature–time attribution patterns. The final model was evaluated on an independent, previously unseen cohort without retraining, fine-tuning, threshold tuning, or modification of the original research data. Across 195,018 independent test windows, the model achieved an AUROC of 0.7067, AUPRC of 0.0391, sensitivity of 0.6363, specificity of 0.6895, and F1-score of 0.0698. At the patient level, 357 of 440 positive patients were detected, corresponding to a sensitivity of 0.8114. SHAP analysis identified hematocrit (Hct) as the most influential global feature, while the first temporal position (t=1) was identified as the dominant temporal position. The model generated 61,720 alerts, including 2,278 true-positive and 59,442 false-positive alerts, demonstrating a substantial alert burden. Overall, the findings indicate that the Improved LSTM captures clinically relevant temporal risk patterns while providing interpretable information about influential features and time positions; however, its low precision and high false-alert burden emphasize the need for further validation and optimization before consideration for clinical deployment.
    URI
    http://suspace.su.edu.bd/handle/123456789/3071
    Collections
    • 2026-2030 [51]

    Copyright © 2022-2025 Library Home | Sonargaon University
    Contact Us | Send Feedback
     

     

    Browse

    All of SUSpaceCommunities & CollectionsBy Issue DateAuthorsTitlesSubjectsThis CollectionBy Issue DateAuthorsTitlesSubjects

    My Account

    LoginRegister

    Copyright © 2022-2025 Library Home | Sonargaon University
    Contact Us | Send Feedback