Time-Series ICU Patient Deterioration Predictor

Fuente: Zenodo
Saved in:
Bibliographic Details
Main Author: Yip, Simon
Format: Recurso digital
Language:English
Published: Zenodo 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866902170597588992
author Yip, Simon
author_facet Yip, Simon
contents <p>Dual-architecture machine learning system for early detection of clinical deterioration in intensive care unit (ICU) patients. The system compares gradient-boosted decision trees (LightGBM) with temporal convolutional networks (TCN) to model complementary aspects of physiological risk using routinely collected clinical observations. Three NEWS2-derived deterioration outcomes are considered: maximum risk level attained during the ICU stay (<code>max_risk</code>), median sustained risk level across the stay (<code>median_risk</code>), and the proportion of time spent in a high-risk state (<code>pct_time_high</code>).</p> <p>Models are trained and evaluated using the PhysioNet MIMIC-IV Clinical Demo v2.2 dataset via two distinct feature-engineering pipelines. The TCN operates on high-resolution timestamp-level temporal features (96-hour windows, 171 features) to capture short-term physiological instability, while the LightGBM model uses patient-level aggregated tabular features (40 features) to characterise longer-term exposure to risk. Comparative evaluation indicates complementary performance profiles: LightGBM exhibits superior calibration and regression fidelity for sustained risk estimation, while TCNs show stronger sensitivity and discrimination for acute deterioration events. Performance is assessed using ROC-AUC, Brier score, and R², alongside interpretability analyses based on SHAP values and saliency methods.</p> <p>The end-to-end pipeline includes clinically validated NEWS2 preprocessing (including CO₂ retainer logic, Glasgow Coma Scale mapping, and supplemental oxygen protocols), comprehensive feature engineering, model training with hyperparameter optimisation, robust metric evaluation, and a command-line inference interface supporting batch prediction and per-patient lookup. Overall, the system demonstrates physiologically plausible predictive behaviour, clinically meaningful interpretability, and a reproducible workflow suitable for extension to full clinical datasets or downstream deployment contexts.</p> <p>This work is intended for research and educational use.</p> <table> <tbody> <tr> <td><strong>Target Outcome</strong></td> <td><strong>Best-Performing Model</strong></td> <td><strong>Key Metric(s) </strong></td> <td><strong>Notes</strong></td> </tr> <tr> <td><code>max_risk </code></td> <td>TCN</td> <td>ROC-AUC = 0.923; </td> <td>Strong acute deterioration detection</td> </tr> <tr> <td><code>median_risk</code></td> <td>LightGBM</td> <td>ROC-AUC = 0.872; Brier Score = 0.065</td> <td>Superior sustained risk calibration</td> </tr> <tr> <td><code>pct_time_high</code> </td> <td>LightGBM</td> <td>R² = 0.793; RMSE = 0.038</td> <td>Higher fidelity estimation of high-risk exposure </td> </tr> </tbody> </table> <p> </p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_18487174
institution Zenodo
language eng
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Time-Series ICU Patient Deterioration Predictor
Yip, Simon
Early Warning Score
ICU
Critical Care
Patient Deterioration
Machine Learning
LightGBM
Temporal Convolutional Network
Time Series Analysis
Clinical Prediction
NEWS2 Score
MIMIC-IV
Neural Network
Artificial Intelligence
Health Informatics
<p>Dual-architecture machine learning system for early detection of clinical deterioration in intensive care unit (ICU) patients. The system compares gradient-boosted decision trees (LightGBM) with temporal convolutional networks (TCN) to model complementary aspects of physiological risk using routinely collected clinical observations. Three NEWS2-derived deterioration outcomes are considered: maximum risk level attained during the ICU stay (<code>max_risk</code>), median sustained risk level across the stay (<code>median_risk</code>), and the proportion of time spent in a high-risk state (<code>pct_time_high</code>).</p> <p>Models are trained and evaluated using the PhysioNet MIMIC-IV Clinical Demo v2.2 dataset via two distinct feature-engineering pipelines. The TCN operates on high-resolution timestamp-level temporal features (96-hour windows, 171 features) to capture short-term physiological instability, while the LightGBM model uses patient-level aggregated tabular features (40 features) to characterise longer-term exposure to risk. Comparative evaluation indicates complementary performance profiles: LightGBM exhibits superior calibration and regression fidelity for sustained risk estimation, while TCNs show stronger sensitivity and discrimination for acute deterioration events. Performance is assessed using ROC-AUC, Brier score, and R², alongside interpretability analyses based on SHAP values and saliency methods.</p> <p>The end-to-end pipeline includes clinically validated NEWS2 preprocessing (including CO₂ retainer logic, Glasgow Coma Scale mapping, and supplemental oxygen protocols), comprehensive feature engineering, model training with hyperparameter optimisation, robust metric evaluation, and a command-line inference interface supporting batch prediction and per-patient lookup. Overall, the system demonstrates physiologically plausible predictive behaviour, clinically meaningful interpretability, and a reproducible workflow suitable for extension to full clinical datasets or downstream deployment contexts.</p> <p>This work is intended for research and educational use.</p> <table> <tbody> <tr> <td><strong>Target Outcome</strong></td> <td><strong>Best-Performing Model</strong></td> <td><strong>Key Metric(s) </strong></td> <td><strong>Notes</strong></td> </tr> <tr> <td><code>max_risk </code></td> <td>TCN</td> <td>ROC-AUC = 0.923; </td> <td>Strong acute deterioration detection</td> </tr> <tr> <td><code>median_risk</code></td> <td>LightGBM</td> <td>ROC-AUC = 0.872; Brier Score = 0.065</td> <td>Superior sustained risk calibration</td> </tr> <tr> <td><code>pct_time_high</code> </td> <td>LightGBM</td> <td>R² = 0.793; RMSE = 0.038</td> <td>Higher fidelity estimation of high-risk exposure </td> </tr> </tbody> </table> <p> </p>
title Time-Series ICU Patient Deterioration Predictor
topic Early Warning Score
ICU
Critical Care
Patient Deterioration
Machine Learning
LightGBM
Temporal Convolutional Network
Time Series Analysis
Clinical Prediction
NEWS2 Score
MIMIC-IV
Neural Network
Artificial Intelligence
Health Informatics
url https://doi.org/10.5281/zenodo.18487174