Savable but Lost Lives when ICU Is Overloaded: a Model from 733 Patients in Epicenter Wuhan, China

Tingting Dan, Yang Li, Ziwei Zhu, Xijie Chen, Wuxiu Quan, Yu Hu, Guihua Tao, Lei Zhu, Jijin Zhu, Hongmin Cai, Hanchun Wen

TL;DR

During the initial COVID-19 outbreak in Wuhan, ICU beds were a critically scarce resource. This study utilizes machine learning to identify patients who clinically required intensive care but did not receive it. The findings reveal a "Missing-ICU (MI)-mortality" rate of 41%, compared to a 32% mortality rate for those admitted to the ICU—quantifying the lives that could have been saved with adequate resources.

The Challenge: Quantifying the "Missing-ICU" Gap

In the early stages of a pandemic, medical systems often face a "sudden turnover" where patients' conditions deteriorate rapidly, overwhelming Intensive Care Units (ICUs). When demand exceeds supply, clinicians are forced to make difficult triage decisions.

Conceptual framework for identifying 'Missing-ICU' patients and estimating mortality gaps using AI.

While the mortality rate of patients in the ICU is often reported, the mortality of those who should have been there but weren't due to bed shortages is rarely quantified. By defining and estimating "Missing-ICU (MI)-mortality," this research provides a data-driven argument for the efficacy of ICU intervention and offers a tool for better resource planning in future health crises.

Methodology: Prognostic Modeling via Ensemble Learning

The researchers conducted a retrospective study using data from 733 patients. To identify who "needed" an ICU, they developed a prognostic classification model through the following steps:

  1. Feature Selection: Analysis of 909 clinical factors (demographics, lab results, and symptoms) identified the top 10 features with the highest Area Under the Curve (AUC) to build a streamlined, high-performance model.
  2. Handling Class Imbalance: Only 25 of the 733 patients were admitted to the ICU. To address this imbalance, the authors employed ensemble learning with balanced sampling, training 29 different Support Vector Machine (SVM) classifiers on balanced subsets and using a majority voting system for final predictions.
  3. Defining MI-Mortality: The model identifies patients predicted to need ICU care. By analyzing the subset of these patients who died without receiving ICU care, the authors calculated the MI-mortality.

Dataset and Clinical Indicators

The study utilized a dataset of 733 in-patients diagnosed with COVID-19 at the Huangpi Hospital of Traditional Chinese Medicine in Wuhan, China (January–March 2020).

  • Features: 909 total features, including 858 continuous variables (e.g., lab values) and 51 categorical variables (e.g., comorbidities).
  • Key Biomarkers: High levels of hs-cTnI (High sensitivity troponin I) and Myoglobin (Mb)—markers of cardiac stress—were the most significant predictors of clinical deterioration.

Key Results: The 9% Survival Advantage

The prognostic model demonstrated high reliability in identifying critical patients, achieving a predictive accuracy of 0.8288 and an AUC of 0.9119 when utilizing the top 10 clinical features.

Performance analysis: (A) ROC curves for individual clinical features; (B) ROC curves for the integrated prognostic model.

Quantifying the Impact of ICU Care: * MI-mortality (Predicted need, no ICU): 41% * ICU-mortality (Actual ICU care): 32%

This 9% survival gap suggests that ICU intervention significantly reduces the risk of death for critical patients, even in high-stress environments. Furthermore, the model could make these predictions using data collected 1–15 days before actual ICU admission, providing a vital early warning window for hospital administrators.

Discussion and Limitations

While the study provides a compelling quantification of "savable lives," the authors acknowledge several limitations. The data is from a single center in Wuhan with a small positive sample size (25 ICU patients), which may affect the model's generalizability. Future work should incorporate cross-institutional samples to validate these findings and refine the early warning system for broader clinical application.

Summary updated

Abstract

Coronavirus Disease 2019 (COVID-19) causes a sudden turnover to bad at some checkpoints and thus needs the intervention of intensive care unit (ICU). This resulted in urgent and large needs of ICUs posed great risks to the medical system. Estimating the mortality of critical in-patients who were not admitted into the ICU will be valuable to optimize the management and assignment of ICU. Retrospective, 733 in-patients diagnosed with COVID-19 at a local hospital (Wuhan, China), as of March 18, 2020. Demographic, clinical and laboratory results were collected and analyzed using machine learning to build a predictive model. Considering the shortage of ICU beds at the beginning of disease emergence, we defined the mortality for those patients who were predicted to be in needing ICU care yet they did not as Missing-ICU (MI)-mortality. To estimate MI-mortality, a prognostic classification model was built to identify the in-patients who may need ICU care. Its predictive accuracy was 0.8288, with an AUC of 0.9119. On our cohort of 733 patients, 25 in-patients who have been predicted by our model that they should need ICU, yet they did not enter ICU due to lack of shorting ICU wards. Our analysis had shown that the MI-mortality is 41%, yet the mortality of ICU is 32%, implying that enough bed of ICU in treating patients in critical conditions.
BibTeX
@inproceedings{dan2021savable,
  author = {Tingting Dan and Yang Li and Ziwei Zhu and Xijie Chen and Wuxiu Quan and Yu Hu and Guihua Tao and Lei Zhu and Jijin Zhu and Hongmin Cai and Hanchun Wen},
  title = {{Savable but Lost Lives when ICU Is Overloaded: a Model from 733 Patients in Epicenter Wuhan, China}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i6.16612},
}
Figure