The unpredictable deterioration of COVID-19 patients into critical illness poses a significant challenge. Early identification of these patients is crucial for timely intervention and improved outcomes. This research aimed to develop a deep learning model capable of predicting the risk of critical illness in COVID-19 patients based on their clinical characteristics at admission. The model's accuracy and generalizability were evaluated using several independent cohorts of patients with varying geographical locations and access to healthcare resources. The importance of this study lies in its potential to improve patient care by allowing for early identification of high-risk individuals, thereby facilitating early intervention and optimizing the allocation of limited healthcare resources. This is particularly relevant in pandemic scenarios like the COVID-19 outbreak, where rapid and efficient triage is critical for managing large influxes of patients with varying severities of illness.

Traditional survival analysis methods, such as the Cox proportional hazards model (CPHM), have been used to predict clinical outcomes. However, these methods often make simplifying assumptions about the relationship between predictor variables and the risk of an event. Deep learning methods, on the other hand, can model complex, non-linear relationships. Previous studies have shown the success of deep learning in various medical applications, including image analysis and prediction of clinical outcomes. This research aims to leverage the advantages of deep learning to improve upon traditional survival analysis techniques, creating a model that is more accurate in predicting the risk of critical illness in COVID-19 patients. The study considers other existing models such as CURB-65, highlighting the proposed model's superior predictive abilities.

A retrospective cohort study was conducted using data from 1590 hospitalized COVID-19 patients from 575 medical centers across China. Data included 74 baseline clinical features. A machine learning algorithm (LASSO) was used for feature selection, identifying ten statistically significant predictors (age, dyspnea, cancer history, COPD, number of comorbidities, neutrophil/lymphocyte ratio, lactate dehydrogenase, direct bilirubin, creatine kinase, and X-ray abnormality). A three-layer feedforward neural network was developed to create a deep learning survival Cox model. The model's performance was evaluated using the concordance index (C-index) and area under the receiver-operating characteristic curve (AUC) on both internal and external validation cohorts. The internal validation involved splitting the training cohort into training and validation sets. Three external validation cohorts from different regions (Wuhan, Hubei province excluding Wuhan, and Guangdong province) were used to assess the model's generalizability. The model's output was incorporated into an online patient triage tool providing personalized risk assessments for COVID-19 patients at admission, including the probability of critical illness within 5, 10, and 30 days.

The deep learning survival Cox model exhibited high predictive performance with a C-index of 0.894 (95% CI: 0.857–0.930) and AUC of 0.911 (95% CI: 0.875–0.945) on the internal validation set. This was superior to the traditional Cox model (C-index 0.876, AUC 0.889) and the CURB-65 model (C-index 0.75). External validation cohorts showed C-indexes of 0.890, 0.852, and 0.967, respectively, demonstrating robust performance across diverse patient populations. Risk stratification based on the model identified three groups: low, medium, and high risk of critical illness, with significantly different probabilities of developing critical illness (0.9%, 7.3%, and 52.9%, respectively). Furthermore, analysis of follow-up data revealed the model's ability to monitor the risk of critical illness during hospital stay, with improved performance (AUC 0.960, C-index 0.935). An online tool based on the model allows healthcare professionals to quickly assess the critical illness risk of admitted COVID-19 patients.

The superior performance of the deep learning model compared to traditional methods highlights the power of deep learning to capture complex, non-linear relationships between clinical features and the risk of critical illness. The model's generalizability across different regions and healthcare settings suggests its broad applicability. Early identification of high-risk patients allows for timely intervention, improved resource allocation, and potentially better patient outcomes. The online tool significantly simplifies the risk assessment process, making it easily accessible to clinicians, thus improving the efficiency and accuracy of patient triage. This study strongly supports the integration of deep learning models into clinical workflows for improved decision-making in the management of COVID-19 patients.

This study successfully developed and validated a deep learning-based survival model for predicting the risk of critical illness in COVID-19 patients. The model's high accuracy and generalizability, coupled with a user-friendly online tool, provide a valuable resource for healthcare professionals. Future research could explore the integration of additional data sources, such as imaging data and time-varying clinical variables, to further enhance the model's predictive power and refine patient risk stratification.

The study's retrospective nature and reliance on existing medical records may limit the generalizability of the findings. Missing data, although addressed through imputation, could still influence the model's performance. The external validation cohorts, while diverse geographically, may not represent all patient populations universally. The study focused on the risk of critical illness and didn’t directly address mortality.