Coronavirus disease 2019 (COVID-19) is a contagious disease caused by the virus SARS-CoV-2, which has had a profound impact on global economies, healthcare systems, and social norms (1). Since the initial case was identified in Wuhan, China, in December 2019 (2), over 777 million individuals have been infected, and more than 7 million have died worldwide (as of 9 February 2025) (3). The average time from exposure to symptom onset is five days, and approximately 5% of patients with COVID-19 experience severe symptoms necessitating intensive care (4). While the diagnosis of COVID-19 is facilitated by the use of rapid antigen tests (RATs) (5) and polymerase chain reaction (PCR) (6) technology, the challenge lies in accurately assessing the severity of the disease based on clinical data and chest x-ray features (7). In 2020 WHO developed a Clinical Progression Scale, patients have been categorized as those with mild disease (ambulatory, not requiring supplemental oxygen), those with moderate disease (hospitalized, might requiring low-flow oxygen), and those with severe COVID-19 (on HFNC, NIV, IMV, or ECMO) (8). Nevertheless, accurate prediction of the prognosis of COVID-19 remains an elusive endeavor. Predicting COVID-19 mortality is important as it has significant implications for the selection of pharmacologic treatments, management strategies, and for family planning and goals of care discussions (9).

Previous clinical decision models have focused on common health data rather than comorbidities, which might have negative impact on accuracy since comorbidities and COVID-19 mortality are correlated (10). The 4C Mortality Score, which was identified as the most promising risk stratification model in numerous systematic reviews (11, 12), is a risk stratification tool that predicts in-hospital mortality rate for hospitalized COVID-19 patients with eight parameters (age, sex, number of comorbidities, respiratory rate, peripheral oxygen saturation, level of consciousness, urea level, and C reactive protein) (13). Risk stratification tool is a method that predict one's risk based on its clinical histories and other factors. The 4C mortality score sum up the scores of the eight parameters and range from 1 to 21, each represent a certain mortality rate respectively. While the 4C Score trained with 35,463 patients showed moderate diagnostic accuracy for mortality with derivation cohort area under the receiver operating characteristic curve (AUC) of 0.79, it performs poorly on other cohorts, with AUC ranging from 0.63 to 0.73 (13). In the present study, we sought to increase the accuracy by incorporating the identity of the comorbidity, not just the number of comorbidities, to the prediction. We developed the Comorbidities and Clinical Indicators on the Mortality Model (CCIMM), a machine learning model that can accurately predict a patient's mortality rate within 30 days of hospitalization, using 79 comorbidities that are readily available at the time of admission.

Our findings demonstrate that XGBoost outperformed other algorithms tested in this study for predicting 30-day mortality among hospitalized COVID-19 patients. The inclusion of vital signs significantly enhanced model performance, achieving an AUC of 0.96 (Table 3). While the results confirmed the importance of dynamic predictors such as pulse rate, oxygen saturation, blood pressure, and respiratory rate, further discussion is warranted to explore broader implications and contextualize these findings within existing clinical and research frameworks.

Machine learning has increasingly been recognized as a transformative tool in addressing complex interactions in high-dimensional clinical datasets. Traditional statistical methods, such as logistic regression and LASSO, often lack the capacity to capture nonlinear relationships and interdependencies, limiting their predictive accuracy (27, 28). Ensemble learning methods, including random forest and gradient-boosted decision trees (GBDT), have demonstrated superior performance in such scenarios (29). XGBoost, as a leading GBDT algorithm, leverages an iterative approach to optimize predictions and effectively handle missing or sparse data (30). This capability, combined with its robustness in incorporating both static and dynamic clinical variables, underscores its relevance for healthcare applications, particularly during a global health crisis like COVID-19.

A key strength of this study lies in its use of SHAP analysis to enhance model interpretability. One of the significant barriers to adopting machine learning in healthcare is the “black-box” nature of many algorithms. By quantifying the contributions of individual features to model predictions, SHAP analysis provides a transparent framework that bridges the gap between advanced computational methods and clinical decision-making. Identifying predictors such as chronic kidney disease, diabetes without complications, and hypertension allows clinicians to better understand the rationale behind predictions. This transparency not only fosters trust but also facilitates the integration of machine learning tools into routine clinical workflows, ensuring that decisions are informed and actionable.

In the early days, clinical diagnosis was purely based on physician's experiences with x-ray images and check-up data, lacking precision and efficiency (8). With the rise of ML, more models have been produced to estimate mortality and severity. According to systematic reviews in recent years (11, 12), the 4C mortality score was identified as the most promising model. The 4C Mortality Score is a risk stratification tool that predicts in-hospital mortality rate for hospitalized COVID-19 patients with eight parameters routinely available at hospital admission: age, sex, number of comorbidities, respiratory rate, peripheral oxygen saturation, level of consciousness, urea level, and C reactive protein (13). Its dataset was stratified from 260 hospitals across England, Scotland, and Wales. The 4C Score showed moderate diagnostic accuracy for mortality with derivation cohort area under the receiver operating characteristic curve (AUC) 0.79. However, it performs poorly on other cohorts, with AUC ranging 0.63∼0.73. Therefore, it cannot be implemented worldwide.

Previous clinical decision models, such as the 4C Mortality Score, have focused on common health data rather than comorbidities, lacking accuracy in assessing severity and mortality. Our research bridges this gap by adding 79 comorbidities to the prediction. We developed the Comorbidities and Clinical Indicators on the Mortality Model (CCIMM), a machine learning model that can accurately predict a patient's mortality rate within 30 days of hospitalization. We came up with two models: one comprising demographic information, vital signs upon admission, and underlying comorbidities; the other, comprising only demographic information and underlying comorbidities. The former one, with vital signs, has the highest AUC of 0.96. It can be used in the emergency room, where patients have unstable vital signs. The latter, without vital signs, has an AUC of 0.93. It can be used in outpatient conditions, where vital signs are rather stable and patients often have chronic illnesses. This model is also helpful at understanding the association between comorbidities and 30-day mortality rate. In addition, both models used federated learning to integrate data from three hospitals, resulting in models suitable for broader population needs.

This research has allowed us to gain insight into the impact of comorbidities and clinical indicators on mortality in COVID-19. This is in contrast to the 4C score, which assesses comorbidities not by the identities of the comorbidities, but by the number of comorbidities, in a categorical fashion separated into three groups- 0, 1 or ≥2. We use machine learning to deal with large amounts of data, testing different algorithms in hopes of getting the best results. CCIMrM can speed up the diagnostic agenda by being an accurate indicator of whether the patient needs antiviral medication. It can prevent hospitals from running out of medical supplies or isolation wards in devastating situations, and serve as an indicator of disease severity for patients and families. The model can also be applied to future epidemics, providing physicians with a quick and easy prediction. Furthermore, the best performing model was trained through federated learning. Our study extends the current understanding of mortality prediction for hospitalized COVID-19 patients by developing an accurate and explainable federated learning model based on comorbidities and clinical indicators.

Interestingly, the findings challenge certain assumptions in current clinical guidelines, particularly in the context of prescribing Paxlovid, which is an oral antiviral medication prescribe to patients with mild to moderate COVID-19. It can decrease the rate of severe COVID-19 or mortality with adjusted hazard ratios of 0.54 (31). While existing guidelines in Taiwan recommend Paxlovid for patients with conditions like diabetes mellitus, cerebrovascular disease, and chronic lung disease (12, 32), our study found these comorbidities to have a neutral or even protective association with 30-day mortality. This discrepancy underscores the need for continuous refinement of evidence-based guidelines using real-world data and advanced analytical techniques. For instance, while chronic kidney disease and BMI ≥30 remain high-priority risk factors, conditions like dementia and rheumatic diseases may warrant reevaluation based on the observed data trends. Additionally, the inconclusive findings regarding cancer highlight the necessity of individualized treatment plans rather than a one-size-fits-all approach to antiviral therapy.

Federated learning emerged as a critical tool in this study, addressing one of the main challenges in collaborative medical research—data privacy. By enabling the training of machine learning models across decentralized datasets, federated learning ensures that patient data remains localized while still contributing to a robust, generalizable model. This approach proved particularly effective in improving performance across external cohorts, even though it led to a slight decline in the AUC for the Taipei cohort. Such trade-offs underscore the potential of federated learning to enhance the scalability and applicability of AI solutions in diverse healthcare settings, where data-sharing restrictions often impede collaborative advancements.

Despite its promising outcomes, the study has limitations that merit further discussion. First, the model's reliance on readily available clinical features, while advantageous for implementation, excludes other potentially valuable data sources, such as imaging and laboratory results. For example, chest x-rays and CT scans are often pivotal in assessing COVID-19 severity, and their integration into future models could significantly enhance predictive accuracy. Multimodal models that combine clinical, imaging, and laboratory data, as well as nuances about the severity of disease, would likely provide a more comprehensive understanding of patient risk profiles.

Second, the relatively small dataset focused exclusively on the Taiwanese population limits the generalizability of the findings. While federated learning mitigated some of these limitations by improving model performance in external cohorts, the demographic and clinical characteristics of the original dataset may still introduce biases. Expanding the dataset to include diverse populations from different geographic regions and healthcare systems would provide a more representative basis for training and validation. Additionally, the observed protective effects of certain comorbidities, such as diabetes and chronic lung disease, may reflect unique population-level characteristics rather than universal trends. Further studies are needed to validate these findings and investigate their underlying mechanisms.

The potential of integrating time-series data into predictive models represents another avenue for future research. COVID-19 is characterized by rapid disease progression, making real-time monitoring of vital signs crucial for early intervention. Incorporating continuous data streams from wearable devices or bedside monitors could enable dynamic risk assessment, allowing clinicians to adjust treatment plans in real time. In addition, as COVID variants evolve and clinical practices change with time, a real-world implementation would require frequent updates to maintain satisfactory performance, especially in a federated setting. Such advancements would move machine learning from static prediction models to dynamic, context-sensitive tools that align closely with the realities of patient care.

Another critical area for exploration is the role of interpretability in enhancing clinician adoption of machine learning tools. While SHAP analysis offers valuable insights into feature importance, further efforts are needed to ensure that these explanations are presented in a manner that is intuitive and clinically relevant. For example, integrating visualizations of SHAP values into electronic health record systems could provide clinicians with real-time feedback on patient risk factors, enabling more informed decision-making. Additionally, user-centered design principles should guide the development of interfaces that present machine learning outputs in ways that align with clinicians' workflows and information needs.

The discrepancies observed between this study's findings and existing clinical guidelines also raise broader questions about how evidence is generated and translated into practice. Machine learning models offer the advantage of being data-driven, allowing for the identification of patterns that may not align with traditional clinical assumptions. However, these insights must be contextualized within the broader framework of clinical expertise and patient care priorities. For instance, while certain comorbidities may show a protective association with mortality in the study population, this does not necessarily negate their significance in other aspects of disease management. Collaborative efforts between data scientists, clinicians, and policymakers are essential to ensure that machine learning insights are translated into guidelines that are both evidence-based and clinically meaningful.

In conclusion, this study highlights the transformative potential of machine learning, particularly XGBoost, in predicting 30-day mortality among hospitalized COVID-19 patients. While the inclusion of vital signs significantly enhanced predictive accuracy, federated learning demonstrated its value in improving generalizability across diverse cohorts. The integration of SHAP analysis provided a transparent framework for understanding model predictions, fostering clinician trust and facilitating practical application. Future research should focus on addressing the limitations of current models by incorporating multimodal data, expanding population diversity, and enhancing real-time monitoring capabilities. By aligning advanced computational techniques with clinical expertise, machine learning has the potential to revolutionize risk stratification and treatment optimization in COVID-19 care, paving the way for more personalized and effective healthcare solutions.