Invasive mechanical ventilation (IMV) is a cornerstone of organ support in severe COVID-19 patients with acute respiratory distress syndrome (ARDS). As widely experienced in ICUs during the SARS-CoV-2 pandemic, IMV frequently causes complications.1,2 Hospital services were overwhelmed not only by the surge of patients, but also by scarce human resources and equipment, lack of sufficient mechanical ventilators being probably the most relevant. In surge scenarios, appropriate triage strategies are therefore needed to allocate IMV or alternatives such as high flow nasal prongs. These strategies should be based on the knowledge and understanding of specific potential predictors3 that could help clinicians to personalize decisions regarding IMV.

There is still considerable controversy regarding who and when to intubate. Several recent studies have addressed the subject,4 although bias cannot be excluded in observational non-randomized trials. A retrospective study suggested that early intubation and IMV is associated with favorable outcomes but included only intubated patients instead of the whole population at risk.

Previous studies have identified covid-19 progression predictors including age, comorbidities, renal function, or immunodeficiency5 using traditional statistical approaches, where collinearity of data cannot be ruled out. Artificial intelligence (AI) is currently being used for COVID-19 risk stratification,6 studying multiple clinical features to increase effectiveness and efficiency in diagnosis, treatment, and prognosis. Self-explainable Machine learning (ML) techniques can help with risk factor selection through ranking methodologies.7 In this context, the utilization of artificial intelligence (AI) holds potential in facilitating the development of a conceptual model aimed at comparing the significance of variables. This can be achieved by employing regularization models8 to enhance predictor selection, followed by the implementation of the Generalized Linear Mixed-effects Model (GLMM)9–11 to construct the said conceptual model. Such an approach becomes particularly relevant when assessing and comparing outcomes across different AI models, enabling a comprehensive evaluation of variable significance. This is a novel methodology, leveraging modern machine learning techniques to provide rigorous and applicable insight into relevant clinical questions when randomized clinical trials are not feasible. From here on, in this paper, we aim to determine if potential predictors for invasive mechanical ventilation (IMV) are also determinants for mortality in COVID-19-associated acute respiratory distress syndrome (C-ARDS) while comparing the significance of variables in both cases.

The results of the present study include some highly relevant clinical results. First, the variable sets predicting IMV, and ICU mortality are different. Whereas oxygenation variables are independent predictors of IMV, ICU mortality is associated with increased age and LDH and the presence of comorbidities. The latter variables may be considered markers of two processes: COVID-19-associated inflammation and ICU-acquired superinfection (see Figure 4 in the Supplementary material). Secondly, the characteristics of pharmacological therapy, including the administration of steroid drugs, has little influence on both the need for IMV and ICU mortality, considering our results. We included in the analysis 64 patients not receiving steroids and 216 receiving this treatment, at the usual 6mg dexamethasone or equivalent daily dose. This is a remarkable finding, because the effect of steroids on mortality identified in a previous trial27 have influenced recommendations, as well as clinical practice, since its publication. It may be speculated that the decision to include and randomize or not at the discretion of the attending physicians, and based on undisclosed criteria, rendered different results by selecting a study subset of COVID-19 cases with different characteristics. In comparison, no inclusion-exclusion criteria for selection process were applied in our “pragmatic” type of cohort. Steroids were given to almost every patient unless a severe contraindication existed, after the results of the RECOVERY trial were made available.

The present study applied a novel methodology (logistic regression with regularization plus GLMM Tree mixed models) to evaluate the relative importance of several variables as predictors of significant clinical events. Using machine learning and a fine-grained longitudinal multifaceted database, we have established relevant variable value thresholds to support clinical decisions. Although the model would perform quite well as predictor for IMV and ICU mortality, with good positive predictive values, it is important to emphasize that this is not a predictive model in the classical sense, but an attempt to pinpoint the most important clinical events that represent turning points during the studied process (in this case, clinical management of patients not initially under IMV). In this sense, we should say that the inclusion of the likelihood ratio as an evaluation factor for comparing performance model was reach great results. However, following the premise of model explainability, we believe it is important to take this element into account as a final selection factor for the set of predictors that best fit daily clinical practice. This study demonstrates that predictor-ranking methodologies using self-explainable machine learning may support therapeutic decision-making using observational data, when randomized clinical trials are unfeasible or unethical.

Regarding with the strengths of our study, we would like to mention the quantity and quality of the data set. Collected data have a high level of detail, leveraging the power of strategically devised electronic health records (EHR), which include relevant information in a highly structured and recoverable format. Every effort was made to configure our EHR to optimally gather all relevant information about COVID-19 patients. Also, our anonymized database is available in the repository along with the script we used for statistical analysis, is highly detailed and has been extensively curated to reflect temporal evolution and to improve data quality as much as possible. In any case, the collection of variables from Electronic Health Records (EHR) may be biased, affecting data quality. Age and gender biases are possible, as well as biases related to the selection and measurement of clinical variables. These biases can lead to incomplete or skewed representations of certain population groups and may impact the validity and generalizability of research findings and clinical decision-making. It is important to be aware of these biases to ensure proper interpretation and use of EHR data.

On the other hand, the limitations of our study results relate mainly to its single-centered nature and require confirmation in a multicenter dataset to gain external validity. Our methodology would be perfectly suited for a multicenter study, including “center” as a random factor in the second (GLMM Tree) part of the process. We suggest that future research applying this methodology could focus on designing clinical studies using observational data to answer relevant clinical questions without the logistic requirements of a randomized clinical trial or for hypothesis-generating purposes. Furthermore, when considering the limitations of using generalized linear mixed effects models (GLMMs) for modeling causation in critical care medicine research, it is important to highlight the absence of explicit causality assumptions. GLMMs primarily focus on association or correlation analysis, lacking the ability to address the assumptions necessary for establishing causal relationships. Specifically, GLMMs do not provide frameworks for the identification of causal effects or account for unmeasured confounding variables, which are crucial considerations in causal inference. In contrast, causal inference methods, such as the potential outcomes framework, explicitly address these assumptions, offering a more comprehensive approach for investigating causality. Therefore, when establishing causal relationships between variables, researchers should carefully consider the limitations of GLMMs and opt for causal inference methods, which provide a more robust approach for investigating causality in critical care medicine research.

In conclusion, different variables predict IMV and ICU mortality in severe COVID-19 patients, suggesting that the therapeutic decision of when to use IMV has little impact on ICU mortality. Our methodology is a valid option to assess therapeutic decisions using observational data when randomized clinical trials are not feasible or ethical.

SM, MA and AN conceived the presented idea. SM and MA contributed equally as first authors. SM and MA developed the theory and performed the computations. AN conducted an independent literature search to identify potentially relevant studies. MS independently reviewed the search results to identify pertinent articles. MS, AN, TF and VY contributed to the interpretation of the results. SM, MA, AN, MS, FL and AC took the lead in writing the manuscript. All authors provided critical feedback and helped shape the research, analysis, and manuscript.

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

The authors declare that they have no conflict of interest.