Machine learning-based model for predicting the occurrence and mortality of nonpulmonary sepsis-associated ARDS

ARDS has consistently been a research hotspot in critical care medicine. The understanding of the pathogenesis of ARDS has deepened over time, and the number of available treatment methods, including lung-protective ventilation, prone positioning, extracorporeal membrane oxygenation, and neuromuscular blockade, continues to grow. However, the rate of mortality in ARDS patients still remains as high as 40%⁶. In this study, the 28-day and 60-day mortality of non-pulmonary sepsis patients in the MIMIC database were 19.5% and 23.7%, respectively. These rates are slightly lower than those reported in the literature, possibly because sepsis patients with lung infections tend to have more severe lung injuries, leading to higher mortality. The causes of ARDS are multifactorial and can be broadly categorized into pulmonary and extrapulmonary factors. The most common pulmonary factor is pulmonary infection, accounting for approximately 60% of cases. The most significant extrapulmonary factor is nonpulmonary sepsis⁷. Sepsis is a life-threatening organ dysfunction caused by a dysregulated host response to infection⁸. Nonpulmonary sepsis involves a dysregulated inflammatory response that can promote the development of ARDS through multiple mechanisms^9,10. Research has indicated that clinical physicians have a relatively low recognition rate for ARDS. The recognition rate for mild ARDS is 51%, whereas that for severe ARDS is 79%¹¹. This is one of the factors contributing to the poor prognosis of ARDS.

ARDS progresses rapidly, with the majority of patients developing ARDS within 12–48 h after admission¹². Currently, there is still limited strong evidence or consensus to guide clinical physicians in the early identification of ARDS¹³. Clinical scores and biomarkers are often used to predict the occurrence of ARDS. The Lung Injury Prediction Score (LIPS) is a relatively common clinical scoring system used for this purpose¹⁴. The LIPS combines patient risk factors, comorbidities, and acute physiological parameters to predict the occurrence of ARDS. The LIPS has a greater negative predictive value and a lower positive predictive value, with studies indicating that the positive predictive value of the LIPS is only 10%¹⁵. Some biomarkers, such as von Willebrand factor, angiopoietin 2, and IL-8, are also used to predict the occurrence of ARDS. These biomarkers can predict the development of ARDS to some extent, but they are often challenging to obtain in clinical practice⁶. In this study, LASSO results indicated that the occurrence of NPS-ARDS is associated with parameters such as patients’ hematocrit, oxygenation index, and mean arterial pressure at admission. Mortality in NPS-ARDS patients was related to indicators such as lactate, prothrombin time, and base excess. These parameters are relatively easy to obtain in clinical practice, enhancing the practicality of our model built using these variables.

Historically, machine learning has been used primarily in natural language processing, financial forecasting, and fields such as physics and astronomy. With the exponential growth in computational power and the availability of vast health care datasets in recent years, machine learning has been applied more extensively in the field of medicine, leveraging its unique advantages⁵. The application of machine learning as a novel approach for data analysis in sepsis primarily includes sepsis definition, early recognition, subtype classification, prediction, and treatment decision-making. Supervised learning is mainly used for building sepsis prediction models, whereas unsupervised learning can be employed for exploring sepsis subtypes¹⁶. In this study, we used five types of machine learning techniques (including K-nearest neighbour, extreme gradient boosting, support vector machine, deep neural network, and decision tree methods) to build models for predicting the occurrence and mortality of nonpulmonary sepsis-associated ARDS.

The MIMIC-IV database is a large medical database established through collaboration between the Beth Israel Deaconess Medical Center (BIDMC) and the Massachusetts Institute of Technology (MIT). It comprises clinical records from over 257,000 patients and provides data for numerous researchers worldwide¹⁷.

We extracted data on non-pulmonary sepsis patients with and without ARDS from the MIMIC database using Structured Query Language, resulting in a total of 11,409 sepsis patients included in the study. Among the remaining 11,409 sepsis patients, 7,632 experienced NPS-ARDS, accounting for 66.9%. NPS-ARDS has a high incidence rate, and establishing an early prediction model for NPS-ARDS can be helpful for the early identification and treatment of this condition. In the baseline data of the two groups of patients, non-pulmonary sepsis patients were predominantly elderly males, with similar distributions of vital signs and infection sites. Patients who developed NPS-ARDS exhibited higher lactate levels. The baseline data of local patients also shared similar characteristics, which enhances the applicability of using local data to validate the model established with MIMIC data.

Early assessment of the direction of disease progression in critically ill patients allows for the timely implementation of preventive measures to mitigate the occurrence of organ dysfunction, achieving a “prevention before disease” effect¹⁸. Additionally, it facilitates communication between health care providers and patients. However, the condition of critically ill patients can change rapidly, and many factors can impact patient outcomes, making it challenging for clinical physicians to anticipate the course of the disease. Early recognition of ARDS is crucial for its timely prevention and management. Clinical physicians often manage to promptly identify ARDS caused by lung infections but may sometimes overlook extrapulmonary conditions, particularly ARDS triggered by nonpulmonary sepsis¹⁹. Machine learning has the ability to analyse large datasets and uncover underlying patterns, making it a novel approach for predicting clinical events in patients²⁰.

This study included sepsis patients from the MIMIC-IV database as the research subjects and established integrated models through machine learning techniques to predict the occurrence of ARDS in nonpulmonary sepsis patients and mortality in these patients after developing ARDS. Model accuracy was validated via a subset of patients from the MIMIC-IV database, and the results indicated that the model built with XGBoost demonstrated high accuracy.

Models established via machine learning methods are mostly validated through retrospective analysis of data from databases, with limited clinical application and even fewer models used for prospective research. Only approximately 2% of published studies conducted prospective validation²¹. In this study, internal validation was performed via MIMIC-IV data, and external validation was conducted via local data. The results demonstrated that the model established with the XGBoost method also exhibited a high level of accuracy. Furthermore, the model in this study was constructed using commonly used clinical indicators, making it highly practical. It can assist clinical physicians in the early identification of nonpulmonary sepsis-induced ARDS, particularly lethal ARDS. This model can provide guidance for identifying high-risk patients and implementing appropriate intervention measures.

Our study focused on patients with non-pulmonary sepsis and established a predictive model for predicting the occurrence and mortality of nonpulmonary sepsis-associated ARDS. After a thorough search, no studies were found that are entirely consistent with this research. Most current studies focus on developing models to predict the occurrence of ARDS in sepsis patients or mortality following ARDS.

In developing a model to predict the occurrence of ARDS in sepsis patients, Shen’s²² study indicated that the model built using Gaussian Naive Bayes (NB) had the highest predictive accuracy (78.6%), while Zhan’s²³ study showed that the model built using logistic regression had the highest accuracy (71.34%). In developing a model to predict mortality in sepsis-associated ARDS patients, Zheng’s²⁴ study demonstrated that the model built using Random Forest had the highest predictive accuracy (74.15%).

Using machine learning techniques, including k-nearest neighbour (KNN), extreme gradient boosting (XGBoost), support vector machine (SVM), deep neural network (DNN), and decision tree (DT) methods, we established predictive models for NPS-ARDS incidence and mortality in sepsis patients. In the internal validation, the model predicted NPS-ARDS and mortality with accuracies of 77.5% and 71.8%, respectively. In the external validation, the model predicted NPS-ARDS and mortality with accuracies of 78.0% and 81.4%, respectively. The accuracy of our model is similar to that of the aforementioned studies.

There are many types of machine learning models, and their principles are complex, making it challenging for clinicians to choose the most appropriate prediction model. Based on similar studies^22,23,24, The above machine learning methods were chosen for our study.

The pathophysiology of ARDS in patients with nonpulmonary sepsis are highly complex, making it challenging to identify the risk factors affecting mortality in such patients. Based on similar studies^22,23,24 and to avoid overlooking any parameters potentially related to outcomes, this study included all commonly accessible clinical parameters, such as patient demographics, laboratory values and vital signs. LASSO was then used to select the parameters related to the outcomes, reducing the model’s complexity and enhancing its practical applicability.

KNN is suitable for simple tasks with small datasets but performs poorly with high-dimensional and large-scale data. XGBoost offers high accuracy and strong capabilities for handling complex data, making it a popular choice for prediction tasks, though it requires complex tuning. SVM handles high-dimensional data well, ideal for tasks needing strong generalization, but it has high computational complexity. DNN excels in large-scale and complex tasks, especially for unstructured data, but requires significant resources and tuning. DT is easy to interpret and good for building preliminary models but is less accurate for complex tasks compared to other models²⁵.

XGBoost achieves higher accuracy because it reduces errors through gradient boosting, prevents overfitting with regularization, handles missing data well, and efficiently selects important features. It also supports parallel processing, which speeds up model training. These factors make it more robust and effective for prediction tasks²⁵.

In this study, starting from clinical practice, we were unable to satisfactorily explain the correlation between the individual parameters used for modelling and the outcomes. In the next phase of our research, our team will employ local interpretable model–agnostic explanations to address this ‘unexplainability’.

Additionally, owing to the imbalance between the samples, this model has a relatively high positive predictive value. In the next phase of our research, our team plans to conduct prospective external validation to further assess the model’s accuracy. Next, we aim to develop a real-time monitoring system that continuously collects patient data to provide real-time predictions, assisting physicians in making improved clinical decisions.