A machine learning method for predicting the probability of MODS using only non-invasive parameters

Highlights

• Only easily obtained noninvasive parameters were used for MODS prediction, allowing the model to be applied in pre-hospital first aid or battlefield environments.

• ML is better than traditional scoring method in predicting MODS and mortality in ICU.

• The optimal features were selected for modeling to balance the number of parameters and predictive performance, which can make the model easily be integrated in emergency equipment and provide more robust advice in pre-hospital first-aid.

ABSTRACT

Objectives

Timely and accurate prediction of multiple organ dysfunction syndrome (MODS) is essential for the rescue and treatment of trauma patients However, existing methods are invasive, easily affected by artifacts and can be difficult to perform in a pre-hospital setting. We aim to develop prediction models for patients with MODS using only non-invasive parameters.

Method

In this study, records from 2319 patients were extracted from the Multiparameter Intelligent Monitoring in Intensive Care Ⅲ database (MIMIC Ⅲ), based on the sequential organ failure assessment (SOFA) score. Seven commonly used machine learning (ML) methods were selected and applied to develop a real-time prediction method for MODS based on full parameters (laboratory parameter. drug and non-invasive parameters, 57 parameters in total) and non-invasive parameters only (17 parameters) and compared with four traditional scoring systems.

Results

The prediction results using LightGBM (LGBM) and Adaboost based on the full parameter modeling were 0.959 for area under receiver operating characteristic curve (AUC), outperforming four traditional scoring systems. The removal of 40 parameters and retaining of 17 non-invasive parameters decreased the AUC value of LGBM by 0.015, which still outperformed all traditional scoring systems.

Conclusions

A real-time and accurate MODS prediction method was developed in this paper based on non-invasive parameters by comparing the performance of four ML methods, which proved to be superior to the traditional scoring systems. This method can help medical staff to diagnose MODS as soon as possible and can improve the survival rate of patients in a pre-hospital setting.

Introduction

Multiple organ dysfunction syndrome (MODS) is defined as the development of potentially reversible physiologic derangement involving two or more organ systems not involved in the disorder after injury. It is the leading cause of morbidity and mortality in critically ill patients [1], [2], [3].

MODS accounts for a considerable proportion of healthcare resources associated with acute trauma treatment [4], [5] and is recognized as the final common route preceding death in critically ill patients [6], [7], [8]. The introduction of early-phase management strategies, such as damage control resuscitation and prediction models, has increased the survival rate of injured patients to reach critical care [9], [10]. Corso F D et al. [11] noted that the primary task of MODS therapy is early recognition and early organ function support.

Several organ dysfunction scoring systems have been developed to describe, quantify and recognize organ dysfunction or failure in intensive care unit (ICU)patients. The three scoring systems commonly used in adult studies are the logistic organ dysfunction score (LODS) [12], the sequential organ failure assessment (SOFA) [13] and the Marshall multiple organ dysfunction score (Marshall) [14]. Pediatric scoring systems have been extrapolated from some of these systems [15]. In addition, Quick SOFA (QSOFA) and Denver score are also commonly used for rapid clinical identification of MODS [16], although QSOFA is mainly used for the prediction of sepsis. All the above scoring methods have good performance, the difference is sensitivity, specificity or balance [[17], [18], [19]].

Traditional scoring systems have the following limitations. First, definitions of MODS are not uniform, even when the same scoring system is used. Rendy L et al. [20] defined MODS based on the SOFA score as score of organs>=1 in >1 system. Shepherd JM et al. [21] defined MODS by the occurrence of a total SOFA score greater than 5, affecting two or more organs. Matthias Fröhlich et al [22] defined that a score of 3 or higher for one of the organ systems meant failure of this system. Yasser Sakr et al. [23] defined organ failure as a SOFA score >2 for any of the six organs/systems evaluated, and multiple organ failure, as more than one failing organ. Second, traditional scoring systems rely on laboratory parameters that require specialized biochemical analysis, which is unavailable in remote areas with field-rescue requirements. Third, traditional scoring systems only consider the current value of the physiological parameters, but a lot of useful multi-dimensional information, such as the change and statistical characteristics of the parameters in the time dimension, is ignored.

The recent introduction of machine learning (ML) in the medical field has produced remarkable efficiency in the diagnosis and decision-making process. Li Han et al. [24] used electrocardiogram (ECG) and phonocardiogram (PCG) signals to identify coronary artery (CAD) disease and achieved a CAD classification accuracy of 95.62%, based on the ECG and PCG recordings cropped to 15 s. However, the proposed method requires a large amount of annotated data for supervised learning. Kashif Naseer Qureshi et al. [25] trained support vector machines, decision trees, bayesian model and KNN by collecting clinical parameters, such as medication history, age, body mass index, gender, oxygen saturation and blood pressure of patients, using four-fold cross-validation. The integrated learning method combined with four ML models realized dynamic warning of the possibility of cardiovascular disease in the future, and the prediction accuracy reached 86.72%. Hans-Christian and Thorsen-Meyer et al. [26] used demographic information and temporal physiological signals of patients to train a recursive neural network with 1-hour time resolution to predict the possibility of death within 90 days, and the area under receiver operating characteristic curve (AUC) of the prediction results ranged from 0.73 to 0.85.

The current research on MODS focuses on three aspects. Firstly, biomarker-based prediction of MODS. Zeng J [27] analyzed the dynamic changes and predictive values of nuclear factor-B (NF-B) combined with IL-6 and tumor necrosis factor- (TNF-) in peripheral blood in MODS. Jin JJ et al. [28] used neutrophil-derived long noncoding RNA IL-7R predicts development of MODS. Secondly, electronic health records-based analysis of the condition and prognosis of MODS, Cole E [5] analyzed characteristics of MODS subtypes in trauma critical care at a population level. Fan BW [29] developed a prediction model which could make the accurate early prediction of the recovery in pediatric sepsis patients from MODS to a mild state. Meanwhile, Thomas Desautels et al. [30] achieved the prediction of sepsis in the ICU ward using a custom-designed ML model using six bedside parameters and age. Thirdly, comparison and usability analysis of different diagnostic criteria or score systems. Typpo KV et al. [31] developed the diagnostic criteria of new and progressive MODS and scoring systems that might be used to assess and monitor the severity and progression of multiple organ dysfunction syndrome in children presented. Williams JM et al. [32] compared the diagnostic accuracy of SIRS and QSOFA for organ dysfunction. Asuroglu Tunc and Ogul Hasan [33] combined Convolutional Neural Networks features with Random Forest to predict SOFA scores of sepsis patients.

However, the application of these modern approaches to predict in patients with MODS has had limited success. Most studies to date have been marred using logistic regression models and univariate statistical methods that do not account for collinearity and complex interactions among [34] and focused on the group of Elderly or children [31,35]. Moreover, many ICU prediction models overemphasize biomarker and laboratory variables [5,29].

In the present study, we aimed to develop a predictive model using only non-invasive parameters to assist medical personnel in the early warning of MODS, which can be widely used in remote areas, emergency public health event sites and battlefield fronts where there is a lack of laboratory parameter testing equipment and specialized clinicians.