Machine learning to predict extubation success using the spontaneous breathing trial, objective cough measurement, and diaphragmatic contraction velocity: Secondary analysis of the COBRE-US trial

Inclusion criteria for the study were adult patients aged 18 and older who needed IMV for over 48 hours and fulfilled the requirements to begin the weaning process. Patients were required to exhibit a robust cough, show no respiratory secretions, and have resolved the acute phase of the condition that necessitated IMV. They must have a stable cardiac state, without the need for vasopressor support or with minimal doses, an acid-base balance, euthermia, hemoglobin >7 g/dl, a Glasgow Coma Scale score >12, and negative Confusion Assessment Method for the ICU (CAM-ICU). Additionally, the patient must maintain a SaO2 >90% with FiO2 ≤0.4, a PaO2/FiO2 ratio >150 mmHg, and a PEEP ≤8 cm H2O. Patients with acute brain injury, neuro-surgical intervention, pregnant women, and those with neuropsychiatric diseases or diaphragmatic paralysis were excluded from the study.

Variables analyzed included sociodemographic variables, the cause of respiratory failure, arterial blood gases before extubation, the mode of ventilation during the weaning process, duration of weaning, days from ICU admission to the start of weaning, and the total length of stay in the ICU. Patients at risk of WF were treated with non-invasive mechanical ventilation (NIMV) or high-flow nasal cannula (HFNC) oxygen therapy.

All patients underwent a 30-minute SBT using a T-piece or pressure support ventilation [6,17]. The criteria for test failure are described in the Appendix 1. The SBT was discontinued in cases of test intolerance [6,7,8,9,10,11,12,13,14,15,16,17].

In the OCM, normal saline (0.9%) was infused via closed suction at the end of inspiration, and peak expiratory flow during the resulting involuntary cough was measured [18]. Ventilatory parameters were set to spontaneous mode without assistance during cough assessment. The objective classification of peak expiratory flow induced by cough is described in the Appendix 2 [18].

A diaphragmatic ultrasound in M-mode was performed to measure DCV (slope, cm/s) from three consecutive normal breaths, with average values used in the analysis [19,20,21]. Assessments were conducted with the patient in a supine position, utilizing either a subcostal or intercostal approach at the midclavicular or anterior axillary lines. The chosen side for measurement depended on technical ease and the clinical judgment of the ultrasound evaluator.

Five ML techniques were used to predict success of SBT and extubation. OCM and DCV were used as predictors for SBT success. On the other hand, OCM, DCV, and successful SBT were used as predictors for extubation success. The study employed five ML techniques, which included two unsupervised methods (hierarchical clustering and k-means clustering) and three supervised methods (support vector machines, decision trees, and neural networks) [22,23]. Each technique underwent training using 10-fold cross-validation. Furthermore, patients were randomly allocated to either the training or test group in a 90/10 ratio. This allocation created a test group that was not involved in training ML techniques [22,23]. This approach enables the assessment of the system's predictive performance on both the training samples and on new samples, mirroring a real-world scenario.

The unsupervised methods were used to let the computer try to find patterns that might reveal interesting associations. Hierarchical clustering groups patients depending on how similar they are, forming pairs and then grouping similar pairs to form a dendrogram [22,23]. On the other hand, k-means clustering groups classify patients by grouping them according to their location in an n-dimensional space. Regarding the supervised methods, the decision tree is like a series of yes or no questions that are easily understood and can be clinically replicated. Support vector machines are a method that involves finding an n-dimensional surface equation that separates the patients into an n-dimensional space [22,23]. And artificial neural networks simulate human brain function by having layers of nodes (neurons) connected and in each node, mathematical operations are applied to the data, leading to a classification probability for each class of patient [22,23].

The classification of patients by each technique, both in the training and test samples, was compared with the actual clinical outcomes. The diagnostic performance of each method was determined using accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) [22,23,24].

Categorical variables are reported as both absolute and relative frequencies [24]. Continuous variables are described using means and standard deviations (SD) or medians and interquartile ranges (IQR), based on their distribution. The Shapiro-Wilk test was used to evaluate the distribution of continuous variables [24].

Data were collected using REDCap software (25). Machine learning models were trained and tested using MATLAB Release 2023a (The MathWorks, Inc., Natick, Massachusetts, USA). Data were analyzed using IBM SPSS software (25, IBM Corp., Chicago, IL, US).

The studies involving human participants were reviewed and approved by the Ethics Committee of the Fundación Neumológica Colombiana (approval number: 201806-23607). Although it is considered minimal risk research, in which no intervention will be carried out, the signature of an informed consent was collected.

Seventy-five percent (261/367) of the cases were attributed to hypoxemia (PaO2 < 60, usual FiO2), 14.9% (52/367) to ventilatory failure due to shock, followed by 6.6% (23/367) due to hypercapnia (pH < 7.25, elevated CO2) (Table 2). Ninety-four percent (345/367) of the population were admitted to the ICU for medical reasons.

The unsupervised methods did not find any interesting patterns and did not have a good classification rate. Among the supervised methods, decision trees achieved the highest accuracy in both the training group (77.3%) and test group (69.6%) (Table 3). The other supervised methods, support vector machines and neural networks, demonstrated the same diagnostic accuracy: 77.3% for the training group and 69.6% for the test group.

Like before, the unsupervised methods did not find any interesting patterns and did not show good classification rate. In this case, the decision trees also demonstrated the best diagnostic accuracy with 85.9% for the training group and 95.7% for the test group (Table 4). The support vector machines and neural networks both achieved a diagnostic accuracy of 85.9% for the training group and 93.5% for the test group.

The observational nature of our study exposes it to the risk of confounding factors. Although our goal was to develop a parsimonious predictive model, we acknowledge the importance of considering additional clinical variables, such as laboratory tests, risk scores, ventilation time, and ventilator settings, to more robustly and clinically substantiate predictions for patients undergoing weaning in ICU, ensuring a more accurate diagnosis by intensivists. Furthermore, dimensionality reduction through principal component analysis could be considered. A limitation of this study is the potential influence of confounding biases on the findings, which should be interpreted within the context of a prospective, multicenter observational study, emphasizing the need for external validation of the results.

This study's strengths are attributed to its prospective design and multicenter approach, derived from the study conducted by Varón-Vega et al. [7]. This post hoc analysis incorporates diaphragmatic ultrasound, OCM, and bedside tests into its predictive models, which have been validated with good reproducibility and predictive capability when used individually or in combination. Additionally, intensive care staff were trained to perform transthoracic ultrasound and tests. However, further research is needed to explore the potential of different models that include other clinical variables.