Identifying Novel Clusters of Patients With Prolonged Mechanical Ventilation Using Trajectories of Rapid Shallow Breathing Index

Abstract

Objective: Patients with prolonged mechanical ventilation (PMV) are comprised of a heterogeneous population, creating great challenges for clinical management and study design. The study aimed to identify subclusters of PMV patients based on trajectories of rapid shallow breathing index (RSBI), and to develop a machine learning model to predict the cluster membership based on baseline variables.

Methods: This was a retrospective cohort study conducted in respiratory care center (RCC) at a tertiary academic medical center. The RCC referral criteria were patients with mechanical ventilation for at least 21 days with stable hemodynamic and oxygenation status. Patients admitted to the RCC from April 2009 to December 2020 were screened. Two-step clustering through linear regression modeling and k-means was employed to find clusters of the trajectories of RSBI. The number of clusters was chosen by statistical metrics and domain expertise. A gradient boosting machine (GBM) was trained, exploiting variables on RCC admission, to predict cluster membership.

Results: A total of 1371 subjects were included in the study. Four clusters were identified: cluster A showed persistently high RSBI; cluster B was characterized by a constant low RSBI over time; Cluster C was characterized by increasing RSBI; and cluster D showed a declining RSBI. Cluster A showed the highest mortality rate (72%), followed by cluster D (63%), C (62%) and B (61%; p = 0.005 for comparison between 4 clusters). GBM was able to predict cluster membership with an accuracy of > 0.95 in ten-fold cross validation. Highly ranked variables for the prediction of clusters included thyroid-stimulating hormone (TSH), cortisol, platelet, free thyroxine (T4) and serum magnesium.

Conclusions: Patients with PMV are composed of a heterogeneous population that can be classified into four clusters by using trajectories of RSBI. These clusters can be easily predicted with baseline clinical variables.

Keywords: ICU; gradient boosting machine; mortality; prolonged mechanical ventilation; rapid shallow breathing index.

Figure 1

Flowchart of patient enrollment and schematic illustration of the analysis workflow. RSBI, Rapid shallow breathing index; RCC, respiratory care center; LIME, local interpretable model-agnostic explanations.

Figure 2

Clustering to identify clusters of patients with prolonged mechanical ventilation. (A) The best number of clusters was chosen by using statistical metrics. Greater values of log likelihood indicate better model fit, whereas lower values of BIC and AIC indicate better model fit. (B) Trajectory characteristics of each cluster. Individual trajectories are represented by black lines and the cluster trajectory is colored. The cluster label and percentage are shown on the top of each panel. (C) Trajectory and 90% confidence interval for each of the ventilator parameters, stratified by the cluster membership. BIC, Bayesian information criterion; AIC, Akaike's information criterion; MV, minute ventilation; Pimax, maximum inspiratory pressure; RR, respiratory rate; RSBI, Rapid shallow breathing index; TV, tidal volume.

Figure 3

Gradient boosting machine training and interpretation. (A) Hyperparameter tuning for the gradient boosting machine model. We used grid search strategy to select hyperparameters with the highest accuracy. (B) Variable importance in the GBM model. Higher importance value indicates greater influence of the variable in differentiating the clusters. (C) LIME interpretation for four sample subjects. The horizontal axis is labeled by the sample ID. The observed cluster membership for patients 1, 2, 4 and 5 were A, B, C and D respectively. Blue (red) color indicates the variable is supporting for (contradicting against) a given cluster. For example, the subject 4 has magnesium <1.92 supporting for cluster C. (D) The iBreakdown explainer for patient #4 showed that there was more support for allocation to cluster C than to other clusters. The feature TSH = 0.023 strongly supports its assignment to cluster C, whereas the APACHE II = 30 on RCC arrival contradicts its assignment to cluster C. The short bar indicates the confidence interval for uncertainty. LIME, local interpretable model-agnostic explanations; HCT, hematocrit; WBC, white blood cell count; BUN, blood urea nitrogen; Cr, creatinine; RDW, red distribution width; MCV, mean corpuscular volume; GCS, Glasgow coma scale; GCSM, motion component of GCS.

Figure 4

Comparisons of the Gradient boosting machine with other models. The performance metrics of accuracy and kappa was reported. The boxplot shows the median and range of the performance metrics in resampled datasets. GBM, gradient boosting machine; LASSO, Least Absolute Shrinkage and Selection Operator; RF, random forest.