During the five-year study period, a total of 3,682 ICU patients were mechanically ventilated for more than four hours. Of these, 1,314 (35.6%) patients were excluded from the analysis (Figure 1).
Figure 1: Flow chart of the ventilated patient population and their outcomes.
The remaining 2,368 patients were included, where 1,895 (80%) patients were divided into the training cohort and 473 (20%) patients into the test cohort. The baseline characteristics are shown in (Table 1). Among the included patients, the ICU readmission rate, general ward mortality, and prolonged hospital stay were 12.7%, 3.1%, and 24.4%, respectively. The median length of stay in the hospital post-discharge from ICU among good versus poor outcome groups was 5.0 days versus 14.9 respectively.
Table 1: Baseline characteristics of the patient population
|
|
|
Total (n=2368)
|
Good Outcome (n=1415)
|
Poor Outcome
(n=953)
|
p-value
|
Demographics
|
|
|
|
|
Age [median (IQR)]
|
63.0 (52.0-72.0)
|
61 (49-71)
|
65 (55-74)
|
0.001**
|
Male gender - n (%)
|
1547 (65.3%)
|
908 (64.2%)
|
639 (67.1%)
|
0.150
|
Admission diagnosis (Medical) - n (%)
|
1371 (57.9%)
|
805 (56.9%)
|
566 (59.4%)
|
0.226
|
CCI [median (IQR)]#
|
2.0 (1.0-2.0)
|
1.0 (0.0-2.0)
|
1.0 (0.0-2.0)
|
<0.001**
|
BMI [median (IQR)]
|
24.0 (20.9-27.7)
|
24.1 (21.0-27.8)
|
23.8 (20.6-27.4)
|
0.241
|
APACHE II at ICU admission [median (IQR)]
|
17.0 (13.0-22.0)
|
17.0 (12.0-23.0)
|
17.0 (13.0-22.0)
|
0.589
|
SOFA Score at ICU admission [median (IQR)] #
|
3.0 (2.0-5.0)
|
3.0 (2.0-5.0)
|
3.0 (2.0-5.0)
|
<0.001**
|
Vasopressor support during ICU stay - n (%)
|
2366 (99.9%)
|
1414 (99.9%)
|
952 (99.9%)
|
0.779
|
Heart Failure as the cause of index ICU admission - n (%)
|
270 (11.4%)
|
183 (12.9%)
|
87 (9.1%)
|
0.004**
|
Pneumonia as the cause of index ICU admission - n (%)
|
473 (20.0%)
|
270 (19.1%)
|
203 (21.3%)
|
0.187
|
Duration of MV, hours [median (IQR)]
|
27.5 (15.0-55.0)
|
20.0 (10.0-43.0)
|
29.0 (6.0-81.0)
|
<0.001**
|
ICU LOS (Days) - [median (IQR)]
|
69.4 (43.2-125.9)
|
62.0 (41.3-100.0)
|
89.9 (47.5-168.7)
|
<0.001**
|
Comorbidities n (%)
|
|
|
|
|
COPD
|
95 (4.0%)
|
72 (5.1%)
|
23 (2.4%)
|
0.001**
|
Other Chronic respiratory diseases - Severe TB/ ILD/ Chest wall deformity/OSA/ Bronchiectasis
|
95 (4.0%)
|
48 (3.4%)
|
47 (4.9%)
|
0.061
|
Stroke
|
175 (7.4%)
|
51 (3.6%)
|
124 (13.0%)
|
<0.001**
|
Chronic Kidney Disease
|
145 (6.1%)
|
63 (4.5%)
|
82 (8.6%)
|
<0.001**
|
Cirrhosis
|
52 (2.2%)
|
31 (2.2%)
|
21 (2.2%)
|
0.984
|
Immunocompromised state
|
5 (0.2%)
|
3 (0.2%)
|
2 (0.2%)
|
0.992
|
ICU discharge
|
|
|
|
|
Heart rate at ICU discharge [median (IQR)]
|
84.0 (74.0-94.0)
|
84.0 (74.0-94.0)
|
85.0 (75.0-95.0)
|
0.040*
|
Respiratory rate at ICU discharge [median (IQR)]
|
19.0 (18.0-22.0)
|
19.0 (18.0-22.0)
|
19.0 (18.0-22.0)
|
0.103
|
GCS at ICU discharge [median (IQR)]
|
15.0 (14.0-15.0)
|
15.0 (15.0-15.0)
|
15.0 (13.0-15.0)
|
<0.001**
|
Motor score of GCS at ICU discharge [median (IQR)] #
|
6.0 (6.0-6.0)
|
6.0 (6.0-6.0)
|
6.0 (6.0-6.0)
|
<0.001**
|
SOFA score at ICU discharge [median (IQR)]
|
2.0 (2.0-4.0)
|
2.0 (2.0-3.0)
|
2.0 (2.0-4.0)
|
0.002**
|
Respiratory secretions required assistance at ICU discharge, n (%)
|
174 (7.3%)
|
106 (7.5%)
|
68 (7.1%)
|
0.741
|
Fluid balance in the last 48 hours at ICU discharge, ml [median (IQR)]
|
2225.0 (-872.0-7024.0)
|
1951.1 (-562.0-6248.2)
|
2620.0 (-1705.0-8372.6)
|
0.978
|
FiO2 at ICU discharge [median (IQR)]
|
30.0 (25.0-35.0)
|
30.0 (25.0-35.0)
|
30.0 (26.0-35.0)
|
0.788
|
SpO2 at ICU discharge [median (IQR)]
|
97.0 (95.0-99.0)
|
97.0 (95.0-99.0)
|
97.0 (95.0-99.0)
|
0.070
|
pH at ICU discharge [median (IQR)]
|
7.43 (7.39-7.47)
|
7.42 (7.38-7.46)
|
7.44 (7.40-7.47)
|
<0.001**
|
PaO2 at ICU discharge [median (IQR)]
|
88.2 (71.0-128.4)
|
89.8 (72.0-137.3)
|
89.8 (70.3-132.4)
|
0.310
|
PaO2/FiO2 ratio at ICU discharge [median (IQR)]
|
316.0 (248.0-386.0)
|
314.0 (247.0-386.0)
|
322.0 (250.0-392.0)
|
0.449
|
PaCO2, mmHg at ICU discharge [median (IQR)]
|
36.1 (32.0-40.4)
|
36.6 (32.5-40.8)
|
35.2 (31.4-39.7)
|
<0.001**
|
Bicarbonate, mmol/L at ICU discharge [median (IQR)]
|
22.0 (20.0-25.0)
|
23.0 (20.0-25.0)
|
22.0 (20.0-25.0)
|
<0.001**
|
Abbreviations: APACHE - Acute Physiology and Chronic Health Evaluation, BMI – Body mass index, CCI - Charlson Comorbidity Index, COPD- chronic obstructive pulmonary disease, FiO2 – fraction of inspired oxygen, ICU – Intensive care unit, ILD – interstitial lung disease, IQR – Interquartile range, MV – Mechanical ventilation, OSA – obstructive sleep apnea, PaO2 - partial pressure of arterial oxygen, PaCO2 - partial pressure of arterial carbon dioxide, SD – Standard Deviation, SOFA - Sequential Organ Failure Assessment, SpO2 – oxygen saturation, TB - tuberculosis
* p-value<0.05; ** p-value<0.01
# Despite the groups displaying similar median values, the significant difference after a median test suggests that the test evaluates the entire distribution, detecting variations in the shape and spread of the data beyond the central tendencies, leading to the identification of statistically significant distinctions between the groups.
A total of 22 Variables excluding variables with poor outcomes occurrences accounting for less than 10% were used in machine learning analysis. Subsequently, the machine learning feature importance ranking technique identified in Figure 2 found the five most important predictor variables (GCS at the time of ICU discharge, duration of MV, ICU LOS, GCS motor response, and CCI) that significantly enhanced our model's predictive accuracy.
Figure 2: Twenty-two most important variables in the XGBoost model. The features represent each variable's relative importance.
XGBoost had the highest AUROC of 0.693 and demonstrated the best precision and accuracy, followed by Random Forest (AUROC 0.679), EBM (AUROC 0.677), CSM Logistic Regression (AUROC 0.667), and lastly, multilayer perceptron (AUROC 0.646) as shown in (Table 2 and Figure 3). The same three ML models were more accurate than the CSM model in the test cohorts (p-value <0.01). Furthermore, at a specificity of 95%, the XGBoost model had the highest sensitivity at 27.3% and additionally highest accuracy at 70.6%.
Table 2: Area under the receiver operating characteristic curve (AUROC) comparison of different machine learning models in the internal validation report.
Model
|
AUROC
|
Standard Deviation
|
Accuracy
|
Sensitivity when Specificity > 95%
|
XGBoost
|
0.693*
|
0.0042
|
0.706
|
0.273
|
EBM
|
0.677*
|
0.0037
|
0.705
|
0.243
|
RF
|
0.679*
|
0.0076
|
0.708
|
0.252
|
Multilayer Perceptron
|
0.646
|
0.0097
|
0.668
|
0.277
|
Logistic Regression
|
0.667
|
0.0068
|
0.703
|
0.238
|
Abbreviations: AUROC - Area under the receiver operating characteristic curve, EBM - Explainable Boosting Machine, XGBoost - Extreme Gradient Boost, RF - Random Forest.
*p-value<0.05
Figure 3: Comparison of area under the receiver operating characteristic curve (AUROC) curves, overview comparison of the five models.
Partial plots were performed and revealed significant associations between specific important feature factors and the increased incidence of poor composite outcome post-ICU discharge. A cutoff GCS lower than 13 with a motor GCS score of less than 5, a duration of MV longer than 100 hours, an extended ICU LOS of more than 400 hours well and CCI scores of 3 or more exhibited a substantial correlation with higher rates of adverse outcomes (Figure 4).
Figure 4: Partial plot of the effect of (A) Glasgow Coma Score (GCS), (B) Mechanical ventilation in hours, (C) ICU length of stay, (D) Charlson Comorbidity Index (CCI), (E) GCS Motor Score on the risk of poorer outcome post-ICU discharge across different value in the XG boosted machine model.