Objectives Current mortality prediction models used in the intensive care unit (ICU) have a limited role for specific diseases such as influenza, and we aimed to establish an explainable machine learning (ML) model for predicting mortality in critically ill influenza patients using a real-world severe influenza data set. Study design A cross-sectional retrospective multicentre study in Taiwan Setting Eight medical centres in Taiwan. Participants A total of 336 patients requiring ICU-admission for virology-proven influenza at eight hospitals during an influenza epidemic between October 2015 and March 2016. Primary and secondary outcome measures We employed extreme gradient boosting (XGBoost) to establish the prediction model, compared the performance with logistic regression (LR) and random forest (RF), demonstrated the feature importance categorised by clinical domains, and used SHapley Additive exPlanations (SHAP) for visualised interpretation. Results The data set contained 76 features of the 336 patients with severe influenza. The severity was apparently high, as shown by the high Acute Physiology and Chronic Health Evaluation II score (22, 17 to 29) and pneumonia severity index score (118, 88 to 151). XGBoost model (area under the curve (AUC): 0.842; 95% CI 0.749 to 0.928) outperformed RF (AUC: 0.809; 95% CI 0.629 to 0.891) and LR (AUC: 0.701; 95% CI 0.573 to 0.825) for predicting 30-day mortality. To give clinicians an intuitive understanding of feature exploitation, we stratified features by the clinical domain. The cumulative feature importance in the fluid balance domain, ventilation domain, laboratory data domain, demographic and symptom domain, management domain and severity score domain was 0.253, 0.113, 0.177, 0.140, 0.152 and 0.165, respectively. We further used SHAP plots to illustrate associations between features and 30-day mortality in critically ill influenza patients. Conclusions We used a real-world data set and applied an ML approach, mainly XGBoost, to establish a practical and explainable mortality prediction model in critically ill influenza patients.

High glycemic variability (GV) is common in critically ill patients; however, the prevalence and mortality association with early GV in patients with sepsis remains unclear. This retrospective cohort study was conducted in a medical intensive care unit (ICU) in central Taiwan. Patients in the ICU with sepsis between January 2014 and December 2015 were included for analysis. All of these patients received protocol-based management, including blood sugar monitoring every 2 h for the first 24 h of ICU admission. Mean amplitude of glycemic excursions (MAGE) and coefficient of variation (CoV) were used to assess GV. A total of 452 patients (mean age 71.4 ± 14.7 years; 76.7% men) were enrolled for analysis. They were divided into high GV (43.4%, 196/452) and low GV (56.6%, 256/512) groups using MAGE 65 mg/dL as the cut-off point. Patients with high GV tended to have higher HbA1c (6.7 ± 1.8% vs. 5.9 ± 0.9%, p < 0.01) and were more likely to have diabetes mellitus (DM) (50.0% vs. 23.4%, p < 0.01) compared with those in the low GV group. Kaplan–Meier analysis showed that a high GV was associated with increased 30-day mortality (log-rank test, p = 0.018). The association remained strong in the non-DM (log-rank test, p = 0.035), but not in the DM (log-rank test, p = 0.254) group. Multivariate Cox proportional hazard regression analysis identified that high APACHE II score (adjusted hazard ratio (aHR) 1.045, 95% confidence interval (CI) 1.013–1.078), high serum lactate level at 0 h (aHR 1.009, 95% CI 1.003–1.014), having chronic airway disease (aHR 0.478, 95% CI 0.302–0.756), high mean day 1 glucose (aHR 1.008, 95% CI 1.000–1.016), and high MAGE (aHR 1.607, 95% CI 1.008–2.563) were independently associated with increased 30-day mortality. The association with 30-day mortality remained consistent when using CoV to assess GV. We found that approximately 40% of the septic patients had a high early GV, defined as MAGE > 65 mg/dL. Higher GV within 24 h of ICU admission was independently associated with increased 30-day mortality. These findings highlight the need to monitor GV in septic patients early during an ICU admission.

Objective: The number of patients requiring prolonged mechanical ventilation (PMV) is increasing worldwide, but the weaning outcome prediction model in these patients is still lacking. We hence aimed to develop an explainable machine learning (ML) model to predict successful weaning in patients requiring PMV using a real-world dataset.Methods: This retrospective study used the electronic medical records of patients admitted to a 12-bed respiratory care center in central Taiwan between 2013 and 2018. We used three ML models, namely, extreme gradient boosting (XGBoost), random forest (RF), and logistic regression (LR), to establish the prediction model. We further illustrated the feature importance categorized by clinical domains and provided visualized interpretation by using SHapley Additive exPlanations (SHAP) as well as local interpretable model-agnostic explanations (LIME).Results: The dataset contained data of 963 patients requiring PMV, and 56.0% (539/963) of them were successfully weaned from mechanical ventilation. The XGBoost model (area under the curve [AUC]: 0.908; 95% confidence interval [CI] 0.864–0.943) and RF model (AUC: 0.888; 95% CI 0.844–0.934) outperformed the LR model (AUC: 0.762; 95% CI 0.687–0.830) in predicting successful weaning in patients requiring PMV. To give the physician an intuitive understanding of the model, we stratified the feature importance by clinical domains. The cumulative feature importance in the ventilation domain, fluid domain, physiology domain, and laboratory data domain was 0.310, 0.201, 0.265, and 0.182, respectively. We further used the SHAP plot and partial dependence plot to illustrate associations between features and the weaning outcome at the feature level. Moreover, we used LIME plots to illustrate the prediction model at the individual level. Additionally, we addressed the weekly performance of the three ML models and found that the accuracy of XGBoost/RF was ~0.7 between weeks 4 and week 7 and slightly declined to 0.6 on weeks 8 and 9.Conclusion: We used an ML approach, mainly XGBoost, SHAP plot, and LIME plot to establish an explainable weaning prediction ML model in patients requiring PMV. We believe these approaches should largely mitigate the concern of the black-box issue of artificial intelligence, and future studies are warranted for the landing of the proposed model.

Background: Machine learning (ML) model is increasingly used to predict short-term outcome in critically ill patients, but the study for long-term outcome is sparse. We used explainable ML approach to establish 30-day, 90-day and 1-year mortality prediction model in critically ill ventilated patients. Methods: We retrospectively included patients who were admitted to intensive care units during 2015–2018 at a tertiary hospital in central Taiwan and linked with the Taiwanese nationwide death registration data. Three ML models, including extreme gradient boosting (XGBoost), random forest (RF) and logistic regression (LR), were used to establish mortality prediction model. Furthermore, we used feature importance, Shapley Additive exPlanations (SHAP) plot, partial dependence plot (PDP), and local interpretable model-agnostic explanations (LIME) to explain the established model. Results: We enrolled 6994 patients and found the accuracy was similar among the three ML models, and the area under the curve value of using XGBoost to predict 30-day, 90-day and 1-year mortality were 0.858, 0.839 and 0.816, respectively. The calibration curve and decision curve analysis further demonstrated accuracy and applicability of models. SHAP summary plot and PDP plot illustrated the discriminative point of APACHE (acute physiology and chronic health exam) II score, haemoglobin and albumin to predict 1-year mortality. The application of LIME and SHAP force plots quantified the probability of 1-year mortality and algorithm of key features at individual patient level. Conclusions: We used an explainable ML approach, mainly XGBoost, SHAP and LIME plots to establish an explainable 1-year mortality prediction ML model in critically ill ventilated patients.

Objective: Pain assessment based on facial expressions is an essential issue in critically ill patients, but an automated assessment tool is still lacking. We conducted this prospective study to establish the deep learning-based pain classifier based on facial expressions.Methods: We enrolled critically ill patients during 2020–2021 at a tertiary hospital in central Taiwan and recorded video clips with labeled pain scores based on facial expressions, such as relaxed (0), tense (1), and grimacing (2). We established both image- and video-based pain classifiers through using convolutional neural network (CNN) models, such as Resnet34, VGG16, and InceptionV1 and bidirectional long short-term memory networks (BiLSTM). The performance of classifiers in the test dataset was determined by accuracy, sensitivity, and F1-score.Results: A total of 63 participants with 746 video clips were eligible for analysis. The accuracy of using Resnet34 in the polychromous image-based classifier for pain scores 0, 1, 2 was merely 0.5589, and the accuracy of dichotomous pain classifiers between 0 vs. 1/2 and 0 vs. 2 were 0.7668 and 0.8593, respectively. Similar accuracy of image-based pain classifier was found using VGG16 and InceptionV1. The accuracy of the video-based pain classifier to classify 0 vs. 1/2 and 0 vs. 2 was approximately 0.81 and 0.88, respectively. We further tested the performance of established classifiers without reference, mimicking clinical scenarios with a new patient, and found the performance remained high.Conclusions: The present study demonstrates the practical application of deep learning-based automated pain assessment in critically ill patients, and more studies are warranted to validate our findings.

Although energy intake might be associated with clinical outcomes in critically ill patients, it remains unclear whether full or trophic feeding is suitable for critically ill patients with high or low nutrition risk. We conducted a prospective study to determine which feeding energy intakes were associated with clinical outcomes in critically ill patients with high or low nutrition risk. This was an investigator-initiated, single center, single blind, randomized controlled trial. Critically ill patients were allocated to either high or low nutrition risk based on their Nutrition Risk in the Critically Ill score, and then randomized to receive either the full or the trophic feeding. The feeding procedure was administered for six days. No significant differences were observed in hospital, 14-day and 28-day mortalities, the length of ventilator dependency, or ICU and hospital stay among the four groups. There were no associations between energy and protein intakes and hospital, 14-day and 28-day mortalities in any of the four groups. However, protein intake was positively associated with the length of hospital stay and ventilator dependency in patients with low nutrition risk receiving trophic feeding. Full or trophic feeding in critically ill patients showed no associations with clinical outcomes, regardless of nutrition risk.