Prediction of ICU length of stay, hospital discharge outcomes and discharge location among ICU-admitted patients diagnosed with viral hepatitis using machine learning: a retrospective cohort study of the MIMIC-IV database

Abstract

Background

Hepatitis, a disease characterised by inflammation of the liver, is a leading global health challenge that contributes to over 1.3 million deaths annually, with hepatitis B and C accounting for many of these fatalities. Intensive care unit (ICU) management of patients is particularly challenging due to the complex clinical care and resource demands. Despite advancements in ICU predictive analytics, limited research has specifically addressed hepatitis patients, creating a gap in optimising care for this population.

Methods

This study focuses on predicting ICU length of stay (LoS), hospital discharge outcomes and discharge location for ICU-admitted viral hepatitis patients using a comparative assessment of machine learning (ML) models. Leveraging data from the Medical Information Mart for Intensive Care-IV database, which includes around 94 500 ICU patient records, this study uses sociodemographic details, clinical characteristics and resource utilisation metrics to develop predictive models such as Random Forest, Logistic Regression, Gradient Boosting Machines and Generalised Additive Model with Negative Binomial Regression.

Results

Among 3875 ICU-admitted hepatitis patients, Random Forest classification outperformed Logistic Regression in predicting discharge outcomes, achieving higher accuracy (0.87 vs 0.82) and greater discriminative ability (area under the receiver operating characteristic curve 0.95 vs 0.89). For ICU LoS prediction, Random Forest regression applied to log-transformed LoS demonstrated strong performance (R² up to 0.82), while the generalised additive model with negative binomial distribution explained approximately 76% of LoS variance. Prediction of discharge location yielded moderate performance across Gradient Boosting and multinomial logistic regression models (accuracy 0.55 and 0.56), reflecting challenges associated with multi-class imbalance. Variable importance analyses across ML models consistently identified medication counts, procedure counts, comorbidity burden, age, race and total LoS as the most influential predictors of discharge outcomes and discharge location.

Conclusions

This study demonstrates the value of ML models for predicting clinical outcomes for hepatitis patients, including ICU LOS and hospital discharge status. The results underscore the influence of factors like race and age, revealing disparities that must be addressed in predictive care strategies. While the models show promise, challenges such as variability in prolonged stays and limited multiclass prediction accuracy point to the need for ongoing refinement and research.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This study used a large, de-identified critical care dataset (Medical Information Mart for Intensive Care, MIMIC-IV) with detailed sociodemographic, clinical and resource utilisation variables suitable for predictive modelling.

• Multiple statistical and machine learning models were implemented and evaluated using a consistent preprocessing, imputation, resampling and validation framework.

• The single-centre nature of the MIMIC-IV database may limit the generalisability of the modelling approach to institutions with different patient populations and clinical practices.

• Diagnosis codes in MIMIC-IV are assigned at the hospital admission level and are not temporally linked to intensive care unit (ICU) admission, limiting precise attribution of ICU admission indications.

• Outcome class imbalance, particularly for infrequent discharge locations, may have affected model stability and performance despite the application of resampling techniques.

Introduction

Hepatitis, an inflammatory liver disease, remains a significant global health challenge, claiming an increasing number of lives each year. According to the WHO 2024 Global Hepatitis Report, viral hepatitis is the second leading infectious cause of death globally, responsible for 1.3 million deaths annually.¹ This number rose from 1.1 million in 2019, with 83% of these deaths attributed to hepatitis B and 17% to hepatitis C.¹ Every day, approximately 3500 people die due to hepatitis B and C infections worldwide.² Despite advancements in diagnostic tools and treatment options, testing and treatment coverage rates have plateaued, signaling a growing public health crisis.¹ In the USA, the most common forms of viral hepatitis are hepatitis A, B, and C, each impacting the liver differently and predominantly affecting distinct populations.^{3 4} Hepatitis B and C account for the majority of the chronic viral hepatitis disease burden,⁵ affecting an estimated 640,000 and 2.4 million individuals, respectively. Marked demographic disparities persist, with higher rates of chronic hepatitis B observed among non-Hispanic Asian and Pacific Islander populations and adults aged 30–49 years,⁶ while hepatitis C is disproportionately observed among non-Hispanic American Indian and Alaska Native populations.⁷ In Massachusetts (for context of the Medical Information Mart for Intensive Care (MIMIC) database), viral hepatitis (particularly hepatitis C virus (HCV)) remains a significant public health concern, with 2580 reported cases of chronic HCV in 2023 and an estimated 1100–2000 cases of chronic hepatitis B reported annually.⁸ Consistent with this substantial disease burden, hepatitis B and C pose severe health risks, often leading to chronic conditions such as cirrhosis and liver cancer, and are the primary contributors to liver-related mortality globally.⁹ Managing an intensive care unit (ICU) for admitted patients with hepatitis is particularly challenging due to the treatment complexities and the resource-intensive nature of care required for severe cases.

One critical issue in ICU settings is the unpredictability of a patient’s length of stay (LoS), a key metric influencing hospital resource management, patient care quality, and healthcare efficiency. Prolonged ICU stays are closely linked to increased hospital costs and heightened resource strain.¹⁰ Research also indicates that longer ICU stays correlate with increased long-term mortality rates, underscoring the need for precise LoS predictions.¹¹ Beyond LoS, discharge outcomes and locations serve as vital metrics for assessing patient recovery, readmission risks, and the burden on healthcare systems.12,15 Sociodemographic and clinical factors such as race, gender, marital status, insurance type, age, and type of hepatitis could significantly influence hospitalization outcomes. Prior ICU studies support this relationship, demonstrating that sociodemographic factors are associated with disparities in LoS, discharge outcomes, and mortality, underscoring the importance of accounting for health inequities in predictive modeling.16,20 Recent years have seen the growing application of machine learning (ML) in healthcare, particularly predictive modeling. ML algorithms provide several advantages over traditional statistical approaches, including their ability to analyze large, complex datasets and identify subtle patterns that conventional techniques may overlook.²¹ Predictive models using ML have been successfully applied in ICU settings to improve forecasts related to patient outcomes, treatment responses, and resource needs.^{22 23} However, there is a noticeable lack of ML models tailored to viral hepatitis patients and their unique clinical profiles.

This study explores and compares ML models to predict the LoS, discharge location and discharge outcomes for ICU-admitted hepatitis patients. These models will incorporate sociodemographic and clinical variables from ICU admission records as key predictors. By addressing this gap, we aim to provide valuable insights for healthcare providers, enhancing resource planning and improving hepatitis patient outcomes in ICU settings.

Materials and methods

Data

The data set for this study was sourced from MIMIC-IV (Medical Information Mart for Intensive Care), version 3.0, a comprehensive and de-identified repository of patient records,24,26 spanning the years 2008–2022. The MIMIC-IV comprises data collected from patients admitted to the emergency department or ICU at the Beth Israel Deaconess Medical Center in Boston, Massachusetts. It includes over 364,000 unique patient records, with a total of 546,028 hospital admissions and nearly 94,500 ICU stays, providing a rich source of information for developing predictive models in healthcare. MIMIC-IV is organized into two main data modules: hosp and icu. The hosp module captures information from the hospital-wide electronic health records, detailing hospitalizations, patient demographics, laboratory results, medication administration, billing data, and more. In contrast, the icu module contains highly detailed clinical data from the ICU, sourced from the MetaVision clinical information system, including treatment plans and monitoring data.24,26 This study was completed prior to the release of MIMIC-IV v3.1. We carefully considered all release notes and determined that data extracted from v3.0 was appropriate for this analysis.

Variables

This study includes sociodemographic, clinical, and resource utilization data. All predictor variables were obtained from structured fields recorded at or prior to ICU admission or accrued during the ICU stay. Sociodemographic variables cover gender, age, race, marital status, and type of insurance. Admission characteristics include the type of admission and admission location. Clinical variables included the number of comorbidities as well as indicators for each hepatitis type. Resource utilization variables include counts of medications, procedures and drugs dispensed. Additional ICU-specific measures capture the volume of fluids prescribed and the number of procedures performed. The three primary clinical outcomes: Total LoS, measured as the total days spent in the ICU; Discharge Location, detailing the destination on discharge from hospital, such as home, hospice or skilled nursing; and Discharge Outcome, a binary variable indicating discharge status (dead/alive).

Patient and public involvement

The study was conducted under the data use agreement governing MIMIC-IV, ensuring compliance with ethical standards for research involving human subjects. Patients or the public were not involved in the design, conduct, reporting or dissemination plans of this research work.

Data processing and preparation

Figure 1 illustrates the data preprocessing workflow for the study, detailing sequential steps taken to create the final dataset, completed using the data management system MySQL.^{27 28} To isolate hepatitis-related hospital admissions, we used the International Classification of Diseases codes specific to viral hepatitis from the hospital billing information and linked them to admission records. These were merged with demographic and length-of-stay data to create a hepatitis-specific cohort. After adding pre-ICU and post-ICU medication and procedure counts and comorbidities counts, the combined dataset initially consisted of 4409 patient records (figure 1). After a data validation process, 534 records showed inconsistency between discharge outcomes and discharge locations, reducing the final dataset to 3875 hepatitis patient records. Data validation and record exclusion procedures were applied uniformly across the study cohort. The final analytic sample was divided into training and testing sets for model development and evaluation.

Figure 1. Data preprocessing workflow for extracting and refining hepatitis patient records from the MIMIC-IV database. ICU, intensive care unit; LoS, length of stay; MIMIC-IV, Medical Information Mart for Intensive Care-IV.

Missing data were handled via multiple imputation by chained equations as appropriate.²⁹ Across study variables, the proportion of missing data ranged from 0.03% to 16.6%. Five imputed datasets were generated over 20 iterations to ensure algorithm convergence. Predictive mean matching was used for numeric variables, while polytomous logistic regression (LR) was applied to categorical variables with more than two levels. Imputation was conducted under the assumption that data were missing at random, supported by the availability of extensive observed covariates related to patient demographics, admission characteristics and clinical severity, which are known to influence documentation patterns in electronic health record data. The discharge outcome variable displayed a significant class imbalance, with 3353 records for the alive category and 522 for death. To address this, the Synthetic Minority Oversampling Technique (SMOTE) was used to balance the classes.³⁰ Finally, an 80:20 split was used to separate the data into training and testing sets. The training subset was used for model development, while the testing subset was reserved for model performance evaluation. This data partitioning strategy was applied consistently across all models to allow comparative assessment of predictive performance. The training and testing datasets were derived from the same source population and healthcare setting, with identical eligibility criteria, predictor definitions and outcome definitions; the only difference between the datasets was the random partitioning used for model development and evaluation. Given the random partitioning from a common source population, no formal comparison of predictor or outcome distributions between the training and testing subsets was performed.

Exploratory univariate analyses

Exploratory univariate analyses were performed to examine associations between predictors and each outcome using Wilcoxon rank-sum tests for binary outcomes, chi-square tests for categorical predictors, Kruskal-Wallis tests for multiclass outcomes and Spearman correlation for LoS. These analyses were used to characterise relationships and confirm face validity of candidate predictors; no variables were excluded solely based on univariate testing and all clinically relevant covariates were retained for model development.

ML models

For predicting the discharge outcome, a LR and Random Forest Classification (RFC) models were used. Discharge location was not considered as a predictor in these models because certain discharge location categories overlapped with death records, providing redundant information. LR is a statistical model used for binary classification, estimating the probability of one of two possible outcomes (in this case alive or dead) based on a set of predictor variables.³¹ LR was chosen for its interpretability, providing clear insights into how each clinical and sociodemographic feature influences discharge outcome. On the other hand, RFC is applied to enhance the accuracy of predicting discharge outcome. The model constructs multiple decision trees using random data samples. The final prediction is determined by the majority vote across all trees. This ensemble approach captures complex, non-linear relationships among variables, offering strong predictive performance and minimising the risk of overfitting.³² Although it does not offer the interpretive insights of LR, RFC’s capability to handle heterogeneous data makes it well-suited for maximising predictive power in clinical scenarios.³³

To predict LoS, we considered two scenarios. In the first case, we considered LoS as a discrete (count) variable,³⁴ using generalised additive model (GAM) with a negative binomial (NB) distribution, denoted by GAM-NB, and as a continuous variable³⁵ in our second case via Random Forest Regression (RFR). For the discrete case, GAM-NB was used to address overdispersion (variance exceeds the mean).³⁶ By incorporating smooth functions, the GAM framework allows for modelling potential non-linear relationships between predictors and LoS.³⁷ RFR, as another ensemble method, constructs multiple decision trees and averages their predictions to capture non-linear relationships and interactions among predictors.³²

For discharge location, a multiclass outcome (>2 levels), we compared a gradient boosting model (GBM) and a multinomial LR (MLR). GBM is an ensemble learning technique that builds multiple decision trees sequentially, with each new tree correcting errors made by previous ones. This model focuses on reducing bias and improving predictive performance through boosting. GBM was chosen for its strong ability to handle complex, non-linear relationships and provide high accuracy.³⁸ On the other hand, the MLR is an extension of LR used when the dependent variable has more than two categories.³⁹ It models the probability of each class as a function of the predictors. This model was included to establish baseline, offering a simpler and interpretable model for understanding the impact of predictor variables on different discharge locations. Predictors were entered into the models using their original measurement scales without normalisation or standardisation. Online supplemental table 1 summarises the models for each clinical outcome.

Model evaluation and resample

The models were evaluated using appropriate performance metrics⁴⁰ based on the type of outcome variable. For LoS, regression models were evaluated using root mean squared error (RMSE), R-squared (R²) and mean absolute error (MAE). The multiclass classification models were assessed using accuracy, the Kappa statistic and the Brier score. For the models based on binary outcome, we used Brier score, accuracy, Kappa, area under the receiver operating characteristic curve (ROC AUC), sensitivity and specificity. Definitions for each of these performance metrics are given in the online supplemental material. To ensure robustness and generalisability, a 10-fold cross-validation was employed,⁴¹ as it provides a balance between computational efficiency and reliability.^{41 42} K-fold cross-validation mitigates the risk of overfitting, provides more reliable estimates of the model’s performance, and enhances the reliability and generalisability of the results, making it a robust validation strategy for ML applications.⁴² Predictions were generated by applying the fitted models to the testing dataset, yielding predicted probabilities for classification outcomes and predicted values for regression outcomes. For classification tasks, predicted probabilities were converted to class labels using a default probability threshold of 0.5, with no outcome-specific threshold optimisation performed.

To compare predictive performance between models, statistical comparisons were conducted between LR and RFC. Performance metrics obtained from identical cross-validation folds were paired between models. Differences in model performance were assessed using paired t-tests across resampling iterations. P values were adjusted for multiple comparisons using the Holm method, and statistical significance was evaluated using a two-sided α level of 0.05.

Software

Data analysis was conducted in R statistical software^{43 44} using various R packages to streamline data manipulation, model development and predictions, and visualisation.3045,53 MySQL²⁷ was used to extract and manage data from the MIMIC-IV database. Statistical significance was assessed at the 5% significance level.

Results

Demographic and descriptive statistics

The study included a total of 3875 ICU-admitted patients with hepatitis-related conditions. The gender distribution was predominantly male, accounting for 68.2% (n=2644) of the sample, while females constituted 31.8% (n=1231). The average age of the participants was 53.26 years (SD=12.77). The racial composition was primarily White (57.3%, n=2220), followed by Black (16.9%, n=653), Hispanic/Latino (6.3%, n=245) and Asian (6.0%, n=234). Other racial groups, such as Indian/Alaska Native and Portuguese, made up a smaller proportion of the population, each representing less than 1% of the sample. These demographic characteristics, along with details on marital status, insurance type, clinical and admission characteristics, are summarised in table 1. The distribution of hepatitis types is shown in figure 2, which reveals that Hepatitis C was the most prevalent condition, affecting 82.2% (n=3186) of the total sample. Hepatitis B was the second most common condition at 19.3% (n=746), while hepatitis A (1.0%, n=37), hepatitis D (9.2%, n=357) and hepatitis E (<1%) were less frequent. Additionally, 8.0% (n=311) of patients experienced hepatic coma. Most of the study participants were either single (48.3%, n=1873) or married (28.5%, n=1106). When looking at insurance type, a similar proportion of participants had either Medicaid (39.8%, n=1542) or Medicare (38.0%, n=1472). As expected, more than 50% of patients were admitted through the emergency ward (59.3%, n=2296). The median number of comorbidities found was 18 (IQR: 12–24), while the number of medications (administered and dispensed) and procedures performed pre-ICU and during ICU ranged (IQR) from 2 (pre-ICU procedures) to a high of 154 (ICU medications). The distribution of hepatitis types across racial groups is depicted in figure 3. Hepatitis C was the most prevalent condition across all racial groups, with White patients contributing the largest proportion. Among Black and Hispanic/Latino patients, hepatitis C was also the most common condition but showed slightly smaller proportions compared with White patients, while hepatitis B was notably more prevalent among Asian patients compared with other racial groups. In addition, hepatitis B disproportionately affects non-White individuals compared with Whites.

Table 1. Baseline demographic, clinical and admission characteristics of study participants.

Variable	N=3875
Gender=Male (%)	2644 (68.2)
Age (mean (SD))	53.26 (12.77)
Race (%)
Indian/Alaska Native	10 (0.3)
Asian	234 (6.0)
Black	653 (16.9)
Hispanic/Latino	245 (6.3)
Other	120 (3.1)
Portuguese	29 (0.7)
Unknown	364 (9.4)
White	2220 (57.3)
Marital status (%)
Divorced	398 (10.3)
Married	1106 (28.5)
Single	1873 (48.3)
Widowed	202 (5.2)
NA	296 (7.6)
Insurance (%)
Medicaid	1542 (39.8)
Medicare	1472 (38.0)
Other	123 (3.2)
Private	674 (17.4)
NA	296 (7.6)
Type of admission (%)
Direct emergency	230 (5.9)
Direct observation	10 (0.3)
Elective	50 (1.3)
Emergency unit observation	5 (0.1)
Emergency ward	2296 (59.3)
Observation admit	548 (14.1)
Surgical same day admission	155 (4.0)
Urgent	581 (15.0)
Number of comorbidities (median (IQR))	18.00 (12.00, 24.00)
Pre-ICU medications count (median (IQR))	82.00 (36.00, 140.25)
Pre-ICU procedures count (median (IQR))	3.00 (2.00, 6.00)
Dispensed medications count (median (IQR))	65.50 (33.00, 118.50)
ICU medications count (median (IQR))	54.00 (20.00, 154.00)
ICU procedures count (median (IQR))	7.00 (3.00, 14.00)

ICU, intensive care unit; NA, not available.

Figure 2. Distribution of hepatitis types among study participants.

Figure 3. Distribution of hepatitis types across racial groups.

The distribution of LoS for hepatitis patients was highly skewed as commonly seen in hospital records, with most patients having a short hospital stay as shown in online supplemental figure 1. The median total LoS was 2 days (IQR: 1–4), with some patients having an extended stay of over 30 days. Variations in LoS were also observed based on admission location (online supplemental figure 2). Patients admitted from emergency admissions had shorter median stays compared with those transferred from other facilities. This figure highlights how admission pathways may influence the duration of ICU stays. Online supplemental figure 3 illustrates the relationship between the total LoS and discharge outcome. Patients with longer LoS were more likely to have died, as indicated by the higher median and wider range of LoS in the death group compared with the alive group. Online supplemental figure 4 on the other hand, highlights the distribution of discharge locations among patients. The majority (51.5%, n=1997) were discharged to home-based care. Skilled nursing or long-term care facilities accounted for 17.7% (n=684). Other notable discharge destinations included rehabilitation centres, hospice care, psychiatric care, other healthcare facilities, and patients leaving the hospital against medical advice, each contributing less than 5% of the sample.

Model results and performance

Discharge outcome (dead/alive)

All models were developed using an analytic cohort of 3875 ICU-admitted hepatitis patients. For binary discharge outcome models, the original dataset included 522 death events and 3353 alive events. To address class imbalance, SMOTE was applied. The LR and RFC models were developed and evaluated to predict discharge outcomes. A complete set of performance metrics for both models are presented in online supplemental table 2. On the test data, LR yielded a Brier Score of 0.1302, an accuracy of 0.818, a Kappa of 0.622, and an ROC AUC of 0.887. Sensitivity and specificity were 0.73 and 0.87, respectively. The RFC model demonstrated a lower Brier score of 0.0875, an accuracy of 0.87, a Kappa of 0.742 and a higher ROC AUC of 0.95. Sensitivity and specificity for the Random Forest model were 0.78 and 0.94, respectively. Cross-validation results are also shown in online supplemental table 2, indicating the mean and SD for each metric across repeated samples. The mean Brier score for RFC was 0.089 (±0.010), while LR had a mean Brier score of 0.129 (±0.011). The RFC model had a mean accuracy of 0.87 (±0.017) and a mean Kappa of 0.729 (±0.037), compared with LR’s mean accuracy of 0.82 (±0.02) and mean Kappa of 0.629 (±0.043).

To evaluate the robustness of the discharge outcome models, two complementary sensitivity analyses were conducted addressing (1) class imbalance handling and (2) alternative representations of comorbidity burden.

First, to assess the impact of class imbalance correction, both LR and RFC models were evaluated without applying SMOTE. In the absence of class balancing, both models demonstrated high specificity but markedly reduced sensitivity for the minority outcome (death), reflecting the underlying imbalance in outcome distribution. Despite this reduction in sensitivity, overall discrimination remained moderate (ROC AUC: 0.74 for LR and 0.79 for RFC), indicating that predictive signal persisted. These findings support the use of SMOTE in the primary analyses to enable clinically meaningful detection of in-hospital mortality.

Second, the comorbidity burden was alternatively summarised using the Charlson Comorbidity Index (CCI) in place of the unweighted comorbidity count. Under this specification, both models retained moderate discriminatory ability on the test set; however, performance metrics differed from the primary analysis. Specifically, accuracy was lower (LR=0.76; RF=0.79), sensitivity was reduced (LR=0.55; RF=0.50) and agreement beyond chance was modest, as reflected by Kappa values (LR=0.25; RF=0.28). These results indicate that predictive performance varies depending on how comorbidity burden is represented, and the CCI-based models are therefore presented as a robustness check rather than replacements for the primary specification.

Returning to the primary SMOTE-based analysis, statistical comparisons were conducted to evaluate differences in performance metrics between the LR and RFC models. The results indicated statistically significant differences (p<0.001) for all key metrics, including Brier score, accuracy, Kappa, ROC AUC, sensitivity and specificity. The RFC model consistently outperformed LR across all evaluated criteria for predicting discharge outcomes for the ICU-admitted hepatitis patients. Figures 4 and 5 display the variable importance for the LR and RFC models, respectively. The top predictors in the LR model (figure 4) were pre-ICU medications count, ICU medications count and ICU procedures count, with sociodemographic factors such as race and age also ranking among the top 10 variables. For the RFC model (figure 5), ICU medications count, ICU procedures count and total LoS were identified as the most influential variables, with race and age again appearing among the top 10 predictors. (online supplemental table 3) summarises the confusion matrix results for both models. Both models showed good sensitivity and specificity values, with RFC performing better with higher precision.

Figure 4. Variable importance for logistic regression for discharge outcome. ICU, intensive care unit.

Figure 5. Variable importance for random forest classification for discharge outcome. ICU, intensive care unit.

Total LoS

The GAM-NB and RFR models were used to predict total LoS as discrete and continuous outcome. The mean LoS in the dataset was 3.69, with a variance of 42.54, indicating overdispersion. On the test data, the GAM-NB model achieved an RMSE of 2.9619, an MAE of 1.4237 and an R² of 0.7594, demonstrating its ability to explain 75.94% of the variance (online supplemental table 4). The cross-validated estimates are RMSE of 5.5212, mean R² of 0.7601 and mean MAE of 1.5634. As shown in online supplemental figure 5, most predictions are closely aligned with actual values; deviations are observed at higher LoS, indicating potential challenges in predicting extended stays. The RFR model, applied after log-transforming the total LoS, resulted in improved performance, achieving an RMSE of 0.3144, an R² of 0.821, and an MAE of 0.2377 on the test data. Cross-validation further confirmed the model’s consistency, with mean RMSE of 0.295 (±0.020), mean R² of 0.838 (±0.021) and mean MAE of 0.229 (±0.014). However, when done on the original scale, the performance metrics were: RMSE of 3.271, R² of 0.755 and MAE of 1.360 (online supplemental table 5).

Figure 6 illustrates the variable importance of the RFR model for predicting LoS. The most influential predictors were the number of ICU medications and procedures, comorbidities count and pre-ICU procedure count. While clinical factors were most prominent, sociodemographic variables such as age and race also emerged as important predictors. Online supplemental figure 6 illustrates the relationship between predicted and actual LoS values for the RFR model. The plot on the left presents the scatter plot for the log-transformed data, where a stronger linear relationship is observed, with predictions closely aligning with actual values. In contrast, the plot on the right displays greater dispersion for higher values.

Figure 6. Variable importance for random forest regression model for total LoS. ICU, intensive care unit; LoS, length of stay.

Discharge location

The GBM was evaluated on the test data, achieving a Brier score of 0.589, an accuracy of 0.558, a Kappa of 0.211 and an ROC AUC of 0.7407 (online supplemental table 6). Cross-validation analysis resulted in mean Brier score of 0.595 (±0.0206), mean accuracy of 0.559 (±0.227), mean Kappa of 0.2077 (±0.0391) and mean ROC AUC of 0.7319 (±0.0282), indicating stable performance across folds. The MLR produced an accuracy of 0.5639 on the test data, with a Kappa value of 0.2037 and Brier score of 0.5896 (online supplemental table 6). Cross-validation yielded consistent results: mean accuracy of 0.5658 (±0.15), mean Kappa of 0.2335 (±0.30), mean Brier score of 0.6 (±0.0352), suggesting comparable predictive performance (online supplemental table 6). Significant predictors identified in this model included gender, marital status, insurance type, type of admission, number of comorbidities and number of ICU procedures and medications. Figure 7 illustrates the variable importance of GBM, highlighting the number of ICU medications, comorbidities, age, total LoS and race as the most influential predictors.

Figure 7. Variable importance for gradient boost model for discharge location. ICU, intensive care unit.

Discussion

This study applied a range of statistical and ML approaches to model and predict ICU LoS, discharge outcomes and discharge location among critically ill patients diagnosed with viral hepatitis. Using routinely collected electronic health record data from the MIMIC-IV database, the models consistently identified measures of treatment intensity, such as medication counts, procedure counts and comorbidity burden, as influential predictors across outcomes. These findings suggest that resource utilisation variables capture important aspects of illness severity and care complexity that extend beyond diagnostic classification alone. Among sociodemographic factors, we found age and race to be important predictors. Notably, race consistently emerged as one of the top predictors across all models, underscoring its critical role in influencing health outcomes for hepatitis patients. Race should not be interpreted as a biological determinant of outcomes, but rather as a proxy for underlying structural and social determinants of health, including differences in access to care, baseline disease severity and healthcare utilisation patterns. This finding is particularly significant as it aligns with observed racial disparities in the raw data, suggesting systemic inequities. Additionally, Hepatitis C emerged as the most prevalent condition in our cohort, aligning with global epidemiological trends that underscore its significant burden on critical care systems.^{54 55}

While predicting LoS, random forest regression demonstrated advantages in capturing complex, non-linear relationships between clinical characteristics and ICU duration. The use of log transformation techniques helped mitigate the influence of extreme values and improved the model’s ability to characterise variability in LoS. In contrast, semiparametric approaches offered strengths in addressing distributional challenges such as overdispersion but were less flexible in modelling complex interactions. These findings suggest that model selection should be guided by analytic goals, with ensemble methods better suited for prediction and interpretable models offering value for inference. Key drivers of ICU LoS included medication burden, procedural intensity, comorbidity burden, pre-ICU interventions and patient age, reflecting underlying disease severity and care complexity. These findings align with prior studies,⁵⁶ which demonstrated that higher procedural and pharmacological interventions often correlate with disease severity and resource intensity, leading to longer ICU stays. Similarly, another study emphasised that prolonged ICU stays are associated with complications, readmissions and sedative use.⁵⁷ Although the models performed well for shorter stays, predicting longer LoS posed challenges due to increased variability and extreme values, as reflected in online supplemental figures 5 and 6. In another study,³⁵ the authors observed similar challenges, reporting that ML models often struggle to predict longer stays due to the inherent variability of extended ICU durations. Our study reflects these findings, with greater error dispersion observed in the upper range of LoS predictions. The performance of RFR in predicting LoS aligns with a prior study,⁵⁸ which reported high predictive accuracy for ICU stays using RFR. Ensemble methods such as RFR excelled in capturing nonlinear interactions and handling heterogeneous predictors,⁵⁹ achieving 65% accuracy for LoS predictions in ICU datasets. From an operational perspective, improved understanding of the drivers of ICU LoS may support more effective capacity planning and allocation of critical care resources for patients with advanced liver disease.

Discharge outcome prediction benefited from the use of ensemble-based classifiers, which were better able to accommodate complex interactions among clinical severity, treatment intensity and healthcare utilisation compared with traditional regression approaches. Predictors reflecting cumulative treatment burden, including medication use, procedural intensity and prolonged ICU exposure, were strongly associated with mortality risk, underscoring the close relationship between disease severity and adverse outcomes. These findings are consistent with prior studies demonstrating that extended ICU LoS is associated with increased mortality, particularly among patients with complex and resource-intensive clinical courses.^{11 57 60} Sociodemographic factors also contributed meaningfully to discharge outcomes, with race emerging as an important predictor, in line with existing evidence of disparities in ICU mortality across demographic groups.⁶¹ From a clinical perspective, early identification of patients at elevated mortality risk may support proactive decision-making, including closer monitoring, timely escalation of care and informed discussions regarding goals of care.

The prediction of discharge location presented unique challenges, primarily due to the small number of records and large number of categories (total of 8). GBM and MLR yielded moderate performance for this multicategorical outcome, with accuracies of 55.87% and 56.39%, respectively. These results highlight the complexities of multi-class prediction in critical care settings with limited data. SMOTE was applied to address class imbalance but resulted in data loss, which limited the model’s performance. Similar challenges were addressed using hierarchical classifiers and SMOTE to manage imbalanced multiclass datasets, suggesting that advanced techniques and larger datasets are critical for improving predictive accuracy.⁶² Additionally, the authors noted that discharge decisions are influenced by subjective factors such as caregiver preferences, resource availability, and social support, which may not be fully captured by clinical and demographic data alone. Top predictors for discharge location include gender, marital status, insurance type, admission type, number of comorbidities, and ICU procedures and medications. These findings align with previous studies⁶³ that highlighted the importance of clinical and patient characteristics in discharge planning. The association of ICU procedures and medications with discharge location reflects the intensity of care required, further emphasizing their relevance in critical care contexts.

Sociodemographic variables, including age and race, consistently emerged as important predictors across models. These associations should be interpreted cautiously and understood as reflecting structural and systemic factors rather than biological determinants. Differences in baseline health status, access to care, cumulative disease burden and patterns of healthcare utilisation may all contribute to these observed relationships. The persistence of sociodemographic predictors after adjustment for clinical and utilisation variables underscores the risk that predictive models trained on observational data may encode existing inequities unless equity-aware modelling strategies are considered.

This study had several limitations that warrant further consideration. First, the reliance on the MIMIC-IV database, which reflects a single institution’s patient population and care practices, may limit the generalisation of the findings to other healthcare settings with differing demographics and protocols. Second, the imbalanced distribution of outcomes, particularly for rare discharge destinations, may have reduced model stability and predictive accuracy. While ML models like RF demonstrated strong performance, their limited interpretability compared with simpler models like LR poses challenges for clinical adoption. Future research should focus on validating these findings across diverse datasets to enhance generalisability. Addressing outcome imbalances through advanced techniques without data loss will be critical for better prediction of rare outcomes like hospice discharge or leaving against medical advice. Moreover, future studies could explore fitting a GAM with a Zero-Inflated NB distribution to account for overdispersion and zero inflation simultaneously. Expanding this work to include real-time predictive tools integrated into Clinical Decision Support Systems could provide dynamic insights for resource allocation and care optimisation. Future work can also selectively apply model-specific explainability methods, such as SHAP for tree-based models, to further explore feature contributions, while continuing to rely on appropriate interpretability strategies for regression-based models. Finally, although not a limitation, it would be recommended and appropriate to consider the most recent version of the MIMIC-IV database to account for all updates and changes.

Conclusions

This study highlights the potential of ML in advancing critical care for hepatitis patients. The models identified key predictors of ICU LoS, discharge outcomes and discharge locations, with clinical factors such as pre-ICU and ICU procedures and medication counts playing a significant role in determining these outcomes. Additionally, sociodemographic factors, including age and race, were consistently identified as important predictors, highlighting the presence of racial disparities among ICU-admitted patients diagnosed with hepatitis. These findings highlight the importance of addressing such disparities while using predictive analytics to optimise resource allocation, improve patient outcomes and guide targeted interventions. Despite its strengths, the study also revealed challenges such as variability in extended LoS and limited data for multi-class predictions, underscoring the need for further advancements in predictive modelling strategies in these areas. Continued research in this area is essential not only to enhance predictive models but also to support broader public health efforts aimed at preventing, detecting and managing hepatitis, ultimately reducing the burden of liver disease and hepatitis-related complications.

Data availability statement

Data may be obtained from a third party and are not publicly available.