Introduction

Liver cancer is one of the most common and fatal cancers worldwide. It is a leading cause of cancer death in both men and women, accounting for a significant proportion of global cancer-related deaths1. In 2020, over 830,000 people died from liver cancer, making it the third leading cause of cancer-related death worldwide2. Although treatment methods for liver cancer have improved in recent years, the prognosis remains poor, with low five-year survival rates across all stages, especially for patients with advanced liver cancer3,4,5,6. Many liver cancer patients require admission to the intensive care unit (ICU) due to complications such as acute liver failure, hepatic encephalopathy, gastrointestinal bleeding, and infections7,8. Among these, liver failure is the primary reason for ICU admission in liver cancer patients. Despite advancements in treatment, the mortality rate for ICU-admitted liver cancer patients remains extremely high, with studies showing a mortality rate exceeding 50%, and even higher for advanced-stage patients9,10. At present, the lack of early prediction and risk stratification of in-hospital mortality remains a major challenge for ICU clinicians11,12,13. Deciding which high-risk and poor-prognosis liver cancer patients should be admitted to the ICU requires considering a range of complex factors, including underlying diseases, liver function reserves, treatment plans, and the preferences of patients and their families14. These critically ill liver cancer patients generally have low long-term survival rates, and the cost of healthcare during hospitalization is high. Therefore, it is necessary to explore risk prediction models to identify high-risk groups among ICU-admitted liver cancer patients. The development of artificial intelligence has significantly improved the predictive models for estimating cancer patient mortality risk. Machine learning (ML), a novel form of artificial intelligence, can leverage large datasets and the rapid development of deep learning to transform measurement results into relevant predictive models, especially for cancer15,16. In recent years, ML has been proven effective in predicting liver cancer susceptibility, recurrence, and patient survival17,18,19. However, there is still limited data on the use of ML models to predict in-hospital mortality risk for liver cancer patients in ICU settings. Therefore, this study aims to establish four ML models, logistic regression, random forest, gradient boosting machine (GBM), and extreme gradient boosting (XGBoost) to predict the in-hospital mortality rate for ICU-admitted liver cancer patients. These models will help develop personalized prevention strategies for critically ill liver cancer patients and assist clinicians in making treatment decisions. Additionally, we aim to compare these four ML models to determine the best model for predicting in-hospital mortality in ICU-admitted liver cancer patients.

Methods

Data source

This retrospective study used data from the Medical Information Mart for Intensive Care (MIMIC-III and MIMIC-IV) databases. MIMIC-III contains data from over 60,000 ICU patients treated at the Beth Israel Deaconess Medical Center (BIDMC), Harvard University, between 2001 and 201220. MIMIC-IV (version 2.0) extends the data range to include ICU patients from 2008 to 201921. Both databases include rich clinical characteristics, laboratory results, vital signs, and medical records, providing high-quality data for this study. Since the data used in this study are from public databases, patient consent and institutional review board (IRB) approval were not required. All research procedures strictly followed the ethical standards set by the 1964 Declaration of Helsinki and its subsequent amendments. After completing an online training course and the Protecting Human Research Participants exam, we were granted permission to extract data from the MIMIC-III and MIMIC-IV databases. Methodologically, the study adheres to the guidelines of the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) statement22.

Fig. 1
figure 1

The flow chart of sample selection.

Full size image

Cohort selection

Patients were excluded from the study if they met any of the following criteria: (1) age under 18 years at the time of first ICU admission; (2) multiple ICU admissions. We selected patients who met the diagnostic criteria from the MIMIC-III and MIMIC-IV databases. Patients with ICD-9 codes starting with 155 or 197.7 and ICD-10 codes starting with C22 or C78 were included in the study. We randomly selected the MIMIC-III database as the training cohort and the MIMIC-IV database as the validation cohort. To prevent data leakage, we only selected samples from MIMIC-IV where the admission occurred after 2014. In the end, 862 patients from the MIMIC-III database were included in the training cohort, and 692 patients from the MIMIC-IV database were included in the validation cohort. The specific process is shown in Fig. 1.

Features and outcome

Baseline characteristics and admission information included age and gender. Comorbidities were collected based on the ICD codes recorded in both databases, including hypertension, diabetes, chronic kidney disease, myocardial infarction, congestive heart failure, atrial fibrillation, valvular disease, chronic obstructive pulmonary disease, stroke, hyperlipidemia, and liver disease. The Charlson Comorbidity Index (CCI)23,24 was also included. Additionally, severity scores were collected, including the Sequential Organ Failure Assessment (SOFA) score25, Oxford Acute Severity of Illness Score (OASIS)26, Acute Physiology Score III (APSII)27, Systemic Inflammatory Response Syndrome (SIRS) score, Logistic Organ Dysfunction Score (LODS)28, and the Model for End-Stage Liver Disease (MELD) score29. Furthermore, the initial vital signs and laboratory results from the 24 h before ICU admission were measured. Detailed laboratory tests and vital signs are shown in Table S1. Since these features were routinely measured at the time of patient admission and are easy to obtain in practice, no further feature selection was performed. The missing data rate for all variables included in the study was kept below 30%. To reduce the impact of missing data on the classification process, we used the Predictive Mean Matching (PMM) method for imputation. PMM predicts the mean of missing values and finds the closest observed values from the dataset to fill in the missing data30,31. This method has performed well in handling missing values and effectively estimates the missing data while reducing potential bias. We calculated the mean and standard deviation for each feature in the training set and performed Z-standardization on the features. The primary outcome of the study was in-hospital mortality, defined by the patient’s status at discharge (whether alive or dead).

Statistical analysis

For normally distributed continuous data, the mean and standard deviation were used for description. For between-group comparisons of normally distributed continuous data, a t-test was used. For data that did not follow a normal distribution, the Wilcoxon Mann-Whitney U test was used. Categorical data were described as frequencies and percentages, and between-group differences were compared using the chi-square test or Fisher’s exact test. A p-value of less than 0.05 was considered statistically significant. Baseline characteristics were reported for the training and validation cohorts. All statistical analyses were performed using R (version 4.0.1).

Model development and evaluation

In this study, we used random stratified sampling to split the MIMIC-III dataset into a training set (80%) and an internal test set (20%). The internal test set was created to compare its performance with that of the external test set (MIMIC-IV) and to prevent overfitting. To address the class imbalance in the training data, we applied the Synthetic Minority Over-sampling Technique (SMOTE) within the training set to balance the number of classes. The internal test set and external test set were kept unchanged to ensure unbiased model evaluation. When testing the model on the external test set, all samples from the MIMIC-III dataset were used as the training set. The study focused on applying 4 popular machine learning algorithms (logistic regression, random forest (RF), extreme gradient boosting (XGBoost), and LightGBM) to predict in-hospital mortality in liver cancer ICU patients.

To evaluate model performance, we used confusion matrix metrics, including accuracy, area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), accuracy, and F1 score. We also compared the performance of models trained with SMOTE to those trained without SMOTE to evaluate the impact of class balancing. In addition, we plotted calibration curves for both SMOTE and non-SMOTE models and used the Brier score to further assess the calibration and overall performance of the models. We used 1,000 bootstrap iterations to calculate the confidence intervals for all performance metrics, including AUROC, AUPRC, accuracy, and F1 score. The model with the highest AUROC in both the internal test set and external test set was considered to have the best predictive performance. The modeling work was carried out using Python (version 3.7.3).

Kaplan-Meier curve

In the external validation cohort, we used each machine learning model to calculate the predicted probability of in-hospital mortality for each patient. The optimal threshold was defined as the cutoff point that achieved the maximum sum of sensitivity and specificity in the training set. Patients were stratified into high-risk and low-risk groups based on this threshold. Kaplan-Meier (KM) survival curves were constructed using the survival package in R, with ICU admission as the starting point and either in-hospital death (event) or discharge (censored) as the endpoint. A log-rank test was performed to compare the survival differences between the two risk groups, and a P-value less than 0.05 was considered statistically significant. This time-to-event analysis was used to evaluate the models’ ability to stratify mortality risk over time and to assess their clinical utility in early prognostic evaluation of critically ill liver cancer patients.

SHAP analysis

To evaluate the importance of individual features, we used feature ranking as a method to assess their impact on the final prediction. The contribution of each input variable to the model output was determined using Shapley values, which are based on cooperative game theory. To address the opaque nature of machine learning algorithms and facilitate clinical interpretation, we applied SHAP and Local Interpretable Model-agnostic Explanations (LIME) to help explain the predictions made by the most effective models. We calculated feature SHAP values for the best-performing models from both the internal and external test sets. Then, we used beeswarm plots to show the relationship between the top 20 features’ values and their SHAP values, based on the global SHAP values from the training, internal test, and external test sets. Additionally, we presented scatter plots of the top 20 features in the training set to help visualize the dependence between feature values and their corresponding SHAP values.

Results

Baseline characteristics

A total of 1,554 patients were included in this study, with 862 patients in the training cohort extracted from the MIMIC-III database and 692 patients in the validation cohort extracted from the MIMIC-IV database. In the training cohort, 226 patients (26.2%) died during hospitalization, while in the validation cohort, 184 patients (26.6%) died. Table 1 shows the baseline characteristics of the training and validation groups. The training cohort had a higher proportion of patients with mild liver disease (26.8% vs. 15.3%), while the validation cohort had a higher proportion of patients with severe liver disease (14.0% vs. 8.7%). In terms of scores, the Charlson Comorbidity Index was significantly lower in the training cohort (2.98 ± 1.35) compared to the validation cohort (3.44 ± 1.26). Additionally, the SIRS score was lower in the validation cohort (2.58 ± 0.87 vs. 2.87 ± 0.92). Although there were no significant differences in SOFA, OASIS, and SAPS II scores between the two groups, the in-hospital mortality rate was similar (26.6% in the validation cohort vs. 26.2% in the training cohort), suggesting comparable disease severity between the two groups. See Table 1 for details.

Table 1 Comparisons of baseline characteristics in all cohorts.
Full size table

Model performance

We used 4 models to predict in-hospital mortality using all features. As shown in Table 2, which reports the results of models trained without SMOTE, the logistic regression model had the poorest predictive performance, followed by XGBoost, LightGBM, and random forest. In the internal test set, the random forest model performed the best, with the highest AUROC of 0.911 (95% CI: 0.855–0.956), AUPRC of 0.823 (95% CI: 0.718–0.905), and accuracy of 0.861 (95% CI: 0.809–0.908), although its F1 score 0.745 (95% CI: 0.632–0.836) was slightly lower than that of logistic regression 0.722 (95% CI: 0.610–0.818). In the external test set, random forest also performed excellently, with an AUROC of 0.857 (95% CI: 0.826–0.889) and accuracy of 0.828 (95% CI: 0.802–0.857), although its F1 score decreased to 0.577 (95% CI: 0.502–0.641). LightGBM and XGBoost performed similarly across both datasets, with LightGBM showing strong performance in the internal test set with an AUPRC of 0.786 (95% CI: 0.662–0.888), but its performance slightly decreased in the external test set with an AUPRC of 0.717 (95% CI: 0.657–0.773).

When comparing models trained with SMOTE to those without SMOTE, we found that models trained without SMOTE consistently outperformed those trained with SMOTE across key performance metrics, particularly AUROC, AUPRC, and Brier score. For example, in the external test set, the random forest model without SMOTE achieved a Brier score of 0.1254, which was lower than the SMOTE-trained version (0.1447), indicating better calibration and overall performance. These findings suggest that applying SMOTE did not improve, and in some cases slightly degraded, the predictive performance and calibration of the models.

Additionally, we performed ROC analysis to further confirm the predictive ability of these four models for in-hospital mortality. As shown in Fig. 2A and B, logistic regression had the worst predictive performance, followed by XGBoost and LightGBM. Overall, the random forest model showed the best predictive performance in both the internal and external test sets. Furthermore, we plotted Kaplan-Meier survival curves for the high-risk and low-risk groups identified by the four models in the external test set (Fig. 3). The log-rank test showed that the high-risk and low-risk groups identified by all four models were effective (P < 0.001), indicating that all four models can be used for early mortality risk prediction in liver cancer patients entering the ICU. Figure S1 shows the predicted probability distributions for the four models in the training, internal test, and external test sets.

The detailed results for models trained with SMOTE are provided in Table S2, while a direct comparison of model performance between SMOTE and non-SMOTE training is presented in Fig. S3. The corresponding calibration curves and comprehensive Brier score analyses are shown in Fig. S4.

Table 2 Performance of the prediction models using all features (without SMOTE).
Full size table
Fig. 2
figure 2

The performance of the four in-hospital mortality predictive models. ROC curves of the four prediction models using all features to predict in-hospital mortality (A) in the internal test set and (B) in the external test set.

Full size image

Feature importance analysis

To identify the important features that affect the model’s output, we performed feature importance analysis using the best-performing random forest model. We calculated the SHAP values for the features in the training set, internal test set, and external test set, and ranked them. Figure 4 shows the top 20 features for each dataset. Although the feature order slightly varies across datasets, many high-ranking features are common, indicating that these features generally have strong predictive power in the model. In the training set, APSIII, SAPSII, and LODS were among the top features, followed by OASIS, average respiratory rate, and average temperature. These features appeared consistently in all datasets, highlighting their importance in predicting in-hospital mortality risk for liver cancer ICU patients. The feature ranking in the internal test set was highly consistent with that of the training set, though certain features, like average SPO2 and minimum heart rate, had a different order. However, APSIII, SAPSII, LODS, and OASIS remained at the top, showing the stability and consistency of these clinical scores across different datasets. The feature importance ranking in the external test set was slightly different, especially with higher rankings for maximum ALB and minimum SBP, but key features like APSIII, SAPSII, and LODS still appeared in the top 20, further confirming the stability of these features in cross-dataset applications. The horizontal axis and color of the points indicate that the random forest model successfully captured the relationship between scores like APSIII, SAPSII, LODS, OASIS, and SOFA, with higher scores being associated with a higher probability of in-hospital mortality. More details on the relationship between feature values and SHAP values are shown in Fig. S2.

Fig. 3
figure 3

Kaplan-Meier survival curves for high-risk and low-risk groups identified by different models.

Full size image
Fig. 4
figure 4

Summary of SHAP values for the top 20 features in the training set, internal test set, and external test set. (A), (B), and (C) represent the training set, internal test set, and external test set, respectively. The color in the plot indicates the size of each feature’s value, with the color gradient from blue to red representing an increase in variable value from low to high. The horizontal axis of the points represents the SHAP values for the corresponding features, reflecting the extent of the feature’s impact on the model’s predictions. Positive SHAP values indicate a positive effect, while negative SHAP values indicate a negative effect.

Full size image

Discussion

In this retrospective study, we developed and validated machine learning algorithms based on clinical features using the public databases MIMIC-III and MIMIC-IV to predict in-hospital mortality in critically ill liver cancer patients. The MIMIC-III and MIMIC-IV datasets represent data from different time periods. We used the MIMIC-III dataset as the training set and the MIMIC-IV dataset as the external test set. This time-based data division strategy, where earlier data is used for model development and later data for validation, helps effectively reduce the risk of data leakage and overfitting. This design makes the model’s development and validation more externally valid because it avoids using future information for training, making it more aligned with real-world clinical settings. The random forest model showed the best overall performance, with an AUROC of 0.844 in the external test set. Using advanced machine learning techniques, we identified several important clinical features related to in-hospital mortality, including scores like APSIII, SAPSII, LODS, OASIS, as well as vital signs such as average respiratory rate, temperature, heart rate, oxygen saturation, and laboratory indicators like minimum lactate, maximum creatinine, minimum WBC, and INR. These results provide insights that need further consideration. In solid tumors, liver cancer patients have a high ICU in-hospital mortality rate. Studies show that the mortality rate due to liver cancer rupture can range from 25% to 75% during the acute phase32,33. ICU admission-related annual medical costs in the U.S. amount to $3 billion, with an average cost of $116,200 per admission34. Advanced cirrhosis patients are likely to be admitted to the ICU due to critical conditions such as sepsis, kidney failure, or respiratory failure. Although some studies report improved prognosis for cirrhosis patients admitted to the ICU, the prognosis remains poor, with mortality rates as high as 45% or even higher. These findings highlight the importance of predicting and intervening in in-hospital mortality for critically ill liver cancer patients. However, clinicians still face challenges in identifying high-risk ICU patients. Therefore, it is crucial to develop and promote reliable predictive models to identify these patients early and provide effective interventions to improve their prognosis.

Given the increasing applicability and effectiveness of supervised machine learning algorithms in disease prediction modeling, the scope of research is expanding. Well-known supervised learning classifiers, such as XGBoost, random forest, LightGBM, and neural networks, have gradually been applied in clinical practice15,35. With the help of machine learning classification, decision support models assisted by machine learning have proven to be more advantageous than traditional linear models. In this study, random forest, LightGBM, and XGBoost demonstrated their ability to capture nonlinear relationships, and their performance surpassed that of logistic regression, which is consistent with previous reports36. As an ensemble learning method, random forest has several significant advantages. By building multiple decision trees and making voting decisions, it effectively avoids overfitting that may occur with a single decision tree, thereby improving the model’s robustness and generalization ability. Additionally, random forest can handle high-dimensional feature spaces and maintain high prediction accuracy, especially when the number of features far exceeds the number of samples. These advantages make random forest the most effective model in this study. We used Kaplan-Meier survival curves to assess the effectiveness of different models in predicting in-hospital mortality risk for liver cancer patients admitted to the ICU. By dividing patients into high-risk and low-risk groups based on each model’s predicted probabilities, we clearly observed that the survival probability of the high-risk group was significantly lower than that of the low-risk group. This indicates that the clinical data routinely collected has the ability to distinguish between high-risk and low-risk liver cancer ICU patients. In addition, we explored the effect of applying the SMOTE to address potential class imbalance. We compared the performance of models trained with SMOTE and those trained without SMOTE. The models without SMOTE consistently achieved better discrimination and calibration, so we ultimately adopted the results without SMOTE for our final analysis. These findings are consistent with previous research37,38, which has reported that resampling techniques such as SMOTE may sometimes degrade model calibration.

We used the visualization capabilities of SHAP to analyze the impact of individual feature values on the model’s output. The scores of APSIII, SAPSII, LODS, and OASIS had the greatest impact. These scoring systems are widely used in clinical practice to assess the severity of illness in critically ill patients. The results show that early in the ICU admission of liver cancer patients, these scores can effectively help predict in-hospital mortality. APSIII mainly considers clinical physiological indicators such as partial pressure of oxygen, blood pressure, temperature, and laboratory results like serum electrolyte levels in the early ICU phase of liver cancer patients. SAPSII primarily assesses acute pathophysiological status, covering a comprehensive evaluation of blood gas analysis, kidney function, liver function, and more. LODS assesses the severity of multiple organ failure and provides strong predictive support for liver cancer ICU patients. It scores based on the function of multiple organs, while the OASIS score combines physiological data and clinical features, offering high sensitivity in predicting mortality risk for critically ill patients.

However, there are several important limitations in this study. First, the retrospective and observational design may introduce unavoidable selection bias and confounding factors, which could affect the generalizability of the findings. Second, the data were exclusively derived from the publicly available MIMIC-III and MIMIC-IV databases, which are limited to a single healthcare system in the United States. Therefore, the model’s performance may not fully represent patient populations in other institutions or countries, and further external validation using multi-center datasets is warranted. Third, our dataset did not include pathological or radiological information related to liver cancer, making it impossible to determine tumor stage or incorporate it into the prediction models. The absence of this information may have introduced bias, as disease stage is an important prognostic factor. Finally, the study focused on machine learning techniques and primarily reported discrimination metrics such as AUROC and AUPRC. Future studies should also emphasize model calibration, clinical utility assessment, and prospective validation before clinical implementation.

Conclusions

In this study, we applied four machine learning methods to predict in-hospital mortality in critically ill liver cancer patients. The results showed that the random forest model performed the best. The scores of APSIII, SAPSII, LODS, and OASIS were the most important factors influencing the model’s prediction of in-hospital mortality. This study provides clinical feature explanations that offer useful reference information for ICU clinicians in predicting patient prognosis.