A tabular prior-data fitted network-based prediction model integrating ferritin and total iron-binding capacity for 30-day all-cause mortality in critically ill cancer patients Authors: Lin-Lin Liu, Lei Han, Zuo-Lin Xiang Published: 2026 ABSTRACT AbstractObjectivesAccurately predicting outcomes for critically ill cancer patients remains challenging. This study aimed to integrate biologically relevant iron metabolism markers into short-term mortality prediction and to develop an interpretable, externally validated model based on a tabular prior-data fitted network (TabPFN) for risk stratification in this population.Materials and MethodsWe conducted a retrospective cohort study using data from Medical Information Mart for Intensive Care IV (MIMIC-IV) (training and internal validation) and eICU Collaborative Research Database (eICU-CRD) (external validation), including critically ill adult patients with cancer. Associations between iron metabolism markers (ferritin, serum iron, and total iron-binding capacity [TIBC]) and 30-day all-cause mortality were assessed using the Kaplan–Meier curves, multivariable Cox regression, and restricted cubic splines. A TabPFN-based prediction model was developed and compared with 8 other machine-learning algorithms, with the least absolute shrinkage and selection operator used for feature selection. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), calibration, and decision curve analysis. SHapley Additive exPlanation values provided model interpretability.ResultsAmong 1137 patients in the MIMIC-IV cohort, 293 (26.0%) died within 30 days. Elevated ferritin (hazard ratio [HR] = 1.18, 95% CI, 1.09-1.27) and decreased TIBC (HR = 0.92, 95% CI, 0.85–0.99) were independently associated with mortality and exhibited non-linear relationships. Serum iron showed no prognostic value. The TabPFN model achieved the best performance, with an AUROC of 0.865 (95% CI, 0.812-0.919) in the internal validation set and 0.772 (95% CI, 0.744-0.814) in the external validation set, along with good calibration (Brier score = 0.121) and clinical net benefit. SHapley Additive exPlanation analysis identified lymphocyte count, platelet count, lactate, and ferritin as the most influential predictors.ConclusionIron metabolism dysregulation—particularly altered ferritin and TIBC levels—has important prognostic value in critically ill cancer patients. By integrating these biomarkers with a TabPFN framework, this study provides an accurate, interpretable, and externally validated tool for 30-day all-cause mortality prediction that may support clinical decision-making. Background Malignant tumors pose a significant global public health challenge. Improvements in diagnostic and therapeutic methods have extended cancer patient survival. This has led to a notable rise in intensive care unit (ICU) admissions due to complications from the disease or adverse treatment effects, including infections, sepsis, and organ failure.1,2 Statistics indicate that approximately 5%-10% of cancer patients experience such life-threatening conditions requiring ICU support.3,4 Intensive care unit admission rates can reach nearly 30% in high-risk groups, including patients with esophageal cancer, acute leukemia, or those undergoing allogeneic stem cell transplantation. Currently, cancer patients constitute 13.5%-21.5% of all ICU admissions.5,6 Managing critically ill cancer patients is challenging due to the significant variability in outcomes. Accurate prognosis assessment is crucial for guiding clinical decision-making—including determining the intensity of treatment, communicating expectations with families, and allocating medical resources effectively. Although general critical illness scoring systems like SOFA and APACHE II exist, they may not adequately capture the unique pathophysiological state of cancer patients, potentially leading to prognostic inaccuracies. Consequently, there is a pressing need in the intersecting fields of intensive care medicine and oncology to develop more precise prognostic models that incorporate cancer-specific biomarkers, such as iron metabolism markers. Iron metabolism dysregulation represents a core pathophysiological link connecting cancer and critical illness. Iron is a vital trace element crucial for essential cellular functions such as oxygen transport, energy metabolism, and DNA synthesis. In oncology, cancer cells exhibit a distinctive “iron addiction” to sustain their rapid proliferation.7 This addiction is characterized by increased expression of iron uptake proteins such as transferrin receptor 1 (TFR1) and decreased expression of the iron export protein ferroportin, resulting in an enlarged intracellular labile iron pool to satisfy elevated metabolic needs.8 This iron-rich state presents a dual challenge: while iron is essential for cell proliferation, free iron ions in the labile iron pool can trigger excessive reactive oxygen species production through the Fenton reaction. This induces oxidative stress, leading to ferroptosis, a distinct type of iron-dependent regulated cell death.8 Therefore, cancer cell survival heavily relies on sophisticated regulatory mechanisms to control iron metabolism and resist ferroptosis, such as upregulating the System Xc-/GPX4 axis to maintain redox balance.9 Iron metabolism markers such as serum iron, ferritin, transferrin, and total iron-binding capacity (TIBC) effectively indicate tumor aggressiveness, metabolic activity, and oxidative stress levels. Traditional linear statistical models often struggle to fully decipher the complex interactions among these variables. Artificial intelligence (AI), especially machine learning (ML), can autonomously detect intricate patterns and non-linear relationships in extensive datasets, creating predictive tools that may surpass traditional models. Machine-learning techniques, with their robust ability to manage high-dimensional and non-linear data, have shown considerable promise in medical prediction.10,11 They not only offer high predictive accuracy but also, through explainability techniques like SHapley Additive exPlanation (SHAP), can reveal the contribution of each predictor to individual risk. This enhances the model’s clinical credibility and may provide clues for uncovering underlying biological mechanisms. These models can integrate multimodal data, such as electronic health records and biomarkers, to identify latent risk patterns, demonstrating potential in predicting sepsis outcomes and forecasting cardiac arrest.12,13 Few studies have evaluated mortality risk in critically ill cancer patients by jointly considering iron metabolism dysregulation and AI-based modeling. Moreover, the prognostic roles of ferritin and TIBC in this highly heterogeneous population, as well as their nonlinear relationships with mortality, remain unclear. In this context, the present study is designed to address these gaps by systematically evaluating iron metabolism markers and incorporating them into an interpretable, externally validated prediction model based on a tabular prior-data fitted network (TabPFN). Specifically, we aimed to characterize the prognostic value of ferritin, serum iron, and TIBC and to develop a machine-learning tool for 30-day all-cause mortality prediction and clinical decision support. Methods Data source and ethics This study utilized data from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, version 3.1,14 and eICU Collaborative Research Database (eICU-CRD).15 Medical Information Mart for Intensive Care IV contains ICU patient data from the Beth Israel Deaconess Medical Center, while eICU-CRD includes data from multiple US institutions. This study obtained CITI Program certification (Record IDs: 72483467, 72483466) and was granted an ethics exemption by the Institutional Review Board. Informed consent from patients was not required as all data were de-identified. The source code is available at https://github.com/Morii0603/iron-based-prediction. Study population Malignant tumor cases were identified using International Classification of Diseases, Tenth Revision (ICD-10) codes C00-C97. Cancer diagnosis was based on the first hospital admission diagnosis. The study involved ICU patients with confirmed cancer diagnoses. Participants were excluded if they were under 18 years, had an ICU stay of less than 24 hours, or lacked measurements for ferritin, serum iron, transferrin, or TIBC. A total of 1137 patients were ultimately included. Additionally, 116 patients were extracted from the eICU database as an external validation cohort using identical inclusion and exclusion criteria. To assess the representativeness of the external validation cohort and the risk of selection bias, we evaluated whether the subgroup of cancer patients who underwent iron metabolism tests could represent cancer patients to some extent. We collected baseline indicators for cancer patients who underwent iron metabolism tests and those who did not and analyzed whether there were statistical differences between the 2 groups. Follow-up time was until 1 year after the patient left the ICU, with the endpoint defined as 30-day all-cause mortality. Only the initial ICU admission record was analyzed for patients with multiple ICU stays. Based on the rule of thumb requiring 10 outcome events per predictor variable, the sample size was deemed sufficient for developing a model with approximately 30 predictors. Variables and data preprocessing This study utilized all available data from MIMIC-IV and eICU-CRD that met the inclusion criteria. Data gathered from MIMIC-IV and eICU-CRD encompassed demographics, comorbidities, vital signs, lab results, disease severity scores, treatment measures, hospital and ICU stay durations, and follow-up survival status. The first measured values after ICU admission were used for vital signs and laboratory indicators. Machine-learning models rely on complete data inputs and are sensitive to missing values. Therefore, variables with over 20% missing data were omitted. For variables with ≤20% missingness, imputation was performed using the random forest algorithm within the training set only; the imputation model was then applied to the internal and external validation sets to prevent data leakage. Additionally, data were strictly screened to remove records with obvious errors or logical inconsistencies, such as unreasonable extreme values for laboratory indicators. Both transferrin and TIBC were extracted as candidate iron metabolism markers. Given the strong linear relationship between transferrin and TIBC (TIBC [µg/dL]≈transferrin [mg/dL]×1.25), including both would introduce multicollinearity. Therefore, only TIBC was retained for subsequent analyses, as it is the more commonly reported parameter in clinical practice and was less affected by assay variability in the dataset. Statistical analysis and software environment Statistical analysis and modeling were performed on the R (version 4.5.2) and Python (version 3.13.11) platforms. Categorical data are presented as counts with corresponding percentages, and group comparisons were conducted using the χ2 test. Continuous variables are expressed as mean ± standard deviation or median (interquartile range) based on normality; group comparisons utilized the t-test or Mann–Whitney U test. All hypothesis tests were 2-sided, with statistical significance set at a P value of less than .05. Ferritin, serum iron, and TIBC were grouped according to quartiles based on the entire cohort. The Kaplan–Meier survival curves and Cox proportional hazards models were employed to examine the relationship between iron metabolism markers and clinical outcomes in cancer patients, assessing their prognostic impact. The Cox regression was adjusted for gender, age, comorbidities, and key laboratory indicators. ML model development The least absolute shrinkage and selection operator (LASSO) regression method was utilized for feature selection. It applies an L1 regularization penalty term to the loss function, shrinking regression coefficients and reducing the coefficients of less contributive variables to zero, thereby performing automatic variable selection. All candidate predictor variables were entered into the LASSO model, with the optimal penalty parameter λ identified through 10-fold cross-validation. The optimal parameter was chosen as the λ value that resulted in the minimum cross-validation error (λmin). Variables with non-zero coefficients at this λ value were the final indicators selected for the prediction model. TabPFN16 is a foundation model based on the transformer architecture, specifically designed for small-sample tabular data prediction. The model acquires general meta-learning capabilities through pre-training on a massive number of synthetic tasks. When faced with a new prediction task, TabPFN can perform rapid inference based on in-context learning via a single forward pass. This study compared it with several traditional ML methods, such as logistic regression, XGBoost, LightGBM, random forest, AdaBoost, decision tree, gradient boosting decision tree, and multi-layer perceptron. To address class imbalance: Since TabPFN internally encodes strong inductive biases for handling various data distributions (including class imbalance), during the inference stage post-processing, it only requires balancing the probabilities based on the training data distribution to improve classification performance. For other ML models, class weights were used to handle the class imbalance problem. All ML models were developed in a Python environment (using libraries such as scikit-learn 1.7.2, xgboost 3.1.2, lightgbm 4.6.0, PyTorch 3.11.0, transformers 4.57.6, and tabpfn 6.3.0). For the TabPFN-based model, we used the pre-trained TabPFN v2.5 classifier checkpoint (′tabpfn-v2.5-classifier-v2.5_default.ckpt′) and applied it directly to the target task without task-specific fine-tuning of the model weights. Since TabPFN is designed to perform well with default configurations and generally requires little hyperparameter tuning, we used the v2.5 default settings, including n_estimators = 8, softmax_temperature = 0.9, and auto_scale_n_estimators = True. To ensure the reproducibility of experiments, random seeds were fixed for all experiments. Model evaluation In the preprocessing phase, numerical features were standardized via z-score normalization using statistics derived from the training folds. Model development was performed on the MIMIC-IV cohort, where hyperparameter tuning was conducted using grid search with 5-fold cross-validation. Area under the curve (AUC) estimates with 95% CIs were derived from this internal cross-validation. The optimal hyperparameters were then used to train the final model on the entire MIMIC-IV dataset. Subsequently, the model’s generalizability was assessed on the external validation set from eICU-CRD cohort using AUC, accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score. To evaluate the stability of the AUC estimate and quantify its uncertainty, the non-parametric Bootstrap resampling technique was employed. One thousand samples were drawn with replacement from the original data, and the AUC was recalculated for each sample, thereby obtaining the distribution of AUC and estimating its 95% CI. Calibration curves were plotted to analyze prediction accuracy and conduct a comprehensive evaluation, verifying its potential application value in clinical decision support and simulation modeling. Explainability Shapley Additive exPlanations utilizes cooperative game theory to assign contribution values to features, improving model decision transparency and interpretability. Permutation SHAP analysis was conducted on the TabPFN model to improve comprehension and trust in its decision-making process. Results Initial correlation between study population baseline characteristics and iron metabolism markers The study included 1137 critically ill patients diagnosed with malignant tumors. Based on 30-day clinical outcomes after ICU admission, patients were categorized into survival (n = 844, 74.0%) and non-survival (n = 293, 26.0%) groups. Baseline characteristics are summarized in Table 1. Compared with the survival group, non-survivors had significantly higher ferritin levels and lower TIBC levels (both P < .001), whereas serum iron levels did not differ significantly between the groups. The patient enrollment flowchart is shown in Figure 1. Table S1 shows that in the eICU cohort, patients who underwent iron metabolism indicator testing were older, with longer hospital stays and ICU stays, suggesting that this subgroup may have more complex clinical conditions or higher intensity of medical resource use. However, the difference in hospital mortality between the 2 groups was not significant (15.2% vs 14.0%, P = .398), indicating that patients who underwent iron metabolism testing did not deviate significantly from those who did not in terms of the risk of the primary outcome. Flowchart of participant selection and model development. The MIMIC-IV cohort included 1137 patients, comprising 293 deaths and 844 survivors, and was used for model training and internal cross-validation. The eICU cohort included 116 patients, comprising 24 deaths and 92 survivors, and was used for external validation. Statistical analyses included Cox regression, Kaplan–Meier analysis, restricted cubic spline analysis, LASSO feature selection, comparison of TabPFN with other machine-learning models, and SHAP-based model interpretation. Table 1.Baseline characteristics of the study population.VariableDeath (n = 293)Survival (n = 844) P Age71.23 ± 11.2968.68 ± 12.63.002Red blood cells, m/µL3.49 ± 0.623.67 ± 0.65<.001Hemoglobin, g/dL9.70 (8.90-10.80)10.30 (9.20-11.55)<.001RDW, %19.50 ± 3.5718.73 ± 3.77.002Hematocrit, %30.60 (28.30-33.90)32.30 (29.40-35.95)<.001Neutrophils, %86.60 (79.70-91.10)85.10 (78.00-91.15).170Lymphocytes, %10.00 (5.80-18.30)15.00 (9.00-25.20)<.001Platelets, K/µL271.08 ± 167.87328.02 ± 174.89<.001ALT, IU/L45.00 (20.00-113.00)37.00 (18.50-89.00).040AST, IU/L172.00 (100.00-398.00)139.00 (92.00-269.50).001Total bilirubin, mg/dL1.00 (0.50-3.00)0.80 (0.50-1.70).001Albumin, g/dL2.97 ± 0.613.21 ± 0.59<.001BUN, mg/dL49.00 (29.00-74.00)34.00 (22.00-60.50)<.001Creatinine, mg/dL2.10 (1.20-3.60)1.30 (0.90-2.30)<.001Glucose, mg/dL194.00 (150.00-283.00)202.00 (152.00-281.50).327Sodium, mEq/L142.00 (138.00-147.00)143.00 (140.00-146.00).011Chloride, mEq/L5.40 (4.20-7.10)4.80 (4.10-5.80)<.001Free calcium, mg/dL5.10 (4.60-5.80)5.00 (4.60-5.60).230Total calcium, mg/dL9.00 (8.60-9.70)9.20 (8.70-9.60).056PaO2, mmHg152.31 ± 99.70154.89 ± 106.22.716PCO2, mmHg48.00 (40.00-60.00)47.00 (41.00-56.00).518PH7.43 (7.38-7.47)7.44 (7.40-7.49).002Lactate, mmol/L3.00 (1.90-5.80)2.30 (1.60-3.85)<.001Anion gap, mEq/L1.00 (-2.00-4.00)2.00 (0.00-5.00)<.001INR1.70 (1.40-2.60)1.50 (1.30-2.00)<.001PT, s18.80 (15.10-28.40)16.50 (14.10-22.45)<.001PTT, s69.73 ± 44.8965.36 ± 46.02.159Heart rate, bpm92.27 ± 16.6189.35 ± 16.15.009Respiration rate, bpm20.66 (18.10-24.19)19.74 (17.64-22.33)<.001Temperature, °C36.78 (36.58-37.01)36.84 (36.68-37.02).002SBP, mmHg113.27 ± 14.81115.96 ± 14.22.006DBP, mmHg61.93 ± 8.9964.04 ± 9.72.001MBP, mmHg75.90 ± 9.0978.53 ± 9.56<.001SpO2, %96.50 (94.62-97.77)96.56 (95.48-97.74).017Weight admit, kg72.30 (61.00-85.00)74.00 (61.50-87.05).402Ferritin, ng/mL898.00 (330.00-2226.00)458.00 (206.50-1072.00)<.001TIBC, μg/dL176.00 (143.00-233.00)203.00 (159.00-254.00)<.001Iron, μg/dL40.00 (22.00-71.00)37.00 (22.00-65.00).144Male174 (59.4%)488 (58.4%).831Hypertension99 (33.8%)275 (32.9%).845Diabetes109 (37.2%)283 (33.9%).341Heart failure92 (31.4%)266 (31.9%).943MI41 (14.0%)88 (10.5%).136Stroke21 (7.2%)49 (5.9%).514Sepsis122 (41.6%)273 (32.7%).007AKI189 (64.5%)426 (51.0%)<.001CKD91 (31.1%)216 (25.9%).101Hyperlipidemia130 (44.4%)350 (41.9%).508Peripheral vascular disease21 (7.2%)41 (4.9%).190Arterial embolization2 (0.7%)11 (1.3%).577CRRT used23 (7.8%)53 (6.3%).455Vasoactive used118 (40.3%)206 (24.7%)<.001Abbreviations: AKI, acute kidney injury; ALT, alanine aminotransferase; AST, aspartate aminotransferase; CKD, chronic kidney disease; CRRT, continuous renal replacement therapy; DBP, diastolic blood pressure; INR, international normalized ratio; MBP, mean blood pressure; MI, myocardial infarction; PT, prothrombin time; PTT, partial thromboplastin time; RDW, red blood cell distribution width; SBP, systolic blood pressure; TIBC, total iron-binding capacity. Baseline characteristics of the study population. Abbreviations: AKI, acute kidney injury; ALT, alanine aminotransferase; AST, aspartate aminotransferase; CKD, chronic kidney disease; CRRT, continuous renal replacement therapy; DBP, diastolic blood pressure; INR, international normalized ratio; MBP, mean blood pressure; MI, myocardial infarction; PT, prothrombin time; PTT, partial thromboplastin time; RDW, red blood cell distribution width; SBP, systolic blood pressure; TIBC, total iron-binding capacity. Comprehensive assessment of iron metabolism markers for prognostic significance Patients were stratified into subgroups according to the quartiles of iron metabolism markers, and subsequent analyses were performed based on these thresholds. The distributions of these markers are presented in Table 2. Table 2.Distribution according to quartiles of iron metabolism-related indicators.VariableAllQ1Q2Q3Q4TIBC195 (152-247)126 (29-152)174 (153-195)217 (196-247)288 (248-751)Ferritin545 (223-1319)122 (8-223)350 (224-545)850 (546-1319)2437 (1321-2905)Iron37 (22-66)16 (5-22)29 (23-37)47 (38-66)104 (67-483)Abbreviation: TIBC, total iron-binding capacity. Distribution according to quartiles of iron metabolism-related indicators. Abbreviation: TIBC, total iron-binding capacity. The associations between iron metabolism indices and patient survival were evaluated. The Kaplan–Meier survival curves (Figure 2) demonstrated that both ferritin and TIBC levels were significantly associated with survival rates. Specifically, elevated ferritin levels were associated with poorer prognosis (P < .001), whereas decreased TIBC levels suggested worse survival outcomes (P < .001). Serum iron levels also showed a significant association with survival (P = .022). The relationship between iron-metabolism indicators at different levels and all-cause mortality, analyzed using Kaplan–Meier curves. (A) Survival curves according to TIBC quartiles. (B) Survival curves according to ferritin quartiles. (C) Survival curves according to serum iron quartiles. Multivariable Cox proportional-hazards regression analysis was performed to quantify the independent prognostic value of these factors, adjusting for potential confounders (Table 3). The analysis revealed that higher ferritin levels were independently associated with an increased risk of mortality (hazard ratio [HR] = 1.18, 95% CI, 1.09-1.27; trend P < .001). Conversely, higher TIBC levels were associated with a reduced mortality risk (HR = 0.92, 95% CI, 0.85-0.99; trend P = .034).Serum iron did not exhibit independent prognostic value (HR = 0.99, 95% CI, 0.92-1.06; trend P = .782). Table 3.The Cox proportional hazard ratios for the iron-metabolism indicators.VariableHR (95% CI)z P TIBC quartile0.92 (0.85-0.99)−2.117.034 Q11.00 (Ref) Q20.84 (0.69-1.03)−1.684.092 Q30.75 (0.61-0.92)−2.705.007 Q40.78 (0.61-1.00)−1.94.052Ferritin quartile1.18 (1.09-1.27)4.289<.001 Q11.00 (Ref) Q21.08 (0.86-1.36)0.656.512 Q31.27 (1.00-1.62)1.986.047 Q41.64 (1.28-2.10)3.878.001Iron quartile0.99 (0.92-1.06)−0.277.782 Q11.00 (Ref) Q20.86 (0.71-1.06)−1.423.155 Q30.79 (0.64-0.97)−2.257.024 Q41.00 (0.80-1.24)−0.027.978Age1.01 (1.01-1.02)3.592.001Gender0.95 (0.82-1.10)−0.69.49Sepsis0.92 (0.79-1.08)−0.999.318AKI1.41 (1.21-1.65)4.387.001Heart failure0.91 (0.78-1.07)−1.171.241Albumin0.81 (0.70-0.93)−3.065.002Sodium0.99 (0.97-1.00)−2.048.041Lactate1.06 (1.04-1.08)5.901.001Lymphocytes0.99 (0.99-0.99)−4.689.001Adjusted for age, gender, sepsis, AKI, heart failure, albumin, sodium, lactate, and lymphocytes.Abbreviations: TIBC, total iron-binding capacity. AKI, acute kidney injury. The Cox proportional hazard ratios for the iron-metabolism indicators. Adjusted for age, gender, sepsis, AKI, heart failure, albumin, sodium, lactate, and lymphocytes. Abbreviations: TIBC, total iron-binding capacity. AKI, acute kidney injury. To explore potential nonlinear relationships, restricted cubic spline analyses were conducted as is shown in Figure 3. Significant nonlinear associations with all-cause mortality were observed for both ferritin (P for nonlinearity <.001) and TIBC (P for nonlinearity = 0.004). For ferritin, HR increased sharply at low levels and plateaued at higher levels (P for overall <.001), with the curve crossing the reference line at approximately 350-500 and reaching a plateau at around 4700-5000 (P for overall <.001), indicating a threshold effect rather than a simple positive linear association. For TIBC, the risk was highest at low levels, lowest at moderate levels, with the curve crossing the reference line at approximately 196 and the estimated minimum risk occurring at around 276-280. A slight but non-significant increase was observed at high levels (P for overall <.001). In contrast, no statistically significant association was found between serum iron levels and mortality risk (P = .052). Restricted cubic spline (RCS) analysis of the associations between iron-metabolism indicators and all-cause mortality. (A) Association between TIBC and all-cause mortality. (B) Association between ferritin and all-cause mortality. (C) Association between serum iron and all-cause mortality. Development and validation of an optimal prediction model incorporating iron metabolism markers Feature selection was performed using LASSO regression (employing the λmin criterion), which screened 31 potential predictors, as shown in Table S2 and Figure S1. Based on these factors, 9 different ML methods (including TabPFN, LightGBM, XGBoost, etc.) were developed to forecast the 30-day all-cause mortality of critically ill cancer patients. During the training phase, 5-fold cross-validation was used to determine the optimal hyperparameters for each model. The performance comparison of the 9 ML models is shown in Figure 4. Among them, the TabPFN model performed best in cross-validation, achieving an area under the receiver operating characteristic curve (AUROC) of 0.865 (95% CI, 0.812-0.919) and an area under the precision-recall curve (AUPRC) of 0.942 (95% CI, 0.907-0.971). After identifying the optimal hyperparameters via 5-fold cross-validation, the models were retrained using the entire training set and then evaluated on the external validation set. The Bootstrap method was used to assess the CIs for AUROC and AUPRC. TabPFN achieved superior performance on the external validation set with an AUROC of 0.772 (95% CI, 0.744-0.814) and an AUPRC of 0.885 (95% CI, 0.866-0.909) compared with other models. Detailed results of other metrics (accuracy, F1 score, etc.) are presented in Table 4. Model performance comparison based on ROC and precision–recall curves. (A) Receiver operating characteristic (ROC) curves for different models in the cross-validation set. (B) Precision–recall curves for different models in the cross-validation set. (C) ROC curves for different models in the external validation set. (D) Precision–recall curves for different models in the external validation set. Table 4.Performance comparison of the models on the external validation set.ModelAccuracySensitivitySpecificityPrecisionNPVF1 scoreAUCTabPFN0.7780.8330.6320.8560.5900.8450.772LightGBM0.7830.8930.4910.8220.6360.8560.757XGBoost0.7920.9470.3860.8020.7330.8690.723Logistic regression0.7150.7400.6490.8470.4870.7900.750GBDT0.7870.9470.3680.7980.7240.8660.743Random forest0.7830.9730.2810.7810.8000.8660.751AdaBoost0.7730.9000.4390.8080.6250.8520.714MLP0.7200.9870.0180.7250.3330.8360.525Decision tree0.7250.8070.5090.8120.5000.8090.595Abbreviations: AUC, Area under the curve; GBDT, gradient boosting decision tree ;MLP, Multilayer Perceptron; NPV, Negative Predictive Value Performance comparison of the models on the external validation set. Abbreviations: AUC, Area under the curve; GBDT, gradient boosting decision tree ;MLP, Multilayer Perceptron; NPV, Negative Predictive Value Model validation and interpretability analysis Figure 5A illustrates the evaluation of the TabPFN model’s predicted probabilities accuracy on the external validation set on the external validation set. The results showed a Brier score of 0.121, indicating that the model not only has good discriminative ability but also shows high consistency between predicted and observed risks. This further proves the reliability of the model’s probabilities output, providing crucial support for its clinical applicability. The clinical decision curve analysis is shown in Figure 5B. The findings suggest that employing this model for clinical decision-making offers substantial net benefits to patients across a wide range of threshold probabilities, highlighting its strong clinical applicability. Evaluation of model performance and case demonstrations. (A) Calibration curve; (B) clinical decision curve; and (C) feature importance ranking in the prediction model. (C) Feature attributions in SHAP analysis of the optimal GBDT model, where each line represents a feature, with SHAP values plotted on the x-axis. Red dots: higher values, blue dots: lower values. (D) Variable importance in SHAP analysis of the optimal TabPFN model shown as bars, illustrating their contribution to the model predictions. (E) SHAP scores highlighting the predicted risk of 30 days all-cause mortality for an individual subject. The SHAP analysis evaluated the contribution of each variable within the TabPFN model. Red dots represent high-risk values, while blue dots signify low-risk values. Figure 5C shows the ranking of feature importance in the prediction model. The SHAP analysis identified lymphocyte count, platelets, lactate, and ferritin as the 4 core features with the highest contribution in this prediction model. This finding confirms clinical intuition and underscores the importance of ferritin as a crucial iron metabolism marker in forecasting the outcomes of critically ill cancer patients, which is consistent with the conclusions about iron metabolism markers in the second section. To highlight the model’s clinical utility, we demonstrated how the model integrates these variables (including the significantly contributing ferritin) for personalized risk prediction through a specific case example (Figure 5D and E), thus completing the cycle from construction and validation to interpretation and application. Discussion In this study, we systematically evaluated the prognostic value of iron metabolism markers and developed a TabPFN-based model for 30-day all-cause mortality prediction in critically ill cancer patients. The major contribution of this work is not merely the construction of a prediction model, but the integration of biologically meaningful iron metabolism biomarkers into an advanced AI framework within a highly heterogeneous and clinically challenging population. We found that ferritin and TIBC were independently associated with 30-day all-cause mortality and exhibited nonlinear relationships with risk, whereas serum iron showed limited prognostic value. Building on these findings, the TabPFN model integrating iron metabolism markers with routine clinical variables achieved superior discrimination in both internal and external validation, with good calibration and clinical utility. The SHAP analysis further supported the importance of ferritin, together with lymphocyte count, platelet count, and lactate, in individualized risk prediction. Collectively, these findings suggest that incorporating iron metabolism dysregulation into AI-based modeling may provide a more precise and interpretable strategy for 30-day all-cause mortality prediction in critically ill cancer patients. Cancer patients constitute about 13.5%-21.5% of ICU admissions.17–19 In the past, particularly patients with solid tumors were often restricted from ICU admission due to prognostic concerns.20 Advancements in cancer treatment and critical care technology have notably enhanced the prognosis for critically ill cancer patients.21 Multiple studies further confirm that these patients can derive clear benefits from intensive care.22,23 Optimal support for cancer patients necessitates a systematic understanding of their common serious complications, stemming from both the disease and its treatments, alongside the implementation of suitable medical management. The growing variety of anticancer drugs and their expanding side effect profiles make it increasingly difficult to keep up with treatment advancements and identify adverse reactions and drug interactions. The care of critically ill cancer patients necessitates multidisciplinary collaboration due to the complexity of their conditions. Clinical evidence shows that establishing multidisciplinary teams can effectively reduce ICU mortality.24 Therefore, unstable cancer patients are best managed by a team encompassing experts from critical care medicine, oncology, palliative care, and other fields. Previous studies indicate that while cancer patients’ ICU treatment outcomes differ significantly, tumor type, stage, or remission status minimally influence short-term ICU survival, mainly affecting long-term prognosis.25,26 This study employed ML to develop a reliable model for predicting 30-day survival of cancer patients in the ICU. Based on the indicators included in the model, the SHAP analysis further identified lymphocyte count, platelet count, lactate level, and ferritin as the 4 most relied-upon predictors by the model. This model combines iron metabolism, inflammatory markers (lymphocytes), coagulation function (platelets), tissue perfusion status (lactate), and various routine vital signs, laboratory tests, and comorbidity data, serving as an effective tool for clinicians to identify high-risk patients and predict their short-term survival probability accurately. As critical care technology advances, the prognostic value of certain clinical indicators previously deemed crucial for predicting poor outcomes has decreased. Neutropenia was once considered a major negative prognostic factor, but its effect on patient outcomes has notably diminished due to the use of potent antimicrobials and granulocyte colony-stimulating factor.27 This study found that neutrophils did not have significant prognostic value among immune and inflammatory indicators. Lymphocyte percentage, as an important indicator of cancer prognosis, has been confirmed in various tumor types. Studies in multiple cancers have suggested that low lymphocyte percentage is an independent indicator of poor prognosis.28,29 The study identified lymphocyte percentage as a protective factor for predicting long-term patient survival. For predicting in-hospital 30-day survival, lymphocyte percentage had the highest SHAP value in the model, indicating its greatest contribution to short-term survival prediction, suggesting its potential value for critically ill cancer patients. Regarding coagulation function indicators, platelet count is an important influencing factor for cancer patients, related to both tumor progression and drug treatment.30,31 This study identified platelet count as a protective prognostic factor for critically ill cancer patients, as demonstrated by both multivariate Cox regression and ML models. The model also suggested that prolonged prothrombin time indicates poor prognosis. Regarding metabolic and renal indicators, lactate accumulation is a direct manifestation of cancer cells’ reliance on aerobic glycolysis (Warburg effect). Stromal cells, including cancer-associated fibroblasts, contribute to significant lactate buildup in the tumor microenvironment due to elevated glycolytic activity. This acidic milieu not only aids tumor cell survival but also triggers gene programs linked to neuro-invasion, thereby promoting cancer’s invasive and metastatic potential. Therefore, high lactate levels indicate poor patient prognosis.32–34 In this study, both multivariate Cox regression and ML results indicated that high lactate is a strong risk factor in the model. Previous studies suggest that combined acute renal failure also indicates poor outcomes.35 The evaluation of critically ill cancer patients in this study identified creatinine and blood urea nitrogen as key indicators of renal function. Additionally, elevated anion gap also suggests poor prognosis. Regarding nutrition, low albumin reflects malnutrition, impaired liver synthetic function, or inflammation-induced capillary leakage and is strongly associated with poor prognosis, consistent with previous literature.36–38 The combination of reduced oxygen saturation, increased carbon dioxide partial pressure, lowered pH, and elevated respiratory rate indicates impaired ventilatory function and acid–base imbalance, suggesting a poor prognosis.39 Additionally, the use of vasoactive drugs marks extreme hemodynamic instability and critical illness, independently associated with poor prognosis, consistent with previous literature.40 Advanced age represents a poor prognostic factor,41 consistent with the article’s conclusions. Our findings confirm that iron metabolism markers—specifically ferritin and TIBC—are independently associated with mortality in critically ill cancer patients. Notably, the restricted cubic spline analysis revealed nonlinear relationships between these markers and mortality risk, suggesting that their prognostic impact cannot be adequately captured by linear models. This complexity supports the use of ML approaches, which are inherently designed to model such nonlinear associations. The function of ferritin in various cancers is intricate and sometimes conflicting: in some cases like esophageal adenocarcinoma and breast cancer,42,43 reducing its levels might increase intracellular labile iron, which can enhance tumor growth and epithelial–mesenchymal transition; conversely, systemic iron overload, such as in hereditary hemochromatosis, is linked to a higher risk of liver, colorectal, and breast cancers.44,45 High pre-transplant ferritin levels significantly predict decreased overall survival in patients with acute leukemia or myelodysplastic syndrome undergoing myeloablative hematopoietic stem cell transplantation.46 This study further strengthens the prognostic value of ferritin. Consistent findings from the Kaplan–Meier analysis, multivariate Cox regression, and ML models demonstrate that elevated ferritin levels independently increase mortality risk. Restricted cubic spline analysis also revealed a nonlinear dynamic relationship with mortality, suggesting the need to pay attention to its dynamic changes in clinical risk assessment. TIBC indicates TIBC of transferrin and serves as a reliable surrogate for transferrin concentration in clinical practice. Studies have clearly shown that TFR1 is a key molecular pathway for iron uptake in many cancer cells.8,47,48 In lung cancer patients, TIBC levels frequently change, while TFR1 is upregulated in several malignancies, such as breast, lung, and bladder cancer, to satisfy the elevated iron requirements of tumor cells. Previous research on lung and colon cancer indicates that elevated TIBC levels could serve as a protective prognostic factor for cancer patients.49,50 In this study, analyses based on multiple prediction models consistently found that high TIBC levels are associated with favorable clinical outcomes, demonstrating a clear protective effect, a result consistent with the aforementioned studies. The analysis indicated no significant link between serum iron levels and patient prognosis. This study has several limitations. First, the data originated from public databases, introducing inherent selection bias and unmeasured confounding despite the sizable sample. Second, iron metabolism markers were captured only at fixed time points, failing to account for their dynamic prognostic impact. Third, although the model demonstrated good discrimination in the internal validation cohort, a reduction in AUROC was observed during external validation. This decline may reflect the substantial heterogeneity between MIMIC-IV and eICU-CRD. The eICU-CRD database represents a multicenter ICU network, in which patient characteristics, disease severity, treatment patterns, laboratory testing practices, data collection procedures, and missing-data mechanisms may differ from those in the development cohort. These differences may lead to shifts in predictor distributions and reduce model transportability. Moreover, iron metabolism markers are not routinely assessed in all ICU patients, and their measurement may depend on clinical judgment, illness severity, and institutional practice, resulting in selective measurement and nonrandom missingness. These factors may partially explain the attenuated performance in the external validation cohort. Nevertheless, the model retained an AUROC of 0.772 in the external validation cohort, indicating acceptable discrimination in an independent real-world multicenter population. Future studies should further validate, recalibrate, and update this model using prospective multicenter cohorts across diverse healthcare systems and sociodemographic subgroups. In clinical deployment, missing predictors should be handled using imputation approaches as adopted here, and the model should serve as a decision-support tool alongside clinician judgment. Conclusion In conclusion, iron metabolism dysregulation—particularly elevated ferritin and reduced TIBC—is closely associated with mortality risk in critically ill cancer patients. By integrating these biologically relevant markers with routine clinical variables within a TabPFN framework, this study developed an accurate, interpretable, and externally validated model for 30-day all-cause mortality prediction. The SHAP analysis further identified ferritin, lymphocyte count, platelet count, and lactate as the most influential predictors, highlighting the model’s alignment with clinical pathophysiology. These findings support a biologically informed, data-driven approach to precision prognostic assessment in this high-risk population. Further multicenter prospective studies are warranted to validate its generalizability and to evaluate the prognostic implications of dynamic changes in iron metabolism markers. Supplementary Material