A sequential machine learning framework for ICU and hospital length of stay prediction from admission-time data

Abstract

Objectives

Accurate prediction of hospital length of stay (LOS) is essential for effective patient flow management, resource allocation, and improved clinical outcomes, particularly in resource-intensive settings such as intensive care units (ICUs). Existing LOS prediction models often rely on post-admission data or focus on homogeneous patient populations, which limits their utility for early operational planning. This study aims to develop a sequential machine-learning framework to predict total hospital LOS using admission-time information. The framework predicts ICU admission and ICU duration as intermediate outcomes, which are then incorporated as features to improve total LOS prediction while also providing standalone operational value for early resource planning.

Methods

We propose a sequential prediction framework that decomposes LOS prediction into staged tasks, enabling early and interpretable clinical and operational decision support. The framework was validated using a large dataset consisting of 1,104,203 patient admissions from 52 Maryland hospitals, including 20 admission diagnoses and 29 comorbidities. Deep learning and classical machine learning models were employed for both classification and regression tasks, and performance was compared with traditional single-stage LOS prediction approaches.

Results

The XGBoost classifier predicted ICU admission with an accuracy of 0.73, sensitivity of 0.76, and specificity of 0.72. Incorporating ICU-specific predictions improved total hospital LOS prediction performance. For LOS classification, the proposed framework achieved an F1 score of 0.701 (a 5.1% relative improvement) and a macro AUC-ROC of 0.76 (a 1.9% relative improvement) compared with single-stage models. Regression analysis resulted in a mean absolute error of 2.62 days for LOS prediction.

Conclusion

The proposed sequential, ICU-informed machine-learning framework enhances the accuracy and interpretability of total hospital LOS prediction using admission-time data. By explicitly modeling ICU admission and ICU LOS as intermediate outcomes, the framework provides actionable insights to support early patient flow management, resource utilization, and operational planning within hospital systems.

Keywords

healthcare management length of stay sequential modeling machine learning ICU admission hospital information systems

Introduction

Hospital length of stay (LOS) reflects both patient severity and hospital efficiency, and measures the time from admission to discharge across units like the intensive care unit (ICU), critical care unit (CCU), and surgical wards.¹ Both unnecessary long or insufficiently short stays can be problematic for patient outcomes and hospital operations. Prolonged stays raise the risk of infections,² miscommunication due to more provider hand-offs,³ and reduced capacity,⁴ especially during crises like pandemics.^5,6 Conversely, short stays may lead to inadequate care and higher readmission risk. Accurately predicting LOS is therefore critical not only for patient care but also for operational planning, as it supports more efficient bed utilization and staff scheduling, helping to improve system-level resource management.⁷

The hospital LOS, depends on a combination of patient-specific and hospital-related factors. From the patient side, individuals with more severe conditions are generally expected to stay longer. While primary diagnosis influences LOS, it is not always the main driver. For example, a patient admitted with intracerebral hemorrhage (ICH) may have prolonged LOS due to a secondary condition like acute kidney injury (AKI). Thus, LOS prediction should consider all diagnoses and comorbidities at admission to be accurate and not misleading. On the hospital side, operational factors like discharge delays, insurance approvals, and hospital capacity all affect LOS. Lack of beds in downstream units prolongs ICU stays and reduces hospital throughput. These complexities motivate our goal to improve LOS prediction at admission by addressing limitations found in prior research. Most LOS studies focus on homogeneous groups, limited to diagnoses like cardiology, diabetes, or neurology.^8–11 In contrast, fewer studies address heterogeneous populations.¹² Studies also vary in whether they treat LOS as a regression or classification task, in the thresholds used for classification, and in the choice of predictive methods and evaluation metrics.

Among studies on homogeneous populations, Hachesu et al.¹³ classified LOS for 2,064 cardiology patients into three groups and achieved 96.4% accuracy using neural networks, support vector machines (SVMs), and decision trees. Tsoukalas et al.¹⁴ used SVMs to classify LOS in 1,492 sepsis patients with thresholds at 4, 8, and 12 days, reporting 0.69–0.82 accuracy and 0.69–0.73 area under curve (AUC). Turgeman et al.¹⁵ applied regression trees to 20,321 cardiology patients, resulting a mean absolute error (MAE) of 1 and coefficient of determination (R²) of 0.79. For heterogeneous populations, Livieris et al.¹⁶ used semi-supervised models on 4,403 inpatients and achieved 63.23–65.30% accuracy. Baek et al.¹⁷ used a 30-day LOS threshold and reported 0.9732 accuracy and 4.68 MAE. Tully et al.¹⁸ applied an XGBoost classifier with SMOTE balancing to predict whether ambulatory surgical patients would stay in the post-anesthesia care unit for three hours or longer, and the model achieved an AUC of 0.712. Rajkomar et al.¹⁹ applied RNNs across six disease groups, achieving 85–86% AUC, these results are hospital-specific with using the data obtained within 24 hours of admission. For regression tasks, R² scores vary: 0.554 Cui et al.,¹² 0.2 Saly et al.,²⁰ and 0.146 Liu et al.²¹ MAEs range from 2.19 to 3.51 days, with root mean squared error (RMSE) of 3.10 reported by Cui et al.¹²

Recent studies have applied machine learning to predict ICU admission (i.e., whether a patient will require ICU care during hospitalization) and ICU length of stay (ICU-LOS), addressing challenges such as non-linearity, class imbalance, and diverse clinical features.²² For example, Peres et al.²³ used a stacking model on 99,492 ICU admissions and achieved an RMSE of 3.82 and AUC of 0.87. In that cohort, 68.6% of patients were admitted through the emergency department, and 52.8% were female. Another recent work Brochini et al.²⁴ modeled ICU and hospital LOS separately for survivors and nonsurvivors but relied on post-admission data. Ma et al.²⁵ proposed a JITL-ELM model using 48-hour physiological data, achieving an AUC of 0.8510. Similarly, Iwase et al.²⁶ used 91 variables within 24 hours post-ICU admission and reported high AUCs for mortality and LOS prediction using Random Forest. Navaz et al.,²⁷ using MIMIC-II data, applied decision tree methods and reported variable accuracy across LOS categories. Lan et al.²⁸ highlights the value of ML in early risk stratification and ICU planning, and developed machine learning models using preoperative data to predict prolonged ICU admission (>24h) after elective non-cardiac surgery. Using a cohort of over 135,000 patients, their gradient boosting model achieved a high performance (AUC-ROC=0.90).

While past studies report good results, they often rely on post-admission data or focus on single diagnoses,^8–11 which limits their generalizability and practical utility. Some models were also trained using data from a single hospital,^16,17 further reducing their ability to generalize across different care settings. This issue of poor generalizability was also emphasized by Gokhale et al.,²⁹ who, after systematically reviewing 15 studies on LOS prediction in general surgery and knee arthroplasty patients, highlighted that only a few studies employed machine learning methods, along with the lack of external validation, as key shortcomings of existing models. In contrast, recent evidence shows that models trained on multi-institutional datasets achieve significantly better generalization Rockenschaub et al.³⁰ To address these gaps, our approach uses only admission-time information, which provides actionable insights at the earliest stage of care. We further disaggregate length of stay (LOS) to capture the effects of unit-level capacity constraints, incorporate 20 admission diagnoses and 29 comorbidities to reflect patient clinical complexity, and leverage data from 52 hospitals to enhance the generalizability of our findings. Together, these features make our approach a more practical tool for early decision-making and system-level hospital management.

Methodology

Our proposed framework (Figure 1) predicts total hospital LOS through three sequential stages: (1) ICU admission prediction, (2) ICU-LOS prediction, and (3) total LOS prediction. The main goal of this framework is to predict total hospital LOS. The model uses a sequential structure in which each stage builds on the predictions from the previous stage. First, ICU admission is predicted, then ICU length of stay (ICU-LOS), and finally total hospital LOS. Because ICU admission and ICU-LOS strongly affect overall hospital stay, incorporating these intermediate predictions improves the accuracy of LOS prediction. In addition, each stage provides operational insights for ICU bed allocation, staffing planning, and discharge coordination, so the intermediate predictions also support clinical and operational decision-making.

Figure 1.

Sequential modeling framework for length of stay prediction. (ICU: Intensive care unit; LOS: Length of stay).

Data preprocessing

All post-admission information was excluded from the dataset. Specifically, we retained only pre-admission data, historical records, and admission diagnoses. The preprocessing steps were as follows:

• Observations with ICU-LOS or total LOS exceeding 60 days (0.27%) were removed.

• Diagnoses or procedures observed in fewer than 0.5% of cases were excluded.

• Categorical variables were one-hot encoded; age was standardized using Z-score normalization.

• Features with both low SHAP importance and high multicollinearity (VIF $>$ 20) were removed.

• Missing values in categorical variables were retained as a separate category prior to one-hot-encoding. Missingness in binary variables were replaced with 0, and rows with missing LOS or AGE values were removed due to their negligible frequency (< 0.1%).

After preprocessing, the final dataset comprised 1,104,203 records and 122 variables, including ICU, ICU-LOS, and total LOS.

Data splitting strategy

To ensure model generalizability and prevent data leakage across stages, we implemented a structured multi-stage data splitting strategy (Figure 2). The dataset was partitioned into four non-overlapping subsets: Subset I (20%) is the training data for Stage 1, Subset II (20%) used as the training data in Stage 2, Subset III.a (30%) is the training data of Stage 3, and Subset III.b (30%) is held-out as the test set for final LOS evaluation. The test set remained strictly isolated throughout all feature selection, hyperparameter tuning, and model training processes across all three stages. The hyperparameter tests were conducted using the respective training subsets. During the final evaluation, the test set (Subset III.b) was processed sequentially through the trained pipeline to obtain unbiased estimates of generalization performance.

Figure 2.

Data splitting approach to prevent data leakage and overfitting.

Stage 1 – ICU admission prediction

The objective is to predict whether a patient will require ICU admission using only admission data (Figure 1).

Input features: 119 admission variables (ICU, ICU-LOS, and total LOS are excluded)

Output: Trained binary classification model (ICU = 0 or 1).

Process:

1. Train classification model on Subset I

2. Generate predicted ICU class (ICU_p) for Subsets II, III.a, and III.b

3. Append ICU_p as a new feature to these subsets

Since the ICU admission variable is highly imbalanced (12.55% positive class). We evaluated SMOTE, up-sampling, and under-sampling techniques; general up-sampling provided optimal performance and was applied exclusively to training data. The models evaluated at this stage include Logistic Regression, AdaBoost, Random Forest, XGBoost, and neural networks with varying architectures.

Stage 2 – ICU-LOS prediction

The objective is to predict the duration of ICU stay using admission data augmented with ICU_p (Figure 1).

Input features: 120 variables (119 original + ICU_p from Stage 1)

Output: ICU-LOS prediction model and predicted ICU duration in days (DaysICU_p) on the training set.

Process:

1. Augment Subset II with ICU_p predictions from Stage 1

2. Train prediction model on augmented Subset II

3. Generate predicted ICU-LOS (DaysICU_p) for Subsets III.a and III.b

4. Append DaysICU_p as a new feature to these subsets

We applied both classification and regression approaches for the DaysICU_p prediction. To avoid error propagation, we trained the ICU-LOS model on all patients, including those predicted as non-ICU in Stage 1. Setting DaysICU_p to zero for predicted non-ICU cases would carry forward misclassification errors from Stage 1. We evaluated linear, tree-based, and neural network models for all prediction tasks. For the classification approach, we defined thresholds to group patients based on the duration of their ICU stay: class 1: ICU-LOS ≤ 1, class 2: 1 < ICU-LOS ≤ 3 days, class 3: ICU-LOS > 3 days.

Stage 3 – Total LOS prediction

The objective is to predict total hospital LOS using pre-admission data augmented with ICU_p and DaysICU_p (Figure 1).

Input features: 121 variables (119 original + ICU_p + DaysICU_p)

Output: LOS prediction model

Process:

1. Augment Subset III.a and III.b with ICU_p and DaysICU_p predictions

2. Train LOS prediction model on augmented Subset III.a

3. Evaluate on augmented Subset III.b (held-out test set)

We implemented both regression and classification approaches because each serves a different purpose: regression offers a more granular estimate of LOS, while classification provides operationally meaningful LOS categories that are directly useful for capacity planning and resource management. In our study, LOS is categorized as follows: class 1: LOS ≤ 7 days, class 2: 7 < LOS ≤ 14 days, and class 3: LOS > 14 days. General up-sampling.³¹ was applied to minority classes in the training set. Models evaluated include Logistic Regression, AdaBoost, Random Forest, XGBoost, and neural networks with varying architectures for the classification approach; and Ridge Regression, XGBoost, Random Forest, AdaBoost, and neural networks with varying architectures for the regression approach.

This sequential, modular design allows each stage to contribute valuable predictive information and helps to build a more accurate and interpretable model for hospital LOS prediction. All modeling experiments were repeated five times with different random seeds and data shuffling to account for variability arising from random parameter initialization. Performance metrics are reported as averages across runs.

Worked example

To illustrate the framework’s operation, we present a step-by-step example using a hypothetical patient case (Table 1).

Table 1.

Worked example: Sequential LOS prediction for a sample patient.

Stage	Input	Output
Patient profile at admission:
67-year-old female, emergency admission, primary diagnosis: heart failure (ICD: I50.9), comorbidities: diabetes mellitus, hypertension, chronic kidney disease, anemia deficiency
Stage 1	119 pre-admission features including: Age=67, Female=1, ATYPE=Emergency, DX_I509=1, DMCX=1, HTN=1, RENLFAIL=1, ANEMDEF=1, …	ICU_p = 1 (binary classification)
Stage 2	120 features: Original 119 features + ICU_p = 1	DaysICU_p = 4.2 days (regression)
Stage 3	121 features: Original 119 features + ICU_p = 1 + DaysICU_p = 4.2	LOS_p = 3 (class 3: $>$ 14 days) Class probabilities: P(Class 1) = 0.15 P(Class 2) = 0.28 P(Class 3) = 0.57
Clinical interpretation:
Based on the patient’s admission characteristics (including age, emergency admission, heart failure diagnosis, and multiple comorbidities) the framework predicts the need for ICU admission with an expected ICU stay of 4.2 days. Incorporating these predictions as new features to the data of patient, the final model predicts an extended hospital stay exceeding 14 days with 57% probability. This early prediction enables proactive resource allocation, discharge planning, and care coordination.

The sequential flow of predictions can be expressed mathematically as:

\begin{align} I C U_{p} = f_{1} (X) \end{align}

(1)

\begin{align} D a y s I C U_{p} = f_{2} (X, I C U_{p}) \end{align}

(2)

\begin{align} L O S_{p} = f_{3} (X, I C U_{p}, D a y s I C U_{p}) \end{align}

(3)

where X represents the original 119 pre-admission features, and f₁, f₂, f₃ denote the trained models for each stage.

Case study and results

We used the 2017–2018 Maryland State Inpatient Databases (SID), Healthcare Cost and Utilization Project (HCUP), Agency for Healthcare Research and Quality.³² The dataset includes patient demographics, socioeconomic data, ICD-10 codes, ICU admissions, ICU-LOS, and total LOS from 52 Maryland community hospitals. Additionally, it considers five groups of variables: demographics, administrative data, admission and comorbidities, and procedures. A detailed list of variables is provided in the supplementary material.

Figure 3 shows the total hospital LOS distribution for all patients, and Figure 4 shows both total hospital LOS and ICU-LOS distributions for ICU patients. Both figures are truncated at 30 days for visual clarity, as cases exceeding this duration represent a small fraction of the dataset and do not meaningfully contribute to the distributional patterns. Across all patients (n=1,104,203), the median age was 58 years, 57% were female, and the average hospital LOS was 4.79 days (SD: 5.57, median: 3), with a right-skewed distribution (skewness: 3.88). ICU admission occurred in 12.55% of patients (n=138,589). Among ICU-admitted patients, the median age was higher at 61 years, 47% were female, and the average hospital LOS increased substantially to 8.06 days (SD: 8.69, median: 5, skewness: 2.52), demonstrating the significant impact of ICU care on overall hospital stay duration. In comparison, non-ICU patients had an average LOS of 4.32 days (SD: 4.78, median: 3).

Figure 3.

Distribution of total LOS for all patients admitted to hospitals.

Figure 4.

Distribution of total LOS and ICU-LOS in ICU patients group.

For ICU-LOS specifically, ICU-admitted patients spent an average of 4.31 days in the ICU (SD: 6.15, median: 2, skewness: 3.99), with the remaining 3.76 days on average spent in other hospital units. The right-skewed distribution indicates that while most ICU patients had brief stays, a subset required prolonged intensive care. Based on this distribution and clinical considerations, we defined three ICU-LOS categories: ICU-LOS ≤ 1 day (34% of ICU patients, n=47,256), representing brief observation or post-operative stabilization; 1 < ICU-LOS ≤ 3 days (33%, n=45,764), representing moderate ICU stays requiring active critical care management; and ICU-LOS > 3 days (33%, n=45,569), representing prolonged intensive care associated with increased complications, higher mortality risk, and substantial resource utilization. The relatively uniform distribution across these three categories provides balanced class representation for predictive modeling while maintaining clinical interpretability.

For total hospital LOS, we selected thresholds of 7 and 14 days. The 7-day threshold is widely employed in LOS prediction studies, and stays longer than 7 days are considered to be associated with increased post-discharge adverse outcomes and complications (^33–35). The 14-day threshold is also well-established for defining prolonged hospitalization, commonly chosen in literature as it marks the point where patients face substantially increased risks of hospital-acquired complications such as infections, delirium, and functional decline, while simultaneously requiring intensified resource utilization including extended nursing care, specialist consultations, and complex discharge planning.^36,37 This resulted in three categories: LOS ≤ 7 days (84% of patients, n=931,228); 7–14 days (11%, n=117,895); and LOS > 14 days (5%, n=55,080).

Figure 5 shows the prevalence and LOS for various admission diagnosis for patients in the ICU and hospital. The lengths of the bars represent the average LOS for each condition and the percentages represent the distribution of diagnoses. The longest ICU and hospital stays are associated with weight loss (WGHTLOSS), pulmonary circulation disorders (PULMCIRC), and paralysis (PARA), with WGHTLOSS linked to the highest LOS overall. The most common ICU diseases are hypertension with complications (HTN-C), electrolyte imbalances (LYTES), and anemia (ANEMDEF), with HTN-C affecting 54.7% of ICU and 48.5% of hospital patients. HTN-C patients average 3.84 ICU days and 5.25 hospital days. This chart provides a clear picture of which admission diagnoses lead to longer hospital or ICU stays and affect more patients, which highlights the need for targeted care and prevention strategies.

Figure 5.

Comparative analysis of ICU and hospital LOS by comorbidities prevalence. Dark blue bars (left) represent the average ICU length of stay (LOS) for patients with each admission diagnosis; light blue bars (right) show the average hospital LOS. Percentages on the bars indicate the prevalence of each diagnosis in the ICU (left) and hospital (right) populations.

Having identified the diagnoses associated with longer ICU and hospital stays, we next focus on building predictive models for ICU admission and LOS. We evaluated seven classification models for predicting ICU admission: Logistic Regression, AdaBoost, Random Forest, XGBoost, and three neural network architectures of increasing complexity (ICU-1, ICU-2, ICU-3), ranging from 7 to 11 layers to capture different levels of feature interactions. Hyperparameters for all models were tuned using grid search. The full list of hyperparameters evaluated and the detailed neural network configurations are provided in Tables 8 and 9 in the Appendix.

In Stage 1, models performance evaluation for ICU admission (Table 2) shows that XGBoost achieved the highest discrimination ability with a test AUC-ROC of 0.8351, outperforming Neural Network ICU-1 (0.8294) and ICU-2 (0.8326). Logistic Regression, included as a linear baseline, achieved 0.7989, confirming that nonlinear models provided better results. XGBoost also demonstrated a balanced clinical performance with sensitivity of 0.77 and specificity of 0.725. Confusion matrix and calibration analysis are provided in the Appendix, Figures 6 and 7. After identifying the best-performing model in Stage 1, we used it in next stages to predict ICU admission (ICU_p).

Table 2.

Stage 1 results: ICU admission classification performance (positive class: ICU admitted). Sensitivity measures correct identification of ICU cases, while specificity measures correct identification of non-ICU cases.

Model	AUC-ROC	Accuracy	Sensitivity	Specificity	Precision	F1-score
Logistic Regression	0.7989	0.6762	0.7741	0.6589	0.2476	0.3751
AdaBoost	0.7912	0.6845	0.7506	0.6702	0.2493	0.3714
Random Forests	0.8107	0.7196	0.7204	0.7169	0.2689	0.3923
XGBoost ^a	0.8351	0.7314	0.7699	0.7251	0.2867	0.4182
NN (ICU-1)	0.8294	0.6752	0.8356	0.6564	0.2598	0.3957
NN (ICU-2)	0.8326	0.7129	0.7911	0.7014	0.2761	0.4098
NN (ICU-3)	0.8221	0.6878	0.8053	0.6712	0.2619	0.3942

^aSelected model for downstream pipeline.

In Stage 2, we evaluated multiple regression models for predicting ICU-LOS using 20% of the data for training (Subset II) and 60% for testing (Subset III). The results provided on Table 3 show that XGBoost achieves the lowest MAE (0.6) among evaluated models. In our multi-stage prediction pipeline, where predicted ICU-LOS is used as an engineered feature for total hospital LOS, reducing prediction error is important. Lower MAE directly translates to less noise introduced into the downstream model, preventing compounding errors across stages. Therefore, XGBoost, which achieves the lowest MAE, is the suitable choice for regression task at this stage. Residual and actual-versus-predicted plots for the ICU-LOS regression model are shown in Appendix (Figure 11) and illustrate the challenges associated with modeling highly skewed ICU-LOS outcomes, including underestimation of long ICU stays, low explained variance, and increasing prediction residuals for patients with prolonged ICU stays.

Table 3.

Stage 2 results: ICU-LOS regression.

Model	MAE	MSE	RMSE	R²	MAPE (%)
Ridge Regression	0.6332	6.7294	2.5941	0.0158	82.05
XGBoost ^a	0.6009	6.5486	2.5590	0.0423	80.74
Random Forest	0.6013	6.5722	2.5636	0.0388	81.52
Gradient Boosting	0.6051	6.5957	2.5682	0.0354	81.40
AdaBoost	0.8068	6.4952	2.5486	0.0501	63.93
NN (DICU-R1)	0.6121	6.7133	2.5910	0.0182	83.94
NN (DICU-R2)	0.6347	6.6686	2.5826	0.0246	80.05
NN (DICU-R3)	0.6152	6.6684	2.5825	0.0246	81.97
NN (DICU-R4)	0.6173	6.7214	2.5924	0.0171	83.90

^aSelected model for the pipeline.

We also explored classification models across three classes: Class 1 (ICU-LOS ≤ 1 day), Class 2 (1 < ICU-LOS ≤ 3 days), and Class 3 (ICU-LOS > 3) for ICU-LOS prediction. We performed hyperparameter optimization and architecture search for XGBoost, Random Forest, AdaBoost, Logistic Regression, and neural network architectures (details in Appendix Tables 10 and 11) using grid search.

Table 12 provided in the Appendix summarizes the ICU-LOS classification results. Among all models, XGBoost shows the best discrimination ability with an AUC-ROC of 0.821 and a balanced accuracy of 0.571, and the best trade-off between sensitivity and specificity across all ICU-LOS categories, outperforming other traditional machine learning models and neural networks. Accordingly, XGBoost (max depth=6, learning rate=0.3, n-estimators=150) was selected for the cascaded prediction pipeline. In order to assess the impact of using our best classification model and best regression model in Stage 2 on the subsequent total LOS prediction model (Stage 3), we conducted a comparative analysis. For this purpose, we fixed the prediction models at stages 1 and 3 (XGBoost) to directly evaluate the effects of employing a regression model versus a classification model at stage 2 on the total LOS prediction (Stage 3). Results provided on Table 4 show both approaches yield similar results, with regression approach showing marginally better results across all metrics. Given the comparable performance, we selected the regression model for Stage 2 because continuous ICU-LOS predictions (DaysICU_p) preserve more granular information about expected ICU duration, which can be more informative for downstream LOS prediction than discrete categorical labels.

Table 4.

Stage 3 results: Comparative analysis of XGBoost (classification) and XGBoost (regression) models for ICU-LOS prediction (Stage 2). The models used in stages 1 and 3 are fixed. Class 1: LOS ≤ 7 days; class 2: 7–14 days; class 3: $> 14$ days.

		Stage 1: XGBoost	Stage 1: XGBoost
		Stage 2: XGBoost (Classification)	Stage 2: XGBoost (Regression)
		Stage 3: XGBoost	Stage 3: XGBoost
	Macro AUC-ROC	0.7628	0.7638
	Overall Accuracy	0.6373	0.6406
	Weighted F1-score	0.6999	0.7015
Class 1	Sensitivity	0.6711	0.6787
	Specificity	0.7349	0.7298
	Precision	0.9321	0.9311
	F1-Score	0.7833	0.7851
Class 2	Sensitivity	0.3802	0.3863
	Specificity	0.7943	0.7938
	Precision	0.1829	0.1829
	F1-Score	0.2481	0.2483
Class 3	Sensitivity	0.5507	0.5419
	Specificity	0.8564	0.8602
	Precision	0.1676	0.1691
	F1-Score	0.2570	0.2578

In Stage 3, after adding the ICU_p and DaysICU_p values, we evaluated LOS prediction with both classification and regression approaches. Hyperparameter configurations tested for all models are provided in the Appendix ( Tables 13 and 14). All models are trained on 30% of data (Subset III.a) and tested on 30% (Subset III.b).

The results show that incorporating the Stage 2 regression output into the Stage 3 model leads to higher performance improvements across most metrics. We achieved slightly higher macro AUC-ROC (0.7638 vs. 0.7628), balanced accuracy (0.5356 vs. 0.5340), overall accuracy (0.6406 vs. 0.6373), and weighted F1-score (0.7015 vs. 0.6999). Calibration analysis (Figure 8) and confusion matrix (Figure 9) are provided in the Appendix.

Table 5 presents the total LOS classification performance, for Stage 3, across all models. XGBoost achieved the highest Macro AUC-ROC (0.764), outperforming Logistic Regression (0.755) and the neural network architectures (0.737—0.743). The severe class imbalance, with 84% of patients in Class 1 (LOS ≤ 7 days), 11% in Class 2 (7—14 days), and only 5% in Class 3 (LOS > 14 days), posed significant challenges for all models. While all models achieved high sensitivity and precision for the majority class (Class 1), performance on minority classes was substantially lower. For Class 2, sensitivity ranged from 0.253 to 0.407, while Class 3 sensitivity ranged from 0.413 to 0.644. Notably, Random Forests achieved the highest overall accuracy (0.682) and weighted F1-score (0.729), but this was driven primarily by strong majority class performance rather than balanced discrimination. XGBoost demonstrated the best trade-off between overall discrimination (AUC-ROC) and minority class detection, achieving the highest F1-scores for both Class 2 (0.248) and Class 3 (0.258) among traditional models. Calibration analysis (Figure 12) and confusion matrix (Figure 10) are provided in the Appendix. The confusion matrix (Figure 10) shows results from a single run of code and confirms that misclassifications primarily occur in longer LOS categories, with the model achieving 67.6%, 38.5%, and 55.1% sensitivity for short, medium, and long-stay classes, respectively.

Table 5.

Stage 3 results: Total LOS classification performance. Class 1: LOS ≤ 7 days (84%); class 2: 7–14 days (11%); class 3: $> 14$ days (5%).

		Logistic reg.	XGBoost^a	AdaBoost	Random forests	LOS-C1	LOS-C2	LOS-C3
	Macro AUC-ROC	0.7552	0.7638	0.7330	0.7555	0.7375	0.7430	0.7401
	Overall Accuracy	0.6360	0.6406	0.5988	0.6823	0.6612	0.6359	0.5391
	Weighted F1-score	0.6977	0.7015	0.6694	0.7293	0.7159	0.6984	0.6211
Class 1	Sensitivity	0.6764	0.6787	0.6344	0.7330	0.7089	0.6839	0.5480
	Specificity	0.6110	0.5984	0.5642	0.6029	0.5736	0.5568	0.5891
	Precision	0.9277	0.9311	0.9256	0.9191	0.9252	0.9300	0.9454
	F1-Score	0.7824	0.7851	0.7528	0.8156	0.8027	0.7882	0.6938
Class 2	Sensitivity	0.3677	0.3863	0.3467	0.4075	0.3707	0.2527	0.4483
	Specificity	0.8531	0.8467	0.8612	0.8578	0.8721	0.8934	0.8316
	Precision	0.1787	0.1829	0.1598	0.1951	0.1610	0.1726	0.1425
	F1-Score	0.2405	0.2483	0.2188	0.2639	0.2245	0.2051	0.2163
Class 3	Sensitivity	0.5261	0.5419	0.5365	0.4129	0.4756	0.6442	0.5838
	Specificity	0.9148	0.9193	0.9171	0.9240	0.9087	0.8895	0.9013
	Precision	0.1587	0.1691	0.1406	0.1972	0.2197	0.1438	0.1661
	F1-Score	0.2438	0.2578	0.2228	0.2669	0.3005	0.2351	0.2586

^aSelected model.

Table 6 presents the total LOS regression performance in Stage 3. We evaluated Ridge Regression, Random Forest, XGBoost, Gradient Boosting, AdaBoost, and three neural network architectures with varying depths (3 to 9 layers). XGBoost achieved the best overall performance with the lowest MAE (2.622), RMSE (2.622), and MAPE (65.20%), along with the highest R² (0.159). Gradient Boosting and neural networks LOS-R1 to LOS-R3 showed competitive results (MAE = 2.64, R² = 0.145), while AdaBoost performed notably worse (R² = 0.059, MAPE = 89.20%). Actual versus predicted total LOS and residual plot for the XGBoost regression model is provided in Appendix Figure 13.

Table 6.

Stage 3 results: Total LOS regression results on the test set with log-transformed LOS.

Model	MAE	MSE	RMSE	R²	MAPE (%)
Ridge Regression	2.6713	5.1710	2.6713	0.1301	66.90
XGBoost ^a	2.6215	5.0841	2.6215	0.1591	65.20
Random Forest	2.6642	5.1549	2.6642	0.1355	66.48
Gradient Boosting	2.6368	5.1279	2.6368	0.1446	65.63
AdaBoost	2.9878	5.3779	2.9878	0.0591	89.20
NN (LOS-R1)	2.6378	5.1262	2.6378	0.1451	67.01
NN (LOS-R2)	2.6910	5.1351	2.6910	0.1422	72.40
NN (LOS-R3)	2.7360	5.3696	2.7360	0.0620	67.93

^aBest model performance.

With the best models selected at each stage, we next evaluate the overall impact of our sequential approach by comparing total LOS prediction across three scenarios, outlined in Table 7.

Table 7.

Comparison of total LOS classification performance across modeling scenarios using XGBoost (depth=6, lr=0.3). Scenario A excludes ICU and DaysICU information, scenario B incorporates predicted ICU and predicted DaysICU, and scenario C includes real ICU and DaysICU values directly. All scenarios are trained on subset III.a (30%) and tested on subset III.b (30%).

		Scenario A	Scenario B	Scenario C
		No ICU/DaysICU	ICU_p + DaysICU_p	With ICU + DaysICU
	Macro AUC-ROC	0.7496	0.7638	0.7859
	Overall Accuracy	0.6198	0.6406	0.6719
	Weighted F1-score	0.6672	0.7015	0.7074
Class 1	Sensitivity	0.6701	0.6787	0.7240
	Specificity	0.5644	0.5631	0.5210
	Precision	0.8989	0.9311	0.9026
	F1-Score	0.7678	0.7851	0.8035
Class 2	Sensitivity	0.3810	0.3863	0.4525
	Specificity	0.3363	0.3316	0.2864
	Precision	0.2318	0.1829	0.2598
	F1-Score	0.2882	0.2483	0.3301
Class 3	Sensitivity	0.5557	0.5419	0.5374
	Specificity	0.3634	0.3581	0.3074
	Precision	0.2072	0.1691	0.3106
	F1-Score	0.3018	0.2578	0.3937

Note. Training Scenario A on combined Subsets I, II, and III.a (70% of data) achieved Macro AUC-ROC of 0.760, overall accuracy of 0.616, and weighted F1-score of 0.670, which remain lower than the performance achieved in Scenario B.

Table 7 compares LOS classification performance across three modeling scenarios to evaluate the contribution of ICU-related information. Scenario A, which excludes ICU and DaysICU features entirely, serves as the baseline with a Macro AUC-ROC of 0.750. Scenario B incorporates the predicted ICU status (ICU_p) and predicted ICU duration (DaysICU_p) from the cascaded pipeline, resulting in an improvement in overall accuracy to 0.64 (95% confidence interval (CI): 0.639–0.642), a 5.1% relative improvement compared to Scenario A, and an increase in macro AUC-ROC to 0.764 (95% CI: 0.763–0.767), corresponding to a relative improvement of 1.9%. Scenario C, which uses actual ICU and DaysICU values, represents the theoretical upper bound with a Macro AUC-ROC of 0.786. Notably, our cascaded approach (Scenario B) captures approximately 39% of the potential improvement between having no ICU information and having perfect ICU information, computed as (0.764 − 0.750)/(0.786 − 0.750). The performance gains are consistent across metrics: overall accuracy improved from 0.620 (Scenario A) to 0.641 (Scenario B), and weighted F1-score increased from 0.667 to 0.702 (95% CI: 0.700–0.703).

In all three scenarios, the LOS model is trained on Subset III.a and tested on Subset III.b. From this perspective, Scenario B benefits from using additional data (Subsets I and II) to create informative intermediate features. This raises an important question: would using all available data, Subsets I, II, and III.a, for direct LOS prediction (Scenario A) outperform the sequential approach (Scenario B), which instead allocates Subsets I and II to generate intermediate features? To answer this, we trained the LOS model on the combined dataset (70% of data) without intermediate predictions and compared it against Scenario B, with both evaluated on Subset III.b. The results show that the larger training set achieved a Macro AUC-ROC of 0.760, overall accuracy of 0.616, and weighted F1-score of 0.670. While this represents an improvement over Scenario A trained on Subset III.a alone (Macro AUC-ROC: 0.750, overall accuracy: 0.620, weighted F1-score: 0.667), it remains below the performance of Scenario B (Macro AUC-ROC: 0.764, overall accuracy: 0.641, weighted F1-score: 0.702). These results demonstrate that while additional training data provides modest gains, the cascaded approach that strategically uses this data to generate informative intermediate features (ICU_p and DaysICU_p) achieves superior predictive performance. This validates our framework design that allocating data to create informative features yields greater benefit than simply expanding the training set for direct LOS prediction, even when these new features are imperfect.

To validate the contribution of the predicted ICU-related features, we conducted feature importance analysis using both XGBoost gain scores and SHAP values provided in the Appendix, Table 15. The analysis shows that ICU_p and DaysICU_p rank among the top 10 most influential features (SHAP ranks 8 and 6, respectively). This confirms that the cascaded pipeline provides meaningful complementary information for LOS prediction.

Model performance at Stage 3 was evaluated across demographic subgroups to assess equity. The model demonstrated consistent performance across racial groups, with Macro AUC-ROC ranging from 0.760 to 0.805. Greater variation was observed across age groups. The model performed best for young adults (18–44 years, AUC 0.808) who had the lowest rate of prolonged stays (2.8%). In contrast, pediatric (0–17 years, AUC 0.708) and elderly patients (80+ years, AUC 0.723) showed lower model performance and higher rates of prolonged LOS (7.1% and 4.7%, respectively). Similarly, the model achieved higher AUC for females (0.787) compared to males (0.732), corresponding to lower prolonged LOS rates in females (4.2% vs 6.1%). Subgroups with more prolonged stays are harder to predict due to greater clinical heterogeneity and the difficulty of predicting rare outcomes in imbalanced distributions. Despite these variations, the model maintained acceptable discrimination (AUC >0.70) across all subgroups.

Discussion

Accurate prediction of ICU outcomes at hospital admission serves multiple critical functions in modern healthcare delivery. As emphasize in the work by Power et al.,³⁸ ICU outcome prediction models are essential for benchmarking institutional performance, supporting quality improvement initiatives, and enabling risk-adjusted comparisons across healthcare facilities with differing case mixes. By generating early, admission-time LOS estimates, our framework further supports hospital-level benchmarking efforts, such as standardized length of stay ratio (SLOSR) efficiency metrics,³⁹ by enabling fair comparisons between observed and expected resource utilization across institutions. Beyond benchmarking, LOS prediction at ICU admission enables proactive resource allocation, facilitates communication with families regarding expected trajectories, and supports discharge planning to optimize patient flow.⁴⁰ Building on prior identification of clinical factors that drive total LOS such as mechanical ventilation and hypomagnesemia that account for longer stays,⁴¹ we further identify diagnosis-specific patterns in total LOS. Conditions such as weight loss and paralysis are associated with longer stays despite lower prevalence, indicating a higher per-patient resource burden. In contrast, common diagnoses such as hypertension with complications and anemia affect a large proportion of patients but are associated with more moderate LOS. These patterns highlight the importance of considering both prevalence and LOS when assessing diagnosis-specific impacts on hospital resource utilization.

Comprehensive review of application of machine learning to predict critical care outcome prediction provided at,²² demonstrates that ensemble methods like XGBoost and neural networks effectively address challenges including non-linear relationships, class imbalances, and missing data. Consistent with these findings, XGBoost emerged as the best-performing model across all stages of our framework, achieving an F1-score of 0.701, representing a 5.1% relative improvement, and a macro AUC-ROC of 0.76, corresponding to a 1.9% improvement over traditional single-stage prediction models, despite deliberately excluding post-admission physiological data commonly used in prior studies. While models using first-24-hour data or dynamic updating during hospitalization can achieve higher accuracy,^22,41 they address a fundamentally different clinical question than admission-time prediction. Our approach prioritizes immediate operational utility over maximal predictive accuracy, recognizing that hospitals require actionable predictions at the moment of admission when critical resource allocation decisions must be made.

While the primary goal of our three-stage sequential framework is to predict total hospital LOS, it also provides advantages for hospital operational planning by decomposing the complex LOS prediction into clinically meaningful intermediate predictions. The first stage’s strong performance (AUC 0.835 for ICU admission prediction) enables immediate bed allocation decisions. For ICU-admitted patients, the second stage provides refined ICU-LOS predictions that inform staffing and discharge coordination. Finally, the third stage integrates these intermediate predictions with admission features to produce total hospital LOS forecasts. This staged approach mirrors actual clinical decision-making pathways and allows different components to be updated or recalibrated independently as clinical practices evolve. If the goal were only to predict ICU duration for ICU-admitted patients, restricting the analysis to that subgroup would be more appropriate. However, our framework generates predictions at admission for all patients with unknown ICU status, to support proactive resource allocation.

Our analysis of comorbidity-specific patterns (Figure 5) reveals substantial variation in expected LOS across diagnostic categories, with cardiovascular conditions and multiple comorbidities associated with longer stays, findings that extend the researchers’ systematic identification of prolonged-stay risk factors at.⁴¹ The framework’s validation across 52 diverse Maryland hospitals demonstrates cross-institutional robustness within a single state healthcare system, though external validation on geographically distinct datasets remains an important priority for future research.

In developing our sequential prediction model, a key decision was whether to use real or predicted values when training later-stage models. For example, we first train a model to predict ICU admission. In the next stage, we use its predictions to train the ICU-LOS model. Our results show that using predicted values, rather than real ones, even during training improves results in subsequent stages. To further refine the models, in stage 2, we predicted ICU-LOS only for patients predicted to require ICU care and assigned a value of zero to others. With this approach, we aimed to address class imbalance by focusing the ICU-LOS model on likely ICU cases. However, it led to reduced accuracy in both ICU-LOS and total LOS predictions. The reason is that misclassified ICU patients had ICU-LOS incorrectly set to zero which negatively impacts subsequent models. We also explored alternative strategy of truncating ICU-LOS to 0–14 days but observed no significant performance improvement.

As shown in Table 7, while our model achieves a Macro AUC-ROC of 0.764, outperforming the 0.69–0.73 range reported by,¹⁴ our approach covers 29 disease groups, offering broader applicability than their single-disease focus. Although¹⁹ achieved higher AUC (85–86%), their models relied on detailed, early-stage clinical data and were hospital-specific. In contrast, our model performs well using only admission-time data and is validated across 52 hospitals, highlighting its generalizability. In regression tasks, our R-squared of 0.16 is in line with reported values in the literature (0.146–0.554), reinforcing the model’s competitive performance and wide utility for hospital-wide LOS prediction. In contrast to many prior ICU and LOS prediction studies that rely on single-center or limited-cohort datasets, our analysis leverages a large, multi-hospital administrative dataset and 1,104,203 admissions, enhancing robustness, heterogeneity, and real-world generalizability.

Conclusion

Accurate prediction of hospital LOS is crucial for improving patient care, optimizing resource use, and reducing healthcare costs. To address this, we developed a structured, multi-stage prediction pipeline that improved total LOS classification performance, increasing the F1 score from 0.667 to 0.701 (a 5.1% improvement) and the macro AUC-ROC from 0.749 to 0.763 (a 1.9% increase). In building this framework, we also introduced effective strategies for data splitting, and preprocessing to manage nonlinearity, class imbalance, and skewness, common challenges in LOS data, offering methodological guidance for similar studies.

Beyond technical contributions, our work provides operational insights. Prolonged ICU stays often result from downstream unit bottlenecks rather than medical need. By disaggregating LOS prediction, our approach helps identify ICU-level demand and supports improved patient flow. Unlike most studies focused on a single diagnosis, our dataset spans 20 admission diagnoses, 29 comorbidities, and draws from 52 hospitals, capturing greater clinical and operational variability. This heterogeneity strengthens the model’s applicability, especially under high-demand conditions. Additionally, we identified diagnoses most associated with prolonged stays, supporting targeted preventive strategies and early interventions.

While our framework shows strong performance, some limitations should be acknowledged. First, the model depends on access to prior patient records within the same hospital system and may require periodic retraining to adapt to changes in treatment protocols and healthcare technologies. Second, laboratory measurements from the initial hours of admission and device-use indicators were not included in this study. While such features have been shown to improve LOS prediction in prior work, our objective was to develop a model based exclusively on admission-time information to support early risk stratification and operational decision-making. Third, our results provides evidence of cross-institutional robustness within a single state healthcare system. However, external validation on geographically distinct datasets with different healthcare systems would be beneficial. Future studies could leverage national datasets (e.g., MIMIC-IV, eICU) or international databases to test generalizability and identify any necessary recalibration. Despite these limitations, our findings demonstrate that a sequential, trajectory-based approach can substantially improve LOS prediction and provide a valuable tool for hospital management.

To support transparency and future research, we have shared our code on GitHub in accordance with TRIPOD-AI guidelines,⁴² and the TRIPOD-AI checklist is provided in the supplementary material.

Supplemental material

Supplemental material - A sequential machine learning framework for ICU and hospital length of stay prediction from admission-time data

Supplemental material for A sequential machine learning framework for ICU and hospital length of stay prediction from admission-time data by Maryam Jafaripakzad, Jorge A. Acuna, Keith Dombrowski, José L. Zayas-Castro in DIGITAL HEALTH

Footnotes

ORCID iDs

Maryam Jafaripakzad

Jorge A. Acuña

Consent to participate

This article does not contain any studies with human participants or animals performed by any of the authors. The case study presented herein utilizes an anonymous dataset, and does not entail the involvement of human or animal subjects. Therefore, patient consent was not required. HCUP Databases Are Limited Data Sets. Under HIPAA, review by an institutional review board (IRB) is not required for use of limited data sets .

Author contributions

All authors listed on the submission made substantial contributions to the conception or design of the work. M.J. was responsible for the development of the prediction models and writing the original draft of the manuscript supported by J. A. In addition, M.J. and J.A. contributed to the analysis and interpretation of the data and results. J. A. and J. Z. contributed to the design of the prediction framework and helped shape the overall research approach. K.D. provided insights into the healthcare system and contributed to aligning the work with real-world clinical impact. All authors edited, wrote, and revised the manuscript critically for important intellectual content. All authors approved the final version of the manuscript and agree to be accountable for all aspects of the work to ensure that any questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

GitHub repository link

Length of stay prediction.

Supplemental material

Supplemental material for this article is available online.

Appendix

Table 8.

Hyperparameter configurations tested for stage 1 (ICU admission prediction).

Model	Hyperparameter	Values tested	Best	Val. AUC
Logistic Regression	C	0.01, 0.1, 1.0*, 10.0	1.0	0.8600
Logistic Regression	max_iter	1000	-
Random Forest	n_estimators	50, 100, 200*	200	0.9193
Random Forest	max_depth	10, 15, 20*	20
XGBoost	max_depth	3, 6, 9*	9	0.9305
	eta	0.1, 0.3*	0.3
	num_boost_round	100	100
AdaBoost	n_estimators	50, 100*	100	0.8434
	learning_rate	1.0, 0.5*	0.5
	max_depth	3, 5*	5

Best configurations (marked with *) were selected.

Table 9.

Neural network architectures evaluated for ICU admission in stage 1 (binary classification).

Model	Model type	Architecture details
NN ICU-1	7-layer NN	[128, 64, 32, 64, 32, 16, 8], PReLU, Dropout 0.3
NN ICU-2	8-layer NN	[128, 256, 256, 128, 64, 64, 32, 16, 8], PReLU, Dropout 0.3
NN ICU-3	10-layer NN	[256, 512, 512, 256, 256, 128, 64, 32, 32, 16, 8], PReLU, Dropout 0.3

Figure 6.

Confusion matrix for ICU admission model–stage 1.

Figure 7.

Calibration curve for ICU admission model–stage 1.

Table 10.

Hyperparameter configurations tested for stage 2 (ICU-LOS).

Approach	Model	Hyperparameter	Values tested	Val. metric
Regression (MAE)
	Random Forest	n_estimators	100, 200*, 300	0.4696 (MAE)
	Random Forest	max_depth	10, 15*, 20
	XGBoost	max_depth	3, 6*, 9	0.4775 (MAE)
		eta	0.1, 0.3*
		num_boost_round	100*
	AdaBoost	n_estimators	50, 100*	0.9515 (MAE)
	AdaBoost	learning_rate	0.5, 1.0*
	Gradient Boosting	n_estimators	100, 200*	0.4677 (MAE)
		learning_rate	0.05, 0.1*
		max_depth	3, 5*
Classification (3 classes: ≤1 day, 1–3 days, >3 days)
	Logistic Regression	C	0.1*, 1.0, 10.0	0.7198 (Acc)
	Random Forest	n_estimators	100, 200, 300*	0.7582 (Acc)
	Random Forest	max_depth	10, 15, 20*
	XGBoost	max_depth	3, 6, 9*	0.8116 (Acc)
		eta	0.1, 0.3*
		num_boost_round	100, 150*
	AdaBoost	n_estimators	50, 100, 200*	0.7461 (Acc)
	AdaBoost	learning_rate	0.5, 1.0*

Best configurations (marked with *) were selected.

Table 11.

Neural network architectures evaluated for days in ICU (DaysICU) prediction in stage 2: classification and regression.

Task	Model	Architecture details
Days in ICU (classification)
NN ICU-LOS-C1	5-layer NN	[128, 128, 128, 64, 32], ReLU, Dropout 0.3, Softmax (3 classes)
NN ICU-LOS-C2	5-layer NN	[128, 128, 128, 64, 32], PReLU, Dropout 0.3, Softmax (3 classes)
NN ICU-LOS-C3	5-layer NN	[128, 256, 128, 64, 64], PReLU, Dropout 0.3, Softmax (3 classes)
Days in ICU (Regression)
NN ICU-LOS-R1	7-layer NN	[128, 128, 128, 128, 64, 64, 64], ReLU, Dropout 0.5
NN ICU-LOS-R2	9-layer NN	[128, 128, 256, 256, 128, 128, 64, 64, 64], ReLU, Dropout 0.5
NN ICU-LOS-R3	7-layer NN	[128, 128, 128, 128, 64, 64, 64], PReLU, Dropout 0.5
NN ICU-LOS-R4	5-layer NN	[128, 128, 64, 64, 32], PReLU, Dropout 0.5

Table 12.

ICU-LOS classification performance (stage 2). Results correspond to the best-performing configuration of each model based on macro AUC-ROC.

		Logistic reg.	XGBoost	AdaBoost	Random forests	DICU-C1	DICU-C2	DICU-C3
	Macro AUC-ROC	0.8109	0.8211	0.8139	0.8169	0.7866	0.7921	0.7933
	Overall Accuracy	0.6718	0.6789	0.6736	0.6883	0.5769	0.6018	0.6188
	Weighted F1-score	0.7565	0.7619	0.7584	0.7685	0.6854	0.7045	0.7176
Class 1	Sensitivity	0.6874	0.6942	0.6895	0.7057	0.5819	0.6096	0.6287
	Specificity	0.5964	0.5984	0.6018	0.6029	0.5499	0.4830	0.4574
	Precision	0.9780	0.9783	0.9775	0.9760	0.9841	0.9830	0.9819
	F1-Score	0.8074	0.8121	0.8087	0.8191	0.7313	0.7525	0.7666
Class 2	Sensitivity	0.4034	0.4695	0.4804	0.4556	0.5499	0.4830	0.4574
	Specificity	0.9780	0.9783	0.9775	0.9760	0.9841	0.9830	0.9819
	Precision	0.0850	0.0897	0.0873	0.0889	0.0715	0.0774	0.0800
	F1-Score	0.1404	0.1507	0.1478	0.1488	0.1265	0.1335	0.1361
Class 3	Sensitivity	0.5964	0.5500	0.5137	0.5386	0.4929	0.5487	0.5619
	Specificity	0.9780	0.9783	0.9775	0.9760	0.9841	0.9830	0.9819
	Precision	0.1560	0.1731	0.1714	0.1802	0.1488	0.1321	0.1328
	F1-Score	0.2472	0.2633	0.2570	0.2690	0.2286	0.2130	0.2148

Figure 8.

Calibration belt for the best-performing ICU-LOS classification model–stage 2.

Figure 9.

Confusion matrix for XGBoost ICU-LOS classification model–stage 2.

Figure 10.

Confusion matrix for XGBoost total LOS classification model–stage 3.

Figure 11.

Actual versus predicted ICU-LOS and residual plot for the XGBoost regression model–stage 2. The plots illustrate the skewed distribution of ICU-LOS and increasing prediction error for longer stays.

Table 13.

Hyperparameters tested for stage 3: total LOS prediction.

Approach	Model	Hyperparameter	Values tested	Val. metric
Classification (3 classes: ≤7 days, 7–14 days, >14 days)
	Logistic Regression	C	0.01, 0.1*, 1.0, 10.0	0.7945 (AUC)
	Logistic Regression	max_iter	1000
	Random Forest	n_estimators	100, 200*, 300	0.7935 (AUC)
		max_depth	10, 15*, 20
		min_samples_split	5
	XGBoost	max_depth	3, 6*, 9	0.8055 (AUC)
		eta	0.1, 0.3*
		num_boost_round	100*
	AdaBoost	n_estimators	50, 100, 100*	0.7879 (AUC)
		learning_rate	0.5*, 1.0
		max_depth	3, 5*
Regression (Continuous Length of Stay)
	Ridge Regression	alpha	0.1, 1.0, 10.0*	2.6933 (MAE)
	Random Forest	n_estimators	100, 200, 300*	2.6365 (MAE)
		max_depth	10, 15, 20*
		min_samples_split	5
	XGBoost	max_depth	3, 6*, 9	2.6013 (MAE)
		eta	0.1, 0.3*
		num_boost_round	100*, 150
	AdaBoost	n_estimators	50, 100, 200	4.6811–6.5107 (MAE)
	AdaBoost	learning_rate	0.5, 1.0
	Gradient Boosting	n_estimators	100, 200*	2.6077 (MAE)
		learning_rate	0.05, 0.1
		max_depth	3, 5*

Best configurations (marked with *) were selected based on MAE (regression) or AUC (classification).

Table 14.

Neural network architectures evaluated for hospital length of stay (LOS) prediction in stage 3: classification and regression.

Task	Model	Architecture details
Hospital LOS (classification)
NN LOS-C1	9-layer NN	[128, 256, 256, 128, 64, 64, 32, 16, 8], PReLU, Dropout 0.5, Softmax (3 classes)
NN LOS-C2	6-layer NN	[128, 256, 256, 128, 64, 32], PReLU, Dropout 0.5, Softmax (3 classes)
NN LOS-C3	8-layer NN	[128, 128, 256, 256, 128, 128, 64, 64], PReLU, Dropout 0.5, Softmax (3 classes)
Hospital LOS (Regression)
NN LOS-R1	3-layer NN	[128, 64, 32], PReLU, Dropout 0.5
NN LOS-R2	7-layer NN	[128, 256, 512, 256, 128, 64, 64], PReLU, Dropout 0.5
NN LOS-R3	9-layer NN	[128, 256, 128, 256, 128, 64, 128, 64, 32], PReLU, Dropout 0.5

Figure 12.

Calibration curves for XGBoost LOS classification model in stage 3. The dashed diagonal represents perfect calibration. Brier scores are shown for each class.

Figure 13.

Actual versus predicted total LOS and residual plot for the XGBoost regression model–stage 3.

Table 15 presents the top features ranked by XGBoost gain and SHAP importance for the final LOS classification model. XGBoost gain measures the average improvement in loss function when a feature is used for splitting, while SHAP values quantify each feature’s average contribution to individual predictions. The predicted ICU admission status (ICU_p) ranks 2nd by gain and 8th by SHAP, whereas predicted ICU duration (DaysICU_p) ranks 21st by gain but 6th by SHAP. This difference reflects that (ICU_p) provides large information gain when used, while DaysICU_p contributes more consistently across predictions. Both predicted features rank among the top 10 by SHAP importance, alongside established clinical predictors such as AGE, confirming their meaningful contribution to LOS prediction. Table 15.

Top features by XGBoost gain and SHAP importance in stage 3.

XGBoost gain			SHAP
Rank	Feature	Gain	Rank	Feature	Mean \|SHAP\|
1	HOSP_210064	471.35	1	AGE	0.121
2	ICU_p	295.78	2	LYTES	0.091
3	ASOURCE_2.0	295.01	3	ANEMDEF	0.074
4	HOSP_213029	251.90	4	HOSP_210009	0.061
5	HOSP_213028	187.22	5	ASOURCE_2.0	0.052
6	HOSP_210009	181.99	6	DaysICU_p	0.049
7	HOSP_210058	169.65	7	WGHTLOSS	0.046
8	HOSP_210002	158.03	8	ICU_p	0.041
9	LYTES	155.02	9	DMCX	0.034
10	WGHTLOSS	132.98	10	ASOURCE_5.0	0.034
21	DaysICU_p	70.77

References

Eskandari

Bahmani

AHA

Mardani-Fard

, et al. Evaluation of factors that influenced the length of hospital stay using data mining techniques. BMC Medical Informatics and Decision Making 2022; 22(1): 1–11. https://doi.org/10.1186/s12911-022-02027-w

Toh

Lim

Yap

, et al. Factors associated with prolonged length of stay in older patients. Singapore Medical Journal 2017; 58(3): 134–139. https://doi.org/10.11622/smedj.2016158

Alder

. Effects of poor communication in healthcare. https://www.hipaajournal.com/effects-of-poor-communication-in-healthcare/ 2023, Accessed 31 January 2025.

Long

Mathews

. The boarding patient: effects of icu and hospital occupancy surges on patient flow. Production and Operations Management 2018; 27(12): 2122–2143. https://doi.org/10.1111/poms.12808

Azzolina

Lanera

Comoretto

, et al. Automatic forecast of intensive care unit admissions: the experience during the covid-19 pandemic in italy. Journal of Medical Systems 2023; 47: 84. https://doi.org/10.1007/s10916-023-01982-9

Lather

Ahmadi

Akindele

, et al. Evaluation of hospital pandemic response decisions to design for operational flexibility. In: Computing in Civil Engineering 2023. ASCE, 2023, pp. 237–245. https://doi.org/10.1061/9780784485248.029

Singh

Farmer

Tully

, et al. Forecasting surgical bed utilization: architectural design of a machine learning pipeline incorporating predicted length of stay and surgical volume. Journal of Medical Systems 2025; 49(1): 67. https://doi.org/10.1007/s10916-025-02201-3

Hammoudeh

Al-Naymat

Ghannam

, et al. Predicting hospital readmission among diabetics using deep learning. Procedia Computer Science 2018; 141: 484–489. https://doi.org/10.1016/j.procs.2018.10.138

Daghistani

Elshawi

Sakr

, et al. Predictors of in-hospital length of stay among cardiac patients: a machine learning approach. International Journal of Cardiology 2019; 288: 140–147. https://doi.org/10.1016/j.ijcard.2019.01.046

10.

Alsinglawi

Alnajjar

Mubin

, et al. Predicting length of stay for cardiovascular hospitalizations in the intensive care unit: machine learning approach. In: 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pp. 5442–5445. IEEE. https://doi.org/10.1109/EMBC44109.2020.9175889

11.

Huang

Juarez

Duan

, et al. Length of stay prediction for clinical treatment process using temporal similarity. Expert Systems with Applications 2013; 40(16): 6330–6339. https://doi.org/10.1016/j.eswa.2013.05.066

12.

Cui

Xie

Shen

, et al. Prediction of the healthcare resource utilization using multi-output regression models. IISE Transactions on Healthcare Systems Engineering 2018; 8(4): 291–302. https://doi.org/10.1080/24725579.2018.1512537

13.

Hachesu

Ahmadi

Alizadeh

, et al. Use of data mining techniques to determine and predict length of stay of cardiac patients. Healthcare Informatics Research 2013; 19(2): 121–129. https://doi.org/10.4258/hir.2013.19.2.121

14.

Tsoukalas

Albertson

Tagkopoulos

. From data to optimal decision making: a data-driven, probabilistic machine learning approach to decision support for patients with sepsis. JMIR Medical Informatics 2015; 3(1): e3445. https://doi.org/10.2196/medinform.3445

15.

Turgeman

May

Sciulli

. Insights from a machine learning model for predicting the hospital length of stay (los) at the time of admission. Expert Systems with Applications 2017; 78: 376–385. https://doi.org/10.1016/j.eswa.2017.02.023

16.

Livieris

Dimopoulos

Kotsilieris

, et al. Predicting length of stay in hospitalized patients using ssl algorithms. In: Proceedings of the 8th International Conference on Software Development and Technologies for Enhancing Accessibility and Fighting Info-Exclusion, pp. 16–22. https://doi.org/10.1145/3218585.3218588

17.

Baek

Cho

Kim

, et al. Analysis of length of hospital stay using electronic health records: a statistical and data mining approach. PLOS ONE 2018; 13(4): e0195901. https://doi.org/10.1371/journal.pone.0195901

18.

Tully

Zhong

Simpson

, et al. Machine learning prediction models to reduce length of stay at ambulatory surgery centers through case resequencing. Journal of Medical Systems 2023; 47: 71. https://doi.org/10.1007/s10916-023-01966-9

19.

Rajkomar

Oren

Chen

, et al. Scalable and accurate deep learning with electronic health records. npj Digital Medicine 2018; 1(1): 18. https://doi.org/10.1038/s41746-018-0029-1

20.

Saly

Yang

Triebwasser

, et al. Approaches to predicting outcomes in patients with acute kidney injury. PLOS ONE 2017; 12(1): e0169305. https://doi.org/10.1371/journal.pone.0169305

21.

Liu

Kipnis

Gould

, et al. Length of stay predictions: improvements through the use of automated laboratory and comorbidity variables. Medical Care 2010; 48(8): 739–744. https://doi.org/10.1097/MLR.0b013e3181e359f3

22.

Atallah

Nabian

Brochini

, et al. Machine learning for benchmarking critical care outcomes. Healthcare Informatics Research 2023; 29(4): 301–314. https://doi.org/10.4258/hir.2023.29.4.301

23.

Peres

Hamacher

Oliveira

FLC

, et al. Data-driven methodology to predict the icu length of stay: a multicentre study of 99,492 admissions in 109 brazilian units. Anaesthesia Critical Care & Pain Medicine 2022; 41(6): 101142. https://doi.org/10.016/j.accpm.2022.101142

24.

Brochini

Liu

Atallah

, et al. Prediction of intensive care length of stay for surviving and nonsurviving patients using deep learning. Critical Care Medicine 2025; 53(4): e794–e804. https://doi.org/10.1097/CCM.0000000000006588

25.

Wang

, et al. Length of stay prediction for icu patients using individualized single classification algorithm. Computer Methods and Programs in Biomedicine 2020; 186: 105224. https://doi.org/10.1016/j.cmpb.2019.105224

26.

Iwase

Nakada

Shimada

, et al. Prediction algorithm for icu mortality and length of stay using machine learning. Scientific Reports 2022; 12(1): 12912. https://doi.org/10.1038/s41598-022-17091-5

27.

Navaz

Mohammed

Serhani

, et al. The use of data mining techniques to predict mortality and length of stay in an icu. In: 2016 12th International Conference on Innovations in Information Technology (IIT), pp. 1–5. IEEE. https://doi.org/10.1109/innovations.2016.7880045

28.

Lan

Zhang

Luo

, et al. Prediction of intensive care unit admission (>24 h) after surgery in elective noncardiac surgical patients using machine learning algorithms. Digital Health 2022; 8: 20552076221110543. https://doi.org/10.1177/20552076221110543

29.

Gokhale

Taylor

Gill

. et al. Hospital length of stay prediction for general surgery and total knee arthroplasty admissions: systematic review and meta-analysis of published prediction models. Digital Health 2023; 9: 20552076231177497. https://doi.org/10.1177/20552076231177497

30.

Rockenschaub

Hilbert

Kossen

, et al. The impact of multi-institution datasets on the generalizability of machine learning prediction models in the icu. Critical Care Medicine 2024; 52(11): 1710–1721. https://doi.org/10.1097/CCM.0000000000006359

31.

Lee

. Resampling methods improve the predictive power of modeling in class-imbalanced datasets. International Journal of Environmental Research and Public Health 2014; 11(9): 9776–9789. https://doi.org/10.3390/ijerph110909776

32.

Healthcare Cost and Utilization Project (HCUP) . Hcup state inpatient databases (sid). Agency for Healthcare Research and Quality, 2018. https://hcup-us.ahrq.gov/sidoverview.jsp. 2017–2018 (Accessed 31 January 2025).

33.

Chrusciel

Girardon

Roquette

, et al. The prediction of hospital length of stay using unstructured data. BMC Medical Informatics and Decision Making 2021; 21(1): 351. https://doi.org/10.1186/s12911-021-01722-4

34.

Zhao

Scales

, et al. Predicting in-hospital length of stay: a two-stage modeling approach to account for highly skewed data. BMC Medical Informatics and Decision Making 2022; 22: 110. https://doi.org/10.1186/s12911-022-01855-0

35.

Symum

Zayas-Castro

. Prediction of chronic disease-related inpatient prolonged length of stay using machine learning algorithms. Healthcare Informatics Research 2020; 26(1): 20–33. https://doi.org/10.4258/hir.2020.26.1.20

36.

Ofori-Asenso

Liew

Mårtensson

, et al. The frequency of, and factors associated with prolonged hospitalization: a multicentre study in victoria, australia. Journal of Clinical Medicine 2020; 9(9): 3055. https://doi.org/10.3390/jcm9093055

37.

Jaotombo

Pauly

Fond

, et al. Machine-learning prediction for hospital length of stay using a french medico-administrative database. Journal of Market Access & Health Policy 2022; 11(1): 2149318. https://doi.org/10.1080/20016689.2022.2149318

38.

Power

Harrison

. Why try to predict icu outcomes? Current Opinion in Critical Care 2014; 20(5): 544–549. https://doi.org/10.1097/MCC.0000000000000136

39.

Peres

Ferrari

Quintairos

, et al. Validation of a new data-driven slosr icu efficiency measure compared to the traditional sru. Intensive Care Medicine 2023; 49(12): 1546–1548. https://doi.org/10.1007/s00134-023-07255-w

40.

Peres

Hamacher

Oliveira

FLC

, et al. Prediction of intensive care units length of stay: a concise review. Revista Brasileira de Terapia Intensiva 2021; 33(2): 183–187. https://doi.org/10.5935/0103-507X.20210025

41.

Peres

Hamacher

Oliveira

FLC

, et al. What factors predict length of stay in the intensive care unit? systematic review and meta-analysis. Journal of Critical Care 2020; 60: 183–194. https://doi.org/10.1016/j.jcrc.2020.08.003

42.

Collins

Moons

KGM

Dhiman

, et al. Tripod+ai statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 2024; 385: e078378. https://doi.org/10.1136/bmj-2023-078378

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.14 MB

0.00 MB