Next Article in Journal
Generative AI-Assisted Automation of Clinical Data Processing: A Methodological Framework for Streamlining Behavioral Research Workflows
Previous Article in Journal
Reimagining Traditional Workspaces Through Digitalisation and Hybrid Perspective: A Systematic Review
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Informatics
Volume 13
Issue 4
10.3390/informatics13040047
informatics-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Concurrent Prediction of Length of Stay, Mortality, and Total Charges in Patients with Acute Lymphoblastic Leukemia Using Continuous Machine Learning

by
Jiahui Ma
1,
Elizabeth Johnson
2,
Bradley M. Whitaker
3,
Faraz Dadgostari
1,
Hansjorg Schwertz
4 and
Bernadette McCrory
1,5,*
1
Department of Mechanical and Industrial Engineering, Norm Asbjornson College of Engineering, Montana State University, Norm Asbjornson Hall, Bozeman, MT 59717, USA
2
Mark & Robyn Jones College of Nursing, Montana State University, Sherrick Hall 212, Bozeman, MT 59717, USA
3
Department of Electrical and Computer Engineering, Norm Asbjornson College of Engineering, Montana State University, Norm Asbjornson Hall, Bozeman, MT 59717, USA
4
Department of Occupational Medicine, Billings Clinic Bozeman, Bozeman, MT 59717, USA
5
BioReD Hub, Montana State University, Bozeman, MT 59717, USA
*
Author to whom correspondence should be addressed.
Informatics 2026, 13(4), 47; https://doi.org/10.3390/informatics13040047
Submission received: 15 January 2026 / Revised: 5 March 2026 / Accepted: 20 March 2026 / Published: 24 March 2026
Browse Figures
Versions Notes

Abstract

Acute lymphoblastic leukemia (ALL) presents significant clinical challenges due to its genetic complexity and high relapse rates. While outcomes like length of stay (LOS), mortality, and total charges (TCs) are critical quality indicators, most existing models rely on static data and separate outcome modeling. This study utilized the HCUP National Inpatient Sample (NIS) to develop a dynamic, concurrent prediction model for prolonged LOS and mortality (PLOSM), alongside a framework for TCs. By integrating temporally updated patient information, the concurrent approach outperformed single-outcome models. Within the first seven days of hospitalization, the model achieved accuracy and precision above 90%, with recall and F1-scores exceeding 80%. Key predictors of these outcomes included age, race, insurance type, financial indicators, and elective surgery status. Notably, both prolonged LOS and mortality were significant drivers of TCs. By bridging predictive modeling and real-time clinical data, this framework enables data-driven decision-making to optimize patient management, enhance safety, and mitigate the financial burden of ALL care.

1. Introduction

Blood cancer is the malignancy of hematopoietic stem cells, leading to the uncontrolled growth of immature stem cells to suppress the production of healthy white blood cells, red blood cells, or platelets [1,2]. Unlike solid tumors forming localized masses, the blood cancer can blast from the bone marrow and spread through the bloodstream, lymph system, and central nervous system [1]. Acute lymphoblastic leukemia (ALL) is a malignant proliferation of lymphoid cells that progresses rapidly and can occur both in children and adults, with a peak incidence between 1 year and 4 years [1,3]. Early symptoms of ALL, such as tiredness, fever, recurrent infections, bone discomfort, and loss of weight, can be subtle, such that they are often mistaken for less serious conditions or side effects of other chronic illnesses or events [1,4].

1.1. Challenges for ALL Treatments

Acute leukemia presents significant challenges in diagnosis and treatment due to its genetic complexity, high relapse rates, drug resistance, and treatment-related toxicities, particularly in the central nervous system [5,6]. Advances in molecular research have improved risk stratification and targeted therapies, but overcoming resistance mechanisms and long-term complications remains a critical focus for future research [5,6]. These challenges place patients with ALL at significant disadvantages to timely diagnosis and early treatment, necessitating discussions surrounding clinical trial involvement much earlier in the disease stage [7]. While conventional chemotherapy yields high cure rates in pediatric ALL, adult patients, often presenting with higher-risk factors, comorbidities, and advanced age, experience significantly lower survival rates and clinical trials originally designed for pediatric patients due to the suboptimal efficacy of these treatments [5,8]. The integration of trial-specific patient safety and health considerations has been historically challenging in clinical care settings, such as inpatient hospital stays or routine outpatient visits. Treatments for ALL may cause known and unknown side effects or detrimental changes in body function, such as cardiovascular disease, secondary malignancies, endocrine complications (obesity, osteoporosis, infertility, and premature menopause), and alterations in cognition [9]. Meanwhile, there is an identified significant financial burden due to misalignment of treatment costs and insurance gaps, which may be amplified during active clinical trial participation if the organization is unfamiliar with clinical trial benefit submissions and patient rights when on a clinical trial [10]. Early financial situation screening, implementing financial navigation programs, and integrating discussions about costs into clinical practice are proposed as strategies to mitigate financial toxicity [10]. Through the Clinical Treatment Act, some standard of care exams, diagnostics, and imaging may be covered, provided the healthcare organization is competent in trial-specific billing and coding [11].

1.2. Data-Driven Methods to Enhance Outcomes of Patients with ALL in Inpatient Clinical Settings

Due to challenges in early identification, rapid progression, limited treatment, and high burdens, care coordination is vital for timely intervention. Unexpected trial-related hospitalizations can complicate treatment evaluation and increase risks for hospital-related complications (HACs) or in-hospital mortality [12]. Deepened understanding and contextualization of the anticipated progression of a patient with ALL can further enhance the potential for facilitating promising treatment options, timely enrollment in clinical trials, and improvement of patient outcomes.
Patient length of stay (LOS) and mortality are crucial metrics for evaluating the efficiency, quality, and cost-effectiveness of inpatient care [13,14,15]. In an inpatient setting, LOS, defined as the duration of an inpatient stay, and mortality, the occurrence of death during inpatient care, have been extensively studied [16,17]. Extended inpatient stays, known as prolonged length of stay (PLOS), can not only constrain resource allocation but also elevate the risk of HACs [18,19,20]. All-cause mortality, which is defined as the total number of deaths from any cause within a specific population over a given period, is the most undesired endpoint for patients in the hospital, making it a crucial metric for hospital assessment [21,22].
Patient total charges (TCs) include the charges for all billed services and procedures during an inpatient stay [23]. TCs for inpatient care encapsulate both the resource consumption, including labor, supplies, and medications, as well as the economic burden imposed on patients and their families during the care process [23]. TCs may vary if a patient with ALL is in a clinical trial versus those who follow a standard of care pathway, as clinical trials require heightened monitoring through specimen collection, procedures, and imaging [12,24]. All three outcomes can help clinical facilities, such as hospitals, identify opportunities to improve patient safety and quality of care, reduce financial toxicity, and improve patient satisfaction with research-integrated clinical care, known as Clinical Research as a Care Option (CRAACO) [24].
Although patient outcome prediction has been widely studied, limited studies have been completed on predicting inpatient outcomes for patients with ALL [25,26,27]. LOS and mortality prediction for other diseases has been a subject of extensive research, with methodologies continually advancing to a noticeable shift towards utilizing more advanced machine learning models in recent years [17,28,29,30,31,32,33,34,35,36,37,38,39,40]. Currently, most prediction models on LOS and in-hospital mortality were performed separately, even though these models were completed in the same research [38,39]. The misleading results of using separate prediction models for associated outcomes, the lack of transparency of the data source, and poor interpretation of the model deteriorate the trust of healthcare professionals to adopt the model [28,39]. While a short length of stay is generally a positive indicator for hospital efficiency, some patients with acute conditions may deteriorate rapidly and pass away shortly after admission. Therefore, the need for concurrent prediction and usage of LOS and in-hospital mortality was identified as another gap in the current research [39]. TCs’ prediction witnessed an improving performance with advanced machine learning models, which was also highly determined by LOS [41,42,43,44]. However, most models utilize data, such as baseline patient characteristics or post-discharge outcomes, to make static predictions [17,28,29,30,34,37,40]. Most patient prediction models for hospitalization were performed at admission; therefore, patient treatment information was overlooked in current models [17,28,29,30,31,32,33,34,35,36,37,38,39,40]. Dynamic prediction models have the potential to revolutionize clinical decision-making by adapting to evolving patient conditions, leading to more accurate and timely predictions [45,46,47].
For patient outcome prediction, classification models, such as logistic regression (LR), K-Nearest Neighbors (KNN), Random Forest (RF), Support Vector Machines (SVM), Gradient Boost (GB), and Neural Networks (NN), have been popularly applied [28,31,34,36,37]. LR has been reported to be most frequently used for classification prediction compared with other machine learning models, but ensemble models (such as RF and GB) have been reported to have better performance in predicting LOS and mortality [31,34,36,37,41,43,44,48]. RF and GB are selected to predict PLOS, mortality, and TCs. RF and GB are powerful ensemble methods that combine multiple decision trees to create robust and accurate predictive models [49,50]. RF is an ensemble machine learning method that utilizes bootstrap sampling and feature randomization to aggregate multiple decision trees to perform classification prediction [50]. GB is an ensemble machine learning method to optimize the model by sequentially adding decision trees through gradient descent to minimize the residuals from the previous model [49].

1.3. Purpose

There is a need to predict more clinically meaningful and reliable patient outcomes, enhance predictive accuracy, and enable timely interventions for patients with ALL. This research proposed the dynamic clinical outcome prediction model, which leverages a comprehensive set of patient-level factors identified through previous examination and literature review.
To develop this model, the following aims will be completed:
  • Adapt the model to dynamically predict patient outcomes using the most up-to-date available patient information.
  • Compare the performance of separate patient outcome predictions for LOS and mortality with both RF and GB in a concurrent prediction model that evaluates both outcomes simultaneously.
  • Implement a continuous prediction method to assess TCs by integrating the most accurate predicted results from LOS and mortality prediction.

2. Methods

This predictive analysis utilized machine learning algorithms to predict inpatient outcomes based on a retrospective database: Healthcare Cost and Utilization Project (HCUP), which is developed by the Agency for Healthcare Quality (AHRQ). The HCUP, hosting the largest collection of administrative and longitudinal hospital care data in the United States, was selected for building the inpatient LOS, mortality, and TCs prediction model [51].
The HCUP National (Nationwide) Inpatient Sample (NIS) is an all-payer inpatient healthcare database and comprises inpatient discharge billing data [51]. The NIS includes four discharge-level files: Core, Severity, Hospital, and Diagnosis and Procedure Groups:
  • Core File: demographics, expected primary payer, TCs, discharge status, financial status, and International Classification of Diseases 10th Revision (ICD-10) coding for diagnoses and procedures [52].
  • Severity File: illness severity and mortality risk for each discharge record, utilizing the All-Patient Refined Diagnosis-Related Group (APRDRG) system, which is assigned using software from 3M Health Information Systems [53].
  • Hospital File: characteristics of each hospital, including hospital location, ownership, and size.
  • Diagnosis and Procedure Groups File: patient’s comorbidities presented at admission were defined by the Elixhauser Comorbidity Software (version 2021.1) Refined for ICD-10-CM in this file [54].
Information related to other outcomes and sampling information from NIS were not included in this study. The clinical trial designation code (ICD-10-CM Z00.6—Encounter for examination for normal comparison and control in clinical research program) was considered in this study to identify patients in clinical trial research, where clinical services are being reimbursed through trial benefits. Data retrieval and slicing were performed using IBM SPSS Statistics (version 29). All statistical analysis and machine learning prediction modeling were performed in Python (version 3.10.0).

2.1. Patient Outcomes Prediction

The 75th percentile of the distribution of the total length of hospital stay for the study population is widely applied in other studies and selected in this study to separate the normal and prolonged hospital stay [55,56]. A combination of levels in PLOS and mortality was used as new levels in the combined outcome, including mortality, normal, PLOS, and both (Figure 1). Individual predictions of PLOS and mortality were performed, whose prediction results were combined and then compared with the prediction of PLOSM. TCs were treated as a continuous outcome. The performance of predicting PLOS and mortality individually against a combined outcome model was evaluated; the better prediction outcomes were then utilized as input features for TC prediction (Figure 1).

2.2. Data Slicing and Cleaning

NIS data were collected from the years 2019 to 2021. Patients with ALL diagnoses were identified using the ICD-10-CM diagnosis codes, including C9100 (ALL not received remission), C9101 (ALL in remission), and C9102 (ALL in relapse). A total of 20,371 discharge records were filtered from the 2019 and 2021 NIS databases (10,452 in 2019 and 9919 in 2021) (Figure 2). Missing values accounted for less than 3% of the total sample size, 180 (n = 20,371). These values were addressed using single-point imputation derived from the full study population: categorical features were imputed using the mode, while continuous variables were imputed using the median (Figure 2). ICD-10 diagnosis (DX) and procedure (PR) codes (up to 40 DX, 25 PR per record, with procedure day) were initially processed as strings. Low-frequency codes (<1%) were removed, and the remaining 253 DX and 29 PR codes were converted to binary features. A team review selected 51 ALL-relevant DX and 29 PR codes (Appendix A Table A1). The initial dataset of 20,371 unique patients was expanded to 23,713 records by restructuring the data into a long format (Figure 2). In this format, each record represents an individual procedure; consequently, patients who underwent multiple procedures appear as multiple distinct entries in the expanded dataset.

2.3. Statistical Analysis

Statistical analyses included descriptive statistics (mean/SD, frequency) and inferential analysis, including Pearson correlation and Analysis of Variance (ANOVA), to examine feature relationships. The significance level was set to 0.05 for all inferential analyses.

2.4. Feature Preparation and Selection

Categorical variables were converted to dummy variables (low-frequency categories excluded). Recursive feature elimination (RFE), combined with 5-fold cross-validation (CV), was used to select relevant features and reduce overfitting (Figure 3) [57,58,59,60].

2.5. Prediction Modeling

Both RF and GB methods are designed for handling high-dimensional datasets and selected for this study (Figure 3) [36]. Five-fold cross-validation was applied to each model to enhance model performance and reduce over-fitting [59]. This study used a dynamic prediction process to continuously model patient outcomes from the first day of hospitalization up to the point defining prolonged LOS. For each prediction day, the feature set was reconstructed using only information available up to that day by masking procedures occurring afterward. The dataset was then randomly split at the hospitalization-record level into 80% training and 20% testing sets. SMOTE resampling, feature selection (RFE), and feature scaling were performed exclusively on the training set to prevent data leakage. Model performance was evaluated using 5-fold cross-validation within the training data and further validated on the independent test set. Models were trained independently for each prediction day. Two ensemble learning algorithms were evaluated, Random Forest (n_estimators = 100, random_state = 42) and Gradient Boosting (n_estimators = 100, random_state = 42), with other hyperparameters kept at default values in scikit-learn. For each prediction day, both models were trained independently and evaluated using 5-fold cross-validation within the training set. Final performance was assessed on an independent 20% held-out test set.

2.6. Model Evaluation and Interpretation

Model performance was assessed using multiple evaluation standards (Figure 3). Model accuracy, precision, recall, and F1-score were used to evaluate classification models for predicting PLOS and mortality [61]. Additionally, the area under the receiver operating characteristic (ROC) curve was computed to evaluate the model’s ability to discriminate between classes [61].
  • Precision = True Positives/(True Positives + False Positives)
  • Recall = True Positives/(True Positives + False Negatives)
  • F1-score = 2 × (Precision × Recall)/(Precision + Recall)
  • Accuracy = (True Positives + False Negatives)/(Total Number of Predictions)
Coefficient of determination (R2) measures the proportion of variance in the dependent variable explained by the model [62]. Root mean square error (RMSE) is the square root of the average of the squared differences between predictions and actual observations [62]. Mean absolute error (MAE) calculates the average absolute difference between predicted and actual values [62]. All three metrics were used to evaluate TC prediction [62].
The feature importance method can quantify the contribution of each individual feature to help identify the most influential features for the prediction outcome [63]. Feature importance was assessed using Gini importance, a measure of how much each feature contributes to reducing impurity in the trees of the ensemble model [64]. A feature importance score was assigned to each feature, and the top 30 important features from each day’s prediction were selected to interpret the prediction performance in this study [63].

3. Results

This study included 20,371 patients with ALL from the 2019 (10,452) and 2021 (9919) NIS databases. LOS (Mean: 8.9, SD: 13.3) and TCs (Mean: 152,358.2, SD: 334,148.1 $) were skewed; mortality was imbalanced (19,887 deceased, 97.6%; 480 discharged, 2.4%) (Figure 4). PLOS was defined as >7 days (75th percentile after outlier removal), resulting in 13,198 prolonged stays (64.8%) and 7173 normal stays (35.2%) (Figure 4). PLOS and mortality were highly associated with TCs (p < 0.001, ANOVA); PLOS was also associated with mortality (p < 0.001, Chi-square). Age (Mean: 28.6, SD: 23.6) had a weak correlation with TCs (r = 0.008, p > 0.05). Patients with prolonged stay (Mean: 31.6, SD: 23.8) were older than those with normal stay (Mean: 27.0, SD: 23.4); deceased patients (Mean: 45.7, SD: 23.5) were older than discharged patients (Mean: 28.2, SD: 24.1).
Patient demographics, financial status, hospital, and medical information were summarized by outcome (Table 1). Low-frequency (<1%) binary features were removed, and potential multicollinearity features were excluded. In total, 136 features were used for modeling.

3.1. Model Performance

3.1.1. Individual Prediction of PLOS and Mortality

For the individual patient outcome prediction of PLOS and mortality, the mortality prediction models with higher AUC scores (RF AUC range: 0.9630–0.9956; GB AUC range: 0.9614–0.9709) outperformed the PLOS prediction (RF AUC range: 0.8913–0.9677; GB AUC range: 0.8641–0.8955) for all days’ prediction (Figure 5).
This superiority was also reflected in accuracy, precision, recall, and F1-scores (Appendix B Table A2). RF generally performed better than GB for both outcomes, exhibiting higher AUC values (Figure 5) and generally higher accuracy, precision, recall, and F1-scores (Appendix B Table A2).

3.1.2. Prediction of Combined Outcome (PLOSM)

As detailed in Figure 1, there were four classes in the combined outcome (PLOSM) of PLOS and mortality (normal: neither prolonged stay nor deceased; PLOS only: prolonged stay without deceased; mortality only: deceased without prolonged stay; and both: prolonged stay and deceased). ROC curves were generated for each class using both RF and GB models (Figure 6). Across all classes and predictions, RF models demonstrated superior performance compared to GB models, as evidenced by higher AUC scores (Figure 6).
Comparing performance metrics, PLOSM prediction achieved higher accuracy, precision, recall, and F1-scores than PLOS prediction, but lower scores than mortality prediction (Appendix C Table A3). Importantly, the PLOSM prediction models exhibited strong performance in predicting the minority classes: mortality only (RF AUC range: 0.9718–0.9978) and the combined PLOSM outcome (RF AUC range: 0.9626–0.9987).

3.1.3. Comparing the Prediction of PLOSM with the Combined Outcome Prediction of PLOS and Mortality

A comparison was conducted between individual predictions of PLOS and mortality and the combined PLOSM prediction. To facilitate this comparison, the results of the individual PLOS and mortality predictions were combined, and performance metrics were recalculated specifically for the testing set (Appendix C Table A3). This combination method precluded the calculation of training set metrics (accuracy, precision, recall, and F1-score) (Appendix C Table A3). Comparing the testing set metrics (Figure 7) revealed that the RF PLOSM prediction generally achieved higher accuracy, F1-scores, and recall. However, the combined PLOS and mortality prediction demonstrated superior performance on day 1 and after day 5 (Figure 7).

3.1.4. Prediction of TCs

TCs’ prediction with RF outperformed the prediction with GB for both training and testing sets across all evaluation metrics, including RMSE, MAE, and R2 (Figure 8). Notably, RF’s RMSE and MAE values decreased as the prediction day advanced, whereas GB’s RMSE and MAE values exhibited an increasing trend (Figure 8). This resulted in RF consistently achieving lower RMSE and MAE values than GB, with the difference between the two models widening as the prediction day increased (Figure 8). While both models showed improved R2 values with increasing prediction day, RF consistently achieved higher R2 values than GB.

3.2. Feature Interpretation

Early time series data proved critical for both PLOSM and TC prediction, with PLOSM models showing maximum feature selection within the first three days, and TC models extending feature relevance to the initial five days (Figure 9). The illness severity and mortality with higher risk scores were the most important features to predict the PLOSM with RF, while the prediction of PLOSM, especially for minority levels of PLOS only and both PLOS and mortality, was essential to predict the TCs for day 1 through day 7 (Figure 9). Other demographic and diagnosis factors, such as age, race, financial status, insurance choices, elective surgeries, major operating room usage, and diagnosis codes, were ranked into the top 10 important predictors for PLOSM and TC prediction (Figure 9). For PLOSM prediction, procedures of Introduction of Other Antineoplastic into Spinal Canal, Percutaneous Approach (3E0R305) and Introduction of Other Antineoplastic into Central Vein, Percutaneous Approach (3E04305) were more important than other procedure predictors, especially on early days at 1, 2, and 3 (Figure 9). However, no features were important to predict the TCs.

4. Discussion

Despite ALL treatment advancements, challenges persist in early diagnosis, treatment complexity, and financial burden [5,6,10,18]. Lack of reliable prediction models impedes timely decisions and resource allocation, impacting patient care. Current models often fail to integrate real-time data or combine LOS and mortality predictions, limiting personalized care [38]. This study addressed these limitations by using HCUP-NIS real-time data to develop dynamic prediction models for concurrent outcomes (PLOS, mortality, and TCs), aiming to improve clinical decisions, resource optimization, and patient safety in ALL.

4.1. Significance of Early Prediction in ALL Treatment

The study findings emphasized the importance of early prediction, particularly within the first three days of hospitalization, in forecasting patient outcomes (Figure 9). Different predictors were found to be significant for forecasting concurrent patient outcomes of PLOS and mortality at each stage, providing valuable insights into the evolving nature of risk factors over time (Figure 9). Notably, the RF model outperformed other methods in predicting concurrent patient outcomes (PLOSM) compared to individual outcome predictions. While RF achieved similar accuracy and precision values, it demonstrated higher AUC, F1, and recall scores (Figure 7, Figure 8 and Figure 9). These findings indicate that the concurrent RF model is superior in predicting each patient’s outcome class, particularly for minority cases, where patient outcomes are highly imbalanced (e.g., a greater proportion of normally discharged patients compared to those with prolonged stays or in-hospital mortality). However, the greater performance drop observed in the four-class PLOSM outcome compared to binary PLOS and mortality prediction, in Appendix C (Table A3), likely reflected increased classification complexity, class imbalance, and overlapping feature distributions among outcome categories. The multiclass task requires distinguishing between combinations of LOS and mortality states, which reduced effective sample size per class and increased boundary complexity, thereby increasing generalization error.
Since inpatient adverse events are less frequent than normal discharges and most prediction models focus on separately predicting LOS and mortality at admission, the ability to anticipate prolonged hospitalization and in-hospital mortality concurrently enables healthcare providers to personalize treatment plans, enhance patient safety, and allocate resources more efficiently [17,28,29,30,31,32,33,34,35,36,37,38,39,40]. These insights are particularly critical for patients with ALL, whose immunocompromised status makes them highly susceptible to hospital-related complications [5,6].
The continuous prediction methodology, which utilizes predicted PLOS and mortality as predictors for TC prediction, enhances the timely forecasting of secondary outcomes when critical features are unavailable at the time of early prediction (Figure 9). Traditional machine learning models for TC prediction rely on post-discharge LOS as a predictor, a variable not available in early-stage assessments [42,44]. This study demonstrates that integrating machine learning pipeline techniques allows for the early prediction of TCs, thereby offering a proactive approach to managing patient care.

4.2. Clinical and Financial Implications

Multiple factors, including disease stage, APRDRG severity risk, APRDRG mortality risk, age, race, geographic location, household income, insurance type, hospital transfer status, medical service line, diagnosis, and treatment, played a significant role in predicting PLOS and mortality, aligning with the important factors from other prediction models [21,34]. Additionally, patient comorbidities (e.g., hypertension and depression), hospital resources and location, clinical trial participation, and predicted PLOSM were all key predictors for TCs (Figure 9). Notably, patients transferred from acute care hospitals significantly influenced PLOSM prediction, aligning with findings that pediatric patients registered in clinical trials exhibited similar risk factors [12]. These findings underscore the importance of identifying key predictors in inpatient care delivery systems to facilitate early interventions tailored to individual disease conditions, diagnoses, and hospital resources. Beyond clinical improvements, this study highlights the economic burden of ALL treatment. Early prediction of TCs provides valuable insights into financial toxicity, enabling timely interventions to mitigate economic strain on patients and families. Insurance type and patient income levels were found to influence both PLOSM and TC predictions (Figure 9).
Data from the Agency for Healthcare Research and Quality (2006–2008) indicated that patients with private insurance were associated with lower risk-adjusted mortality rates than those with other insurance types in U.S. hospitals [65]. Furthermore, patients with private insurance or higher income levels generally had faster access to healthcare services than those with lower income or alternative insurance coverage [66]. Clinical trial enrollment was another crucial predictor for TCs, with greater participation observed among patients with private insurance [67]. These insights reinforce the necessity of financial navigation programs and policy adjustments to enhance care quality and accessibility, aligning with the CRAACO framework [24].
The implementation of dynamic machine learning models enables real-time updates and continuous patient risk assessment, thereby enhancing clinical decision-making [45,47]. Despite the complexity of ALL treatment, modern first-line therapy of standard- and high-dose chemotherapy, hematopoietic stem-cell transplantation, and new targeted therapy improved the long-term outcome for patients with ALL [68]. Given the ever-evolving nature of ALL treatment, continuous updates to patient outcome prediction models are essential. By integrating the latest treatment strategies into prediction models, hospitals can account for new developments and refine patient risk assessments accordingly [45,47]. Daily patient outcome predictions offer a nuanced understanding of predictor effects, facilitating personalized treatment planning and more efficient hospital resource allocation [45,47].
To illustrate how the prediction model performed across days for each individual patient, the prediction results for a subset of records of patients with ALL enrolled in clinical trial research (n = 434) were summarized, as shown in Figure 10. Since the dataset was pivoted by the procedure received day (Figure 9), the subset was expanded to 774 records (22% from the testing set and 78% from the training set). The model performed well for most records, especially for two (deceased with short hospital stay) and three (prolonged hospital stay and deceased) patient outcomes. Some inaccurate predictions shown in the earlier days’ predictions were able to be corrected in later days’ predictions, such as the true patient outcome label was 1, and the day 1 through day 7 prediction pattern was 0111111. However, there were some patients who ended up with prolonged hospital stay prediction results when they were discharged normally from the hospital (e.g., 0001, 0011, 001011, 000001, and 0000001). This patient cohort presents a valuable opportunity for care teams to reassess disease progression and facilitate close monitoring of patients’ health by family members or caregivers post-discharge. A ten-year-old patient, with a less severe risk and mortality score at admission, received an antineoplastic drug through the central vein on day 1 before achieving remission for ALL. The discrepancy between his hospital record and prediction results on day 7 (11111110) may show that there is a possibility of discharging the patient on day 7, since he only had an eight-day hospitalization. The retrospective nature of HCUP data, coupled with its collection methodology, may result in incomplete patient records, specifically regarding diagnoses and procedures. Critically, the absence of vital signs and other important clinical indicators limits the depth and accuracy of this prediction model. Therefore, this dynamic patient outcome prediction model is suggested to be integrated into hospital electronic health records (EHR) to make daily predictions on patient risk of adverse events (e.g., PLOS and mortality) and estimation of TCs to streamline patient monitoring, facilitate timely interventions, and reallocate the resources for each patient.

5. Limitations and Future Studies

This retrospective analysis of HCUP data is subject to limitations inherent in administrative databases, including potential inaccuracies in procedure coding and the absence of granular clinical data, such as vital signs, which are essential for understanding dynamic patient treatment. Local EHR integration is needed for hospital adoption, especially in rural settings. Generalizability is limited due to U.S. data. Imputation and data imbalance may affect accuracy. Future work should include external validation to re-evaluate the inaccurate prediction for certain patient cohorts and explore deep learning for improved predictions. Real-time EHR integration is crucial for clinical decision support. Potential covariate effects require combining prediction with causal analysis to ensure effective improvements in patient safety and financial outcomes [69].

6. Conclusions

This study provided an effective method to predict multiple patient outcomes in a dynamic and continuous way, combined with machine learning models. By anticipating potential outcomes, healthcare professionals can personalize treatment plans based on individual risk factors, improving the efficacy of care for each patient. Moreover, hospitals optimize resource allocation within healthcare facilities, ensuring that high-risk patients have access to necessary intensive care, specialized nursing, and other resources when they need them most. Timely predictions foster efficient communication among care teams, leading to better-informed decisions and improved patient care. Therefore, developing predictive models for patient outcomes like length of stay, mortality, and TCs is vital to support informed decision-making, optimize clinical workflows, and ultimately improve patient safety and well-being in ALL treatments.

Author Contributions

E.J., B.M.W., F.D., H.S. and B.M. Conceptualization, J.M., E.J., B.M.W. and B.M.; methodology, J.M. and B.M.W.; software, J.M.; validation, J.M., E.J., B.M.W., F.D., H.S. and B.M.; formal analysis, J.M., E.J. and B.M.W.; investigation, J.M.; resources, E.J.; data curation, E.J.; writing—original draft preparation, J.M.; writing—review and editing, J.M., E.J., B.M.W., F.D., H.S. and B.M.; visualization, J.M.; supervision, E.J., B.M.W. and B.M.; project administration, B.M.; funding acquisition, E.J., B.M.W. and B.M. All authors have read and agreed to the published version of the manuscript.

Funding

Funding for this study was made possible with internal fund, library and the Office of Sponsored Programs at Montana State University.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable. The data was purchased from the HCUP database. All patient records were de-identified by HCUP.

Data Availability Statement

The data was purchased through the HEALTHCARE COST & UTILIZATION PROJECT (HCUP). The link to purchase is https://hcup-us.ahrq.gov/db/nation/nis/nisdde.jsp (accessed on 1 March 2025).

Conflicts of Interest

No conflicts of interest declared by any of the authors.

Appendix A

Table A1. ICD-10 Codes Description and Frequency.
Table A1. ICD-10 Codes Description and Frequency.
ICD CodesDescriptionFrequencyRelative Frequency
ICD-10-DX codes
A00–A09Intestinal infectious diseases9124.48%
A30–A49Other bacterial diseases207510.19%
B00–B09Viral infections characterized by skin and mucous membrane lesions4512.21%
B25–B34Other viral diseases8834.33%
B35–B49Mycoses12546.16%
B95–B97Bacterial and viral infectious agents215510.58%
C76–C80Malignant neoplasms of ill-defined, other secondary and unspecified sites2231.09%
C81–C96Malignant neoplasms of lymphoid, hematopoietic and related tissue10495.15%
D60–D64Aplastic and other anemias and other bone marrow failure syndromes11,47956.35%
D65–D69Coagulation defects, purpura and other hemorrhagic conditions419820.61%
D70–D77Other disorders of blood and blood-forming organs526125.83%
D80–D89Certain disorders involving the immune mechanism459622.56%
E40–E46Malnutrition211610.39%
E50–E64Other nutritional deficiencies8724.28%
E70–E88Metabolic disorders879243.16%
G60–G65Polyneuropathies and other disorders of the peripheral nervous system241511.86%
G89–G99Other disorders of the nervous system263012.91%
I10–I1AHypertensive diseases557727.38%
I20–I25Ischemic heart diseases9894.85%
I26–I28Pulmonary heart disease and diseases of pulmonary circulation3851.89%
I30–I5AOther forms of heart disease244011.98%
I60–I69Cerebrovascular diseases5102.50%
I70–I79Diseases of arteries, arterioles and capillaries2461.21%
I80–I89Diseases of veins, lymphatic vessels and lymph nodes, not elsewhere classified10265.04%
I95–I99Other and unspecified disorders of the circulatory system9514.67%
M50–M54Other dorsopathies8884.36%
M86–M90Other osteopathies4082.00%
N10–N16Renal tubulo-interstitial diseases2871.41%
N17–N19Acute kidney failure and chronic kidney disease258012.67%
N25–N29Other disorders of kidney and ureter2761.35%
N30–N39Other diseases of the urinary system8874.35%
R00–R09Symptoms and signs involving the circulatory and respiratory systems302014.82%
R10–R19Symptoms and signs involving the digestive system and abdomen448522.02%
R20–R23Symptoms and signs involving the skin and subcutaneous tissue8053.95%
R25–R29Symptoms and signs involving the nervous and musculoskeletal systems4512.21%
R30–R39Symptoms and signs involving the genitourinary system7343.60%
R40–R46Symptoms and signs involving cognition, perception, emotional state and behavior6503.19%
R50–R69General symptoms and signs831040.79%
R70–R79Abnormal findings on examination of blood, without diagnosis317615.59%
R90–R94Abnormal findings on diagnostic imaging and in function studies, without diagnosis4832.37%
T36–T50Poisoning by, adverse effect of and underdosing of drugs, medicaments and biological substances992248.71%
T80–T88Complications of surgical and medical care, not elsewhere classified19289.46%
U00–U49Provisional assignment of new diseases of uncertain etiology or emergency use3691.81%
Y83–Y84Surgical and other medical procedures as the cause of abnormal reaction of the patient, or of later complication, without mention of misadventure at the time of the procedure12045.91%
Y90–Y99Supplementary factors related to causes of morbidity classified elsewhere224911.04%
Z16–Z16Resistance to antimicrobial drugs2511.23%
Z20–Z29Persons with potential health hazards related to communicable diseases518625.46%
Z40–Z53Encounters for other specific health care823340.42%
Z66Do not resuscitate status7133.50%
Z69–Z76Persons encountering health services in other circumstances5692.79%
Z77–Z99Persons with potential health hazards related to family and personal history and certain conditions influencing health status12,89563.30%
ICD-10-PR codes
30233N1Transfusion of Nonautologous Red Blood Cells into Peripheral Vein, Percutaneous Approach246012.08%
30233R1Transfusion of Nonautologous Platelets into Peripheral Vein, Percutaneous Approach16698.19%
30243N1Transfusion of Nonautologous Red Blood Cells into Central Vein, Percutaneous Approach9144.49%
30243R1Transfusion of Nonautologous Platelets into Central Vein, Percutaneous Approach5592.74%
3E03305Introduction of Other Antineoplastic into Peripheral Vein, Percutaneous Approach13856.80%
3E04305Introduction of Other Antineoplastic into Central Vein, Percutaneous Approach425620.89%
3E0430MIntroduction of Antineoplastic, Monoclonal Antibody, into Central Vein, Percutaneous Approach2291.12%
3E0436ZIntroduction of Nutritional Substance into Central Vein, Percutaneous Approach3191.57%
3E0G76ZIntroduction of Nutritional Substance into Upper GI, Via Natural or Artificial Opening2331.14%
3E0R305Introduction of Other Antineoplastic into Spinal Canal, Percutaneous Approach568827.92%
XW04351Introduction of Blinatumomab Antineoplastic Immunotherapy into Central Vein, Percutaneous Approach, New Technology Group 14592.25%
02H633ZInsertion of Infusion Device into Right Atrium, Percutaneous Approach4182.05%
02HV33ZInsertion of Infusion Device into Superior Vena Cava, Percutaneous Approach301814.82%
02PYX3ZRemoval of Infusion Device from Great Vessel, External Approach1980.97%
03HY32ZInsertion of Monitoring Device into Upper Artery, Percutaneous Approach2051.01%
0JH60WZInsertion of Totally Implantable Vascular Access Device into Chest Subcutaneous Tissue and Fascia, Open Approach6573.23%
0JH63XZInsertion of Tunneled Vascular Access Device into Chest Subcutaneous Tissue and Fascia, Percutaneous Approach3571.75%
009U3ZXDrainage of Spinal Canal, Percutaneous Approach, Diagnostic242411.90%
009U3ZZDrainage of Spinal Canal, Percutaneous Approach2661.31%
5A09357Assistance with Respiratory Ventilation, Less than 24 Consecutive Hours, Continuous Positive Airway Pressure1790.88%
5A1955ZRespiratory Ventilation, Greater than 96 Consecutive Hours2681.32%
B01B1ZZFluoroscopy of Spinal Cord using Low Osmolar Contrast2081.02%
B5181ZAFluoroscopy of Superior Vena Cava using Low Osmolar Contrast, Guidance2101.03%
B518ZZAFluoroscopy of Superior Vena Cava, Guidance2301.13%
B548ZZAUltrasonography of Superior Vena Cava, Guidance4412.16%
079T3ZXDrainage of Bone Marrow, Percutaneous Approach, Diagnostic2861.40%
07DR3ZXExtraction of Iliac Bone Marrow, Percutaneous Approach, Diagnostic318715.64%
8E0ZXY6Isolation2341.15%
0BH17EZInsertion of Endotracheal Airway into Trachea, Via Natural or Artificial Opening3601.77%

Appendix B

Table A2. Individual Prediction of PLOS and Mortality Evaluation.
Table A2. Individual Prediction of PLOS and Mortality Evaluation.
Training (Cross Validation)Testing
OutcomeModelDayAccuracyF1-ScorePrecisionRecallAccuracyF1-ScorePrecisionRecall
MortalityRF10.99130.99110.99400.98860.98070.97730.97740.9807
MortalityRF20.99080.99060.99340.98810.98450.98300.98300.9845
MortalityRF30.99120.99100.99480.98760.98060.97750.97760.9806
MortalityRF40.99220.99200.99460.98980.98170.97840.97980.9817
MortalityRF50.99210.99190.99480.98940.98140.97950.97920.9814
MortalityRF60.99230.99220.99510.98950.98320.98100.98210.9832
MortalityRF70.99240.99220.99490.98980.98350.98080.98320.9835
MortalityRF>70.99460.99450.99580.99330.98690.98610.98660.9869
PLOSRF10.86720.86210.87350.85960.83240.83050.82980.8324
PLOSRF20.86740.86360.87430.86000.83450.83380.83350.8345
PLOSRF30.86410.86220.86450.86410.85810.85800.85800.8581
PLOSRF40.86320.86220.86150.86490.86060.86070.86080.8606
PLOSRF50.86750.86760.86480.87200.85580.85620.85720.8558
PLOSRF60.86620.86680.86230.87160.86260.86300.86440.8626
PLOSRF70.86300.86340.86040.86660.87500.87520.87580.8750
PLOSRF>70.89860.90110.88660.91800.90510.90480.90520.9051
MortalityGB10.98360.98360.97700.99040.96760.96940.97150.9676
MortalityGB20.98210.98210.97400.99070.96820.97180.97740.9682
MortalityGB30.97930.97950.96960.98990.95980.96630.97640.9598
MortalityGB40.97800.97820.96740.98940.95780.96380.97260.9578
MortalityGB50.97770.97790.96680.98940.95430.96120.97160.9543
MortalityGB60.97600.97610.96660.98610.95260.95960.97020.9526
MortalityGB70.97150.97160.96250.98130.94850.95720.97060.9485
MortalityGB>70.95790.95850.94210.97580.93050.94160.96060.9305
PLOSGB10.83310.82640.83720.82580.80630.80650.80680.8063
PLOSGB20.82000.81580.82140.81710.78940.79070.79280.7894
PLOSGB30.80980.80740.80770.81240.78400.78520.78760.7840
PLOSGB40.80670.80570.80270.81140.78510.78620.78840.7851
PLOSGB50.80210.80270.79720.81060.78470.78580.78890.7847
PLOSGB60.79560.79800.78750.80910.78650.78770.79270.7865
PLOSGB70.78160.78170.78110.78240.78820.78900.79230.7882
PLOSGB>70.82800.83210.82030.84670.81610.81580.81570.8161

Appendix C

Table A3. PLOSM Prediction and Combined Outcome Prediction Evaluation.
Table A3. PLOSM Prediction and Combined Outcome Prediction Evaluation.
Training (Cross Validation)Testing
OutcomeModelDayAccuracyF1-ScorePrecisionRecallAccuracyF1-ScorePrecisionRecall
PLOSMRF10.92920.92870.93230.92920.81750.66890.78930.6088
PLOSMRF20.93050.93020.93290.93050.81940.72070.83810.6659
PLOSMRF30.92840.92830.93020.92840.83700.72600.83540.6668
PLOSMRF40.92750.92750.92850.92750.83030.68080.80360.6215
PLOSMRF50.92920.92910.93010.92920.84240.72350.84710.6671
PLOSMRF60.92850.92840.92880.92850.83870.74830.87250.6912
PLOSMRF70.92930.92920.92930.92930.84160.65850.88880.6042
PLOSMRF>70.94590.94590.94680.94590.89160.78870.93180.7469
PLOS + Mortality *RF1----0.81750.59710.78820.5425
PLOS + Mortality *RF2----0.82450.69440.79190.6430
PLOS + Mortality *RF3----0.84130.65350.79070.5939
PLOS + Mortality *RF4----0.84780.66510.83830.6014
PLOS + Mortality *RF5----0.84090.69470.82340.6411
PLOS + Mortality *RF6----0.85020.72780.86040.6633
PLOS + Mortality *RF7----0.86220.67280.88970.6155
PLOS + Mortality *RF>7----0.89470.74460.92900.7011
PLOSMGB10.88810.88580.88810.88920.73710.51760.51680.5576
PLOSMGB20.86970.86770.86970.87020.75140.57810.57300.6210
PLOSMGB30.86380.86180.86380.86310.74090.52970.50120.6258
PLOSMGB40.85480.85310.85480.85380.75210.57700.54150.6574
PLOSMGB50.85720.85520.85720.85560.74210.53580.50400.6249
PLOSMGB60.85440.85280.85440.85230.73940.56430.55220.6258
PLOSMGB70.84990.84820.84990.84740.72250.57140.55710.6542
PLOSMGB>70.85930.85800.85930.85860.75290.54780.54640.6095
PLOS + Mortality *GB1----0.78160.49110.52280.5145
PLOS + Mortality *GB2----0.76650.52580.56330.5799
PLOS + Mortality *GB3----0.75300.53600.52730.5946
PLOS + Mortality *GB4----0.75300.47890.46660.5408
PLOS + Mortality *GB5----0.75120.50290.53690.5590
PLOS + Mortality *GB6----0.74950.49940.50430.5514
PLOS + Mortality *GB7----0.74720.49720.50040.5687
PLOS + Mortality *GB>7----0.75290.51930.53160.5862
* This was the outcome of combining the individual prediction of PLOS and Mortality, which was compared with the PLOSD. The Training metrics can be referred to the Mortality’s and PLOS’s metrics.

References

  1. Meenaghan, T.; Dowling, M.; Kelly, M. Acute leukaemia: Making sense of a complex blood cancer. Br. J. Nurs. 2012, 21, 78–83. [Google Scholar] [CrossRef]
  2. Blumenreich, M.S. The White Blood Cell and Differential Count. In Clinical Methods: The History, Physical, and Laboratory Examinations; Walker, H.K., Hall, W.D., Hurst, J.W., Eds.; Butterworth Publishers: Boston, MA, USA, 1990. [Google Scholar]
  3. Malard, F.; Mohty, M. Acute lymphoblastic leukaemia. Lancet 2020, 395, 1146–1162. [Google Scholar] [CrossRef] [PubMed]
  4. Black, G.B.; Boswell, L.; Harris, J.; Whitaker, K.L. What causes delays in diagnosing blood cancers? A rapid review of the evidence. Prim. Health Care Res. Dev. 2023, 24, e26. [Google Scholar] [CrossRef]
  5. Kansagra, A.; Dahiya, S.; Litzow, M. Continuing challenges and current issues in acute lymphoblastic leukemia. Leuk. Lymphoma 2018, 59, 526–541. [Google Scholar] [CrossRef]
  6. Lonetti, A.; Iacobucci, I.; Masetti, R. Successes and Challenges for Diagnosis and Therapy of Acute Leukemia. J. Oncol. 2019, 2019, 3408318. [Google Scholar] [CrossRef]
  7. Fardin, M.F.; Akter, M. Review on blood cancer and their types. J. Mol. Pharm. Regul. Aff. 2021, 3, 18–26. [Google Scholar]
  8. Paul, S.; Kantarjian, H.; Jabbour, E.J. Adult Acute Lymphoblastic Leukemia. Mayo Clin. Proc. 2016, 91, 1645–1666. [Google Scholar] [CrossRef]
  9. Douvas, M.G.; Riegler, L.L. Meeting Challenges in the Long-Term Care of Children, Adolescents, and Young Adults with Acute Lymphoblastic Leukemia. Curr. Hematol. Malig. Rep. 2022, 17, 15–24. [Google Scholar] [CrossRef] [PubMed]
  10. Abrams, H.R.; Durbin, S.; Huang, C.X.; Johnson, S.F.; Nayak, R.K.; Zahner, G.J.; Peppercorn, J. Financial toxicity in cancer care: Origins, impact, and solutions. Transl. Behav. Med. 2021, 11, 2043–2054. [Google Scholar] [CrossRef]
  11. Lu, B.R. Clinical Treatment Act; Congress, T., Ed.; House—Energy and Commerce: Washington, DC, USA, 2019. [Google Scholar]
  12. Ma, J.; Johnson, E.A.; McCrory, B. Predicting risk factors for pediatric mortality in clinical trial research: A retrospective, cross-sectional study using a Healthcare Cost and Utilization Project database. J. Clin. Transl. Sci. 2023, 7, e211. [Google Scholar] [CrossRef]
  13. Newgard, C.D.; Fleischman, R.; Choo, E.; John Ma, O.; Hedges, J.R.; John McConnell, K. Validation of length of hospital stay as a surrogate measure for injury severity and resource use among injury survivors. Acad. Emerg. Med. 2010, 17, 142–150. [Google Scholar] [CrossRef]
  14. English, M.; Mwaniki, P.; Julius, T.; Chepkirui, M.; Gathara, D.; Ouma, P.O.; Cherutich, P.; Okiro, A.E.; Snow, R.W. Hospital Mortality—A neglected but rich source of information supporting the transition to higher quality health systems in low and middle income countries. BMC Med. 2018, 16, 32. [Google Scholar] [CrossRef]
  15. Nguyen, Q.; Wybrow, M.; Burstein, F.; Taylor, D.; Enticott, J. Understanding the impacts of health information systems on patient flow management: A systematic review across several decades of research. PLoS ONE 2022, 17, e0274493. [Google Scholar] [CrossRef]
  16. Huntley, D.A.; Cho, D.W.; Christman, J.; Csernansky, J.G. Predicting length of stay in an acute psychiatric hospital. Psychiatr. Serv. 1998, 49, 1049–1053. [Google Scholar] [CrossRef] [PubMed]
  17. Zhao, P.; Liu, C.; Zhang, C.; Hou, Y.; Zhang, X.; Zhao, J.; Sun, G.; Zhou, J. Using Machine Learning to Predict the In-Hospital Mortality in Women with ST-Segment Elevation Myocardial Infarction. Rev. Cardiovasc. Med. 2023, 24, 126. [Google Scholar] [CrossRef]
  18. Nghiem, S.; Afoakwah, C.; Scuffham, P.; Byrnes, J. Benchmarking hospital safety and identifying determinants of hospital-acquired complication: The case of Queensland cardiac linkage longitudinal cohort. Infect. Prev. Pract. 2022, 4, 100198. [Google Scholar] [CrossRef]
  19. Rosman, M.; Rachminov, O.; Segal, O.; Segal, G. Prolonged patients’ In-Hospital Waiting Period after discharge eligibility is associated with increased risk of infection, morbidity and mortality: A retrospective cohort analysis. BMC Health Serv Res 2015, 15, 246. [Google Scholar] [CrossRef]
  20. Doctoroff, L.; Herzig, S.J. Predicting Patients at Risk for Prolonged Hospital Stays. Med. Care 2020, 58, 778–784. [Google Scholar] [CrossRef] [PubMed]
  21. Novitski, P.; Cohen, C.M.; Karasik, A.; Shalev, V.; Hodik, G.; Moskovitch, R. All-Cause Mortality Prediction in T2D Patients. In Proceedings of the 18th International Conference on Artificial Intelligence in Medicine (AIME); Electr Network; University of Minnesota: Minneapolis, MN, USA, 2020. [Google Scholar]
  22. Parshuram, C.S.; Dryden-Palmer, K.; Farrell, C.; Gottesman, R.; Gray, M.; Hutchison, J.S.; Helfaer, M.; Hunt, E.A.; Joffe, A.R.; Lacroix, J.; et al. Effect of a Pediatric Early Warning System on All-Cause Mortality in Hospitalized Pediatric Patients: The EPOCH Randomized Clinical Trial. JAMA 2018, 319, 1002–1012. [Google Scholar] [CrossRef] [PubMed]
  23. Hamavid, H.; Birger, M.; Bulchis, A.G.; Lomsadze, L.; Joseph, J.; Baral, R.; Bui, A.L.; Horst, C.; Johnson, E.; Dieleman, J.L. Assessing the Complex and Evolving Relationship between Charges and Payments in US Hospitals: 1996–2012. PLoS ONE 2016, 11, e0157912. [Google Scholar] [CrossRef]
  24. Van de Beek, H. Clinical Research as a Care Option: Optimizing Approaches. Appl. Clin. Trials 2019, 28, 32–35. [Google Scholar]
  25. Kashef, A.; Khatibi, T.; Mehrvar, A. Treatment outcome classification of pediatric Acute Lymphoblastic Leukemia patients with clinical and medical data using machine learning: A case study at MAHAK hospital. Inform. Med. Unlocked 2020, 20, 100399. [Google Scholar] [CrossRef]
  26. Thakkar, S.G.; Fu, A.Z.; Sweetenham, J.W.; Mciver, Z.A.; Mohan, S.R.; Ramsingh, G.; Advani, A.S.; Sobecks, R.; Rybicki, L.; Kalaycio, M.; et al. Survival and predictors of outcome in patients with acute leukemia admitted to the intensive care unit. Cancer 2008, 112, 2233–2240. [Google Scholar] [CrossRef]
  27. Cleaver, A.L.; Beesley, A.H.; Firth, M.J.; Sturges, N.C.; O’Leary, R.A.; Hunger, S.P.; Baker, D.L.; Kees, U.R. Gene-based outcome prediction in multiple cohorts of pediatric T-cell acute lymphoblastic leukemia: A Children’s Oncology Group study. Mol. Cancer 2010, 9, 105. [Google Scholar] [CrossRef] [PubMed]
  28. Stone, K.; Zwiggelaar, R.; Jones, P.; Mac Parthaláin, N. A systematic review of the prediction of hospital length of stay: Towards a unified framework. PLoS Digit. Health 2022, 1, e0000017. [Google Scholar] [CrossRef] [PubMed]
  29. Gokhale, S.; Taylor, D.; Gill, J.; Hu, Y.; Zeps, N.; Lequertier, V.; Prado, L.; Teede, H.; Enticott, J. Hospital length of stay prediction tools for all hospital admissions and general medicine populations: Systematic review and meta-analysis. Front. Med. 2023, 10, 1192969. [Google Scholar] [CrossRef] [PubMed]
  30. Medeiros, N.B.; Fogliatto, F.S.; Rocha, M.K.; Tortorella, G.L. Forecasting the length-of-stay of pediatric patients in hospitals: A scoping review. BMC Health Serv. Res. 2021, 21, 938. [Google Scholar] [CrossRef]
  31. Lequertier, V.; Wang, T.; Fondrevelle, J.; Augusto, V.; Duclos, A. Hospital Length of Stay Prediction Methods: A Systematic Review. Med. Care 2021, 59, 929–938. [Google Scholar] [CrossRef]
  32. Farimani, R.M.; Karim, H.; Atashi, A.; Tohidinezhad, F.; Bahaadini, K.; Abu-Hanna, A.; Eslami, S. Models to predict length of stay in the emergency department: A systematic literature review and appraisal. BMC Emerg. Med. 2024, 24, 54. [Google Scholar] [CrossRef]
  33. Gokhale, S.; Taylor, D.; Gill, J.; Hu, Y.; Zeps, N.; Lequertier, V.; Teede, H.; Enticott, J. Hospital length of stay prediction for general surgery and total knee arthroplasty admissions: Systematic review and meta-analysis of published prediction models. Digit. Health 2023, 9, 20552076231177497. [Google Scholar] [CrossRef]
  34. Bacchi, S.; Tan, Y.; Oakden-Rayner, L.; Jannes, J.; Kleinig, T.; Koblar, S. Machine learning in the prediction of medical inpatient length of stay. Intern. Med. J. 2022, 52, 176–185. [Google Scholar] [CrossRef]
  35. Peres, I.T.; Hamacher, S.; Oliveira, F.L.C.; Thomé, A.M.T.; Bozza, F.A. What factors predict length of stay in the intensive care unit? Systematic review and meta-analysis. J. Crit. Care 2020, 60, 183–194. [Google Scholar] [CrossRef]
  36. Naemi, A.; Schmidt, T.; Mansourvar, M.; Naghavi-Behzad, M.; Ebrahimi, A.; Wiil, U.K. Machine learning techniques for mortality prediction in emergency departments: A systematic review. BMJ Open 2021, 11, e052663. [Google Scholar] [CrossRef]
  37. Mpanya, D.; Celik, T.; Klug, E.; Ntsinjana, H. Predicting mortality and hospitalization in heart failure using machine learning: A systematic literature review. IJC Heart Vasc. 2021, 34, 100773. [Google Scholar] [CrossRef]
  38. Clark, D.E.; Ryan, L.M. Concurrent Prediction of Hospital Mortality and Length of Stay from Risk Factors on Admission. Health Serv. Res. 2002, 37, 631–645. [Google Scholar] [CrossRef]
  39. Awad, A.; Bader, M.; McNicholas, J. Patient length of stay and mortality prediction: A survey. Health Serv. Manag. Res. 2017, 30, 105–120. [Google Scholar] [CrossRef]
  40. Du, X.; Wang, H.; Wang, S.; He, Y.; Zheng, J.; Zhang, H.; Hao, Z.; Chen, Y.; Xu, Z.; Lu, Z. Machine Learning Model for Predicting Risk of In-Hospital Mortality after Surgery in Congenital Heart Disease Patients. Rev. Cardiovasc. Med. 2022, 23, 376. [Google Scholar] [CrossRef] [PubMed]
  41. Gowd, A.K.; Agarwalla, A.; Beck, E.C.; Rosas, S.; Waterman, B.R.; Romeo, A.A.; Liu, J.N. Prediction of total healthcare cost following total shoulder arthroplasty utilizing machine learning. J. Shoulder Elb. Surg. 2022, 31, 2449–2456. [Google Scholar] [CrossRef] [PubMed]
  42. Rao, A.R.; Raunak, J.; Mrityunjai, S.; Rahul, G. Predictive interpretable analytics models for forecasting healthcare costs using open healthcare data. Healthc. Anal. 2024, 6, 100351. [Google Scholar] [CrossRef]
  43. Lu, Y.; Lavoie-Gagne, O.; Forlenza, E.M.; Pareek, A.; Kunze, K.N.; Forsythe, B.; Levy, B.A.; Krych, A.J. Duration of Care and Operative Time Are the Primary Drivers of Total Charges After Ambulatory Hip Arthroscopy: A Machine Learning Analysis. Arthrosc. J. Arthrosc. Relat. Surg. 2022, 38, 2204–2216.e3. [Google Scholar] [CrossRef]
  44. Muhlestein, W.E.; Akagi, D.S.; McManus, A.R.; Chambless, L.B. Machine learning ensemble models predict total charges and drivers of cost for transsphenoidal surgery for pituitary tumor. J. Neurosurg. 2019, 131, 507–516. [Google Scholar] [CrossRef]
  45. Placido, D.; Thorsen-Meyer, H.C.; Kaas-Hansen, B.S.; Reguant, R.; Brunak, S. Development of a dynamic prediction model for unplanned ICU admission and mortality in hospitalized patients. PLoS Digit. Health 2023, 2, e0000116. [Google Scholar]
  46. Gupta, A.; Liu, T.; Crick, C. Utilizing time series data embedded in electronic health records to develop continuous mortality risk prediction models using hidden Markov models: A sepsis case study. Stat. Methods Med. Res. 2020, 29, 3409–3423. [Google Scholar]
  47. Jenkins, D.A.; Martin, G.P.; Sperrin, M.; Riley, R.D.; Debray, T.P.; Collins, G.S.; Peek, N. Continual updating and monitoring of clinical prediction models: Time for dynamic prediction systems? Diagn. Progn. Res. 2021, 5, 1. [Google Scholar] [CrossRef] [PubMed]
  48. Karnuta, J.M.; Churchill, J.L.; Haeberle, H.S.; Nwachukwu, B.U.; Taylor, S.A.; Ricchetti, E.T.; Ramkumar, P.N. The value of artificial neural networks for predicting length of stay, discharge disposition, and inpatient costs after anatomic and reverse shoulder arthroplasty. J. Shoulder Elb. Surg. 2020, 29, 2385–2394. [Google Scholar] [CrossRef]
  49. Natekin, A.; Knoll, A. Gradient boosting machines, a tutorial. Front. Neurorobot. 2013, 7, 21. [Google Scholar] [CrossRef]
  50. Rigatti, S.J. Random Forest. J. Insur. Med. 2017, 47, 31–39. [Google Scholar] [CrossRef]
  51. Stulberg, J.J.; Haut, E.R. Practical Guide to Surgical Data Sets: Healthcare Cost and Utilization Project National Inpatient Sample (NIS). JAMA Surg. 2018, 153, 586–587. [Google Scholar] [CrossRef]
  52. ICD-10-CM/PCS MS-DRG; v37.0; Definitions Manual; Centers for Medicare & Medicaid Services: Baltimore, MD, USA, 2019.
  53. Averill, R.F.; Goldfield, N.; Hughes, J.S.; Bonazelli, J.; McCullough, E.C.; Steinbeck, B.A.; Mullin, R.; Tang, A.M. All Patient Refined Diagnosis Related Groups (APR-DRGs) Version 20.0; M.H.I. Systems: Tokyo, Japan, 2003. [Google Scholar]
  54. Elixhauser Comorbidity Software Refined for ICD-10-CM DIAGNOSES; v2024.1; Agency for Healthcare Research and Quality (AHRQ): Rockville, MD, USA, 2023.
  55. Lee, S.Y.; Lee, S.H.; Tan, J.H.H.; Foo, H.S.L.; Phan, P.H.; Kow, A.W.C.; Lwin, S.; Seah, P.M.Y.; Mordiffi, S.Z. Factors associated with prolonged length of stay for elective hepatobiliary and neurosurgery patients: A retrospective medical record review. BMC Health Serv. Res. 2018, 18, 5. [Google Scholar] [CrossRef] [PubMed]
  56. Fetene, D.; Tekalegn, Y.; Abdela, J.; Aynalem, A.; Bekele, G.; Molla, E. Prolonged Length of Hospital Stay and Associated Factors Among Patients Admitted at a Surgical Ward in Selected Public Hospitals Arsi Zone, Oromia, Ethiopia, 2022. medRxiv 2022. [Google Scholar] [CrossRef]
  57. Murdoch, W.J.; Singh, C.; Kumbier, K.; Abbasi-Asl, R.; Yu, B. Interpretable machine learning: Definitions, methods, and applications. arXiv 2019, arXiv:1901.04592. [Google Scholar] [CrossRef]
  58. Darst, B.F.; Malecki, K.C.; Engelman, C.D. Using recursive feature elimination in random forest to account for correlated variables in high dimensional data. BMC Genet. 2018, 19, 65. [Google Scholar] [CrossRef]
  59. Berrar, D. Cross-Validation. In Reference Module in Life Sciences Encyclopedia of Bioinformatics and Computational Biology; Ranganathan, S., Gribskov, M., Nakai, K., Christian Schönbach, C., Eds.; Elsevier: Amsterdam, The Netherlands, 2019; Volume 1, pp. 542–545. [Google Scholar] [CrossRef]
  60. Awad, M.; Fraihat, S. Recursive Feature Elimination with Cross-Validation with Decision Tree: Feature Selection Method for Machine Learning-Based Intrusion Detection Systems. J. Sens. Actuator Netw. 2023, 12, 67. [Google Scholar] [CrossRef]
  61. Hossin, M.; Sulaiman, M.N. A review on evaluation metrics for data classification evaluations. Int. J. Data Min. Knowl. Manag. Process 2015, 5, 1. [Google Scholar]
  62. Chicco, D.; Warrens, M.J.; Jurman, G. The coefficient of determination R-squared is more informative than SMAPE, MAE, MAPE, MSE and RMSE in regression analysis evaluation. Peerj Comput. Sci. 2021, 7, e623. [Google Scholar] [CrossRef]
  63. Musolf, A.M.; Holzinger, E.R.; Malley, J.D.; Bailey-Wilson, J.E. What makes a good prediction? Feature importance and beginning to open the black box of machine learning in genetics. Human Genet. 2022, 141, 1515–1528. [Google Scholar] [CrossRef]
  64. Pudjihartono, N.; Fadason, T.; Kempa-Liehr, A.W.; O’Sullivan, J.M. A Review of Feature Selection Methods for Machine Learning-Based Disease Risk Prediction. Front. Bioinform. 2022, 2, 927312. [Google Scholar] [CrossRef]
  65. Spencer, C.S.; Gaskin, D.J.; Roberts, E.T. The Quality Of Care Delivered To Patients Within The Same Hospital Varies By Insurance Type. Health Aff. 2013, 32, 1731–1739. [Google Scholar]
  66. Roll, K.; Stargardt, T.; Schreyögg, J. Effect of Type of Insurance and Income on Waiting Time for Outpatient Care. Geneva Pap. Risk Insur.-Issues Pract. 2012, 37, 609–632. [Google Scholar]
  67. Obeng-Gyasi, S.; O’Neill, A.; Zhao, F.; Kircher, S.M.; Lava, T.R.; Wagner, L.I.; Miller, K.D.; Sparano, J.D.; Sledge, G.W.; Carlos, R.C. Impact of insurance and neighborhood socioeconomic status on clinical outcomes in therapeutic clinical trials for breast cancer. Cancer Med. 2021, 10, 45–52. [Google Scholar] [PubMed]
  68. Bassan, R.; Hoelzer, D. Modern Therapy of Acute Lymphoblastic Leukemia. J. Clin. Oncol. 2011, 29, 532–543. [Google Scholar] [CrossRef] [PubMed]
  69. Prosperi, M.; Guo, Y.; Sperrin, M.; Koopman, J.S.; Min, J.S.; He, X.; Rich, S.; Wang, M.; Buchan, I.E.; Bian, J. Causal inference and counterfactual prediction in machine learning for actionable healthcare. Nat. Mach. Intell. 2020, 2, 369–375. [Google Scholar] [CrossRef]
Informatics 13 00047 g001
Figure 1. Patient grouping by individual and combined patient outcomes.
Figure 1. Patient grouping by individual and combined patient outcomes.
Informatics 13 00047 g001
Informatics 13 00047 g002
Figure 2. Data selection process, cleaning, and restructuring process. (a) Patient record selection process by ALL-related ICD-10 codes from NIS in 2019 and 2021. (b) Feature missingness and imputation methods summary. (c) Dataset extension methodology illustration to show how the dataset was extended into a long format by multiple ICD-10 procedure codes in each patient record. Note: Blue error represents the dataset was transformed from short format to the long format.
Figure 2. Data selection process, cleaning, and restructuring process. (a) Patient record selection process by ALL-related ICD-10 codes from NIS in 2019 and 2021. (b) Feature missingness and imputation methods summary. (c) Dataset extension methodology illustration to show how the dataset was extended into a long format by multiple ICD-10 procedure codes in each patient record. Note: Blue error represents the dataset was transformed from short format to the long format.
Informatics 13 00047 g002
Informatics 13 00047 g003
Figure 3. Patient outcome prediction process. (a) A machine learning technique flowchart to show how data balancing, feature selection, prediction model, and model evaluation and interpretation were performed. (b) Categorization of static and dynamic features utilized for prediction.
Figure 3. Patient outcome prediction process. (a) A machine learning technique flowchart to show how data balancing, feature selection, prediction model, and model evaluation and interpretation were performed. (b) Categorization of static and dynamic features utilized for prediction.
Informatics 13 00047 g003
Informatics 13 00047 g004
Figure 4. Distribution of Patient Outcomes.
Figure 4. Distribution of Patient Outcomes.
Informatics 13 00047 g004
Informatics 13 00047 g005
Figure 5. ROC Curve across Days for Individual PLOS and Mortality Prediction.
Figure 5. ROC Curve across Days for Individual PLOS and Mortality Prediction.
Informatics 13 00047 g005
Informatics 13 00047 g006
Figure 6. ROC Curve across Days for PLOSM Prediction in each class.
Figure 6. ROC Curve across Days for PLOSM Prediction in each class.
Informatics 13 00047 g006
Informatics 13 00047 g007
Figure 7. Model Performance Metrics for Combined and PLOSM Outcome.
Figure 7. Model Performance Metrics for Combined and PLOSM Outcome.
Informatics 13 00047 g007
Informatics 13 00047 g008
Figure 8. Model Performance Metrics for TCs. RMSE—Root mean square error is the square root of the average of the squared differences between predictions and actual observations; MAE—Mean absolute error is the average absolute difference between predicted and actual values; R2—Coefficient of determination is the proportion of variance in the dependent variable explained by the model.
Figure 8. Model Performance Metrics for TCs. RMSE—Root mean square error is the square root of the average of the squared differences between predictions and actual observations; MAE—Mean absolute error is the average absolute difference between predicted and actual values; R2—Coefficient of determination is the proportion of variance in the dependent variable explained by the model.
Informatics 13 00047 g008
Informatics 13 00047 g009
Figure 9. Selected Features Summary for PLOSM and TC Prediction. (a) Number of features selected for predicting PLOSM for each day. (b) Number of features selected for predicting TCs for each day. (c) Selected static features with importance values were shown for each day’s PLOSM prediction. (d) Selected procedure features with importance values were shown for each day’s PLOSM prediction. (e) Selected static features with importance values were shown for each day’s TC prediction.
Figure 9. Selected Features Summary for PLOSM and TC Prediction. (a) Number of features selected for predicting PLOSM for each day. (b) Number of features selected for predicting TCs for each day. (c) Selected static features with importance values were shown for each day’s PLOSM prediction. (d) Selected procedure features with importance values were shown for each day’s PLOSM prediction. (e) Selected static features with importance values were shown for each day’s TC prediction.
Informatics 13 00047 g009
Informatics 13 00047 g010
Figure 10. Summary of Patient Outcome Prediction Patterns for Patients with ALL Enrolled in Clinical Trial Research. Note: The color of the dot is the pattern calculated in the colored cell of the table on the bottom right corner. The red line in the pattern formation was to separate the true label and the prediction label.
Figure 10. Summary of Patient Outcome Prediction Patterns for Patients with ALL Enrolled in Clinical Trial Research. Note: The color of the dot is the pattern calculated in the colored cell of the table on the bottom right corner. The red line in the pattern formation was to separate the true label and the prediction label.
Informatics 13 00047 g010
Table 1. Feature summaries (n = 20,371).
Table 1. Feature summaries (n = 20,371).
VariablePLOS 1 (n = 7173)Mortality 1 (n = 480)TC ($) 2
Minor (<18 years of age)2868 (14.08)88 (0.43)139.62K (368.51K, 43.23K)
20193617 (17.76)237 (1.16)143.16K (320.64K, 52.41K)
20213556 (17.46)243 (1.19)162.16K (347.72K, 61.59K)
Demographic information
Female3140 (15.41)204 (1.00)153.73K (344.77K, 58.02K)
Race/Ethnicity
White3687 (18.10)256 (1.26)132.48K (286.21K, 51.99K)
Black579 (2.84)43 (0.21)143.99K (301.82K, 52.59K)
Hispanic2173 (10.67)137 (0.67)188.32K (418.74K, 65.19K)
Asian or Pacific Islander334 (1.64)19 (0.09)160.08K (291.52K, 65.44K)
Native American63 (0.31)1 (0.01)119.53K (201.62K, 57.86K)
Other337 (1.65)24 (0.12)170.26K (345.47K, 67.39K)
Insurance and financial information
Expected primary payer
Medicare1115 (5.47)146 (0.72)139.06K (255.48K, 66.88K)
Medicaid2426 (11.91)120 (0.59)160.85K (365.29K, 55.91K)
Private insurance3013 (14.79)170 (0.83)145.97K (311.71K, 54.25K)
Self-pay256 (1.26)18 (0.09)162.33K (376.78K, 58.61K)
No charge41 (0.20)3 (0.01)107.24K (194.46K, 51.68K)
Other322 (1.58)23 (0.11)187.09K (454.80K, 57.25K)
Area 3
Urban4237 (20.80)281 (1.38)164.43K (343.39K, 63.82K)
Transitional2028 (9.96)135 (0.66)138.75K (330.60K, 48.45K)
Rural908 (4.46)64 (0.31)129.44K (296.36K, 47.23K)
Median household income
0–25th percentile2038 (10.00)144 (0.71)156.65K (372.86K, 55.23K)
26–50th percentile1683 (8.26)117 (0.57)146.11K (316.77K, 54.87K)
51–75th percentile1832 (8.99)128 (0.63)152.47K (350.77K, 56.75K)
76–100th percentile1620 (7.95)91 (0.45)153.37K (279.23K, 60.98K)
Hospital information
Location/teaching status of hospital
Rural37 (0.18)5 (0.02)130.77K (240.87K, 52.40K)
Urban. nonteaching157 (0.77)28 (0.14)153.92K (349.08K, 56.60K)
Urban. teaching6979 (34.26)447 (2.19)205.19K (338.87K, 57.69K)
Control/ownership of hospital
Government. non-federal1235 (6.06)80 (0.39)166.80K (368.99K, 65.20K)
Private. not-for-profit5517 (27.08)375 (1.84)119.42K (252.96K, 45.44K)
Private. investor-owned421 (2.07)25 (0.12)136.84K (355.77K, 90.74K)
Region of hospital
Northeast1231 (6.04)71 (0.35)166.80K (368.99K, 65.20K)
Midwest1282 (6.04)78 (0.38)119.42K (252.96K, 45.44K)
South2743 (13.47)209 (1.03)136.84K (296.84K, 50.47K)
West1917 (9.41)122 (0.60)193.88K (409.97K, 74.67K)
Bed-size of hospital 4
Small857 (4.21)53 (0.26)119.54K (327.76K, 43.76K)
Medium1345 (6.6)109 (0.54)160.82K (383.67K, 54.54K)
Large4971 (24.40)318 (1.56)156.96K (318.96K, 61.38K)
Medical Information
Admission on weekend1262 (6.20)98 (0.48)172.63K (357.39K, 64.65K)
Elective surgery1941 (9.53)78 (0.38)150.83K (368.14K, 50.77K)
Injury (incidence)249 (1.22)28 (0.14)182.39K (347.40K, 70.38K)
Service Line (based on ICD-10)
Surgical862 (4.23)83 (0.41)434.47K (747.21K, 194.99K)
Medical6258 (30.72)394 (1.93)133.37K (273.98K, 53.11K)
Transfer into the hospital
Not transferred in5912 (29.02)358 (1.76)143.15K (332.04K, 53.24K)
From an acute care hospital1054 (5.17)97 (0.48)239.41K (353.58K, 137.21K)
From another health faculty207 (1.02)25 (0.12)185.02K (278.09K, 75.65K)
Risk Mortality
No class specified2 (0.01)1 (0.01)1346.15K (2493.40K, 138.43K)
Minor likelihood of dying1620 (7.95)8 (0.04)97.20K (156.19K, 41.76K)
Moderate likelihood of dying2622 (12.87)11 (0.05)106.37K (179.28K, 49.11K)
Major likelihood of dying1767 (8.67)63 (0.31)218.28K (346.93K, 98.24K)
Extreme likelihood of dying1162 (5.70)397 (1.95)512.87K (863.67K, 243.41K)
Risk Severity
No class specified2 (0.01)1 (0.01)1346.15K (2493.40K, 138.43K)
Minor loss of function88 (0.43)0 (0)82.09K (133.43K, 40.37K)
Moderate loss of function1169 (5.74)5 (0.02)66.19K (105.04K, 35.50K)
Major loss of function3256 (15.98)34 (0.17)133.22K (201.46K, 62.75K)
Extreme loss of function2658 (13.05)440 (2.16)373.36K (638.22K, 173.41K)
Emergency department record2906 (14.27)229 (1.12)147.17K (296.75K, 61.49K)
Major operating room1341 (6.58)94 (0.46)396.17K (672.25K, 201.50K)
Enrolled in clinical trial287 (1.41)6 (0.03)253.49K (583.56K, 109.19K)
Comorbidity_Acquired immune deficiency syndrome20 (0.1)2 (0.01)272.57K (382.86K, 88.84K)
Comorbidity_Alcohol abuse55 (0.27)3 (0.01)151.91K (263.87K, 75.48K)
Comorbidity_Dementia33 (0.16)9 (0.04)116.04K (272.80K, 58.48K)
Comorbidity_Depression761 (3.74)44 (0.22)185.82K (387.38K, 73.14K)
Comorbidity_Diabetes with chronic complications834 (4.09)67 (0.33)194.11K (346.32K, 85.67K)
Comorbidity_Diabetes without chronic complications391 (1.92)28 (0.14)135.58K (219.23K, 34.11K)
Comorbidity_Drug abuse121 (0.59)6 (0.03)206.22K (374.59K, 37.25K)
Comorbidity_Hypertension. complicated745 (3.66)84 (0.41)222.92K (466.18K, 88.50K)
Comorbidity_Hypertension. uncomplicated1694 (8.32)116 (0.57)174.59K (312.60K, 72.87K)
Comorbidity_Chronic pulmonary disease660 (3.24)50 (0.25)147.02K (305.98K, 60.90K)
Comorbidity_Obesity910 (4.47)76 (0.37)203.48K (411.54K, 82.96K)
Comorbidity_Peripheral vascular disease366 (1.80)26 (0.13)128.06K (321.99K, 48.75K)
Comorbidity_Hypothyroidism478 (2.35)30 (0.15)153.07K (285.46K, 64.58K)
Comorbidity_Other thyroid disorders79 (0.39)3 (0.01)192.54K (348.39K, 75.10K)
1 Frequency (relative frequency %); 2 Mean (Standard Deviation, Median); 3 Regroup of the six-category urban–rural classification scheme for the U.S. counties developed by the National Center for Health Statistics (NCHS). Urban: “Central/Fringe” counties of metro areas of ≥ 1 million population; Transitional: Counties in metro areas of 50,000–999,999 population; Rural: Micropolitan/Nonmetropolitan counties; 4 Bed-size categories are based on hospital bedsand are specific to the hospital’s location and teaching status.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Ma, J.; Johnson, E.; Whitaker, B.M.; Dadgostari, F.; Schwertz, H.; McCrory, B. Concurrent Prediction of Length of Stay, Mortality, and Total Charges in Patients with Acute Lymphoblastic Leukemia Using Continuous Machine Learning. Informatics 2026, 13, 47. https://doi.org/10.3390/informatics13040047

AMA Style

Ma J, Johnson E, Whitaker BM, Dadgostari F, Schwertz H, McCrory B. Concurrent Prediction of Length of Stay, Mortality, and Total Charges in Patients with Acute Lymphoblastic Leukemia Using Continuous Machine Learning. Informatics. 2026; 13(4):47. https://doi.org/10.3390/informatics13040047

Chicago/Turabian Style

Ma, Jiahui, Elizabeth Johnson, Bradley M. Whitaker, Faraz Dadgostari, Hansjorg Schwertz, and Bernadette McCrory. 2026. "Concurrent Prediction of Length of Stay, Mortality, and Total Charges in Patients with Acute Lymphoblastic Leukemia Using Continuous Machine Learning" Informatics 13, no. 4: 47. https://doi.org/10.3390/informatics13040047

APA Style

Ma, J., Johnson, E., Whitaker, B. M., Dadgostari, F., Schwertz, H., & McCrory, B. (2026). Concurrent Prediction of Length of Stay, Mortality, and Total Charges in Patients with Acute Lymphoblastic Leukemia Using Continuous Machine Learning. Informatics, 13(4), 47. https://doi.org/10.3390/informatics13040047

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop