A Multimodal Machine Learning Model in Pneumonia Patients Hospital Length of Stay Prediction

Abstract

Hospital overcrowding, driven by both structural management challenges and widespread medical emergencies, has prompted extensive research into machine learning (ML) solutions for predicting patient length of stay (LOS) to optimize bed allocation. While many existing models simplify the LOS prediction problem to a classification task, predicting broad ranges of hospital days, an exact day-based regression model is often crucial for precise planning. Additionally, available data are typically limited and heterogeneous, often collected from a small patient cohort. To address these challenges, we present a novel multimodal ML framework that combines imaging and clinical data to enhance LOS prediction accuracy. Specifically, our approach uses the following: (i) feature extraction from chest CT scans via a convolutional neural network (CNN), (ii) their integration with clinically relevant tabular data from patient exams, refined through a feature selection system to retain only significant predictors. As a case study, we applied this framework to pneumonia patient data collected during the COVID-19 pandemic at two hospitals in Naples, Italy—one specializing in infectious diseases and the other general-purpose. Under our experimental setup, the proposed system achieved an average prediction error of only three days, demonstrating its potential to improve patient flow management in critical care environments.

1. Introduction

Since the healthcare environment faced the emergency brought by the COVID-19 pandemic, researchers and doctors understood how crucial it is to properly organize hospital beds availability in order not to have overcrowded hospital wards and fatal delays. For this reason, in order to prevent these unpleasant and dramatic situations, providing a solution which enables medical staff to acknowledge beforehand how long patients affected by diseases like pneumonia are required to stay in hospital is crucial. Even though there have already been studies on the estimation of hospital length of stay (LOS) with machine learning (ML) approaches, only with the advent of the COVID-19 global pandemic the urge to find a system for this task became prominent, with most of the works released since 2020. The implementation of a prediction system for hospital admissions, as well as other technical solutions aimed at speeding up the hospital sector functioning, could represent an important step towards greater resilience of the national health system. Regarding Italy, the COVID-19 health emergency demonstrated how unprepared the country was to manage such a massive influx of patients, highlighting the potential support such a system could have provided. Planning solutions in advance generally allows for the prevention of many problems and a faster application of timely solutions in emergency contexts.

This specific task, given the limited amount of tabular data and the inherent challenges of regression in this context—where daily precision is crucial—it is particularly difficult to address with existing machine learning models. Most current solutions for length of stay (LOS) prediction rely on classification rather than regression models, with targets that discretize the number of hospital days into ranges. For instance, some models from the literature [1] categorize LOS into short (1–3 days), medium (4–7 days), and long (8+ days) ranges. However, such categorical predictions may be insufficient in settings wherein daily management of hospital capacity is essential to ensure bed availability. Beyond the demands heightened by COVID-19, an automated system capable of optimizing hospital resources could enhance patient care during hospitalization and recovery while reducing stress for healthcare staff.

For this reason, the aim of the proposed work is to provide a solution which, starting from heterogeneous data containing images and tabular data, can understand how long the patient should stay in the hospital. The implemented architecture is designed to deliver accurate LOS predictions, even in the face of limited and diverse data. Importantly, it exploits data heterogeneity as an advantage, integrating information from both medical imaging and clinical records. This unique feature distinguishes our approach from existing methods, which typically rely on a single data type [2,3,4,5]. Moreover, the architecture has the potential for broader application beyond LOS prediction. Given that patient features in the dataset capture their current health status, the system, with targeted modifications in its final layers, could also function as an outcome prediction tool, particularly in predicting patient survival following therapy. This extension may also provide insight into the severity of a patient’s condition. Thus, the system could serve a dual purpose: acting as a classifier for outcome prediction and as a regression model for LOS estimation.

The data used in this study were collected during the COVID-19 pandemic in two Neapolitan hospitals when computed tomography (CT) scans were commonly performed on patients’ chests. As such, the patients may have been diagnosed with either generic pneumonia or COVID-19. However, this variability does not affect the primary goal of this research, which is to analyze generic chest CT scans and clinical data to predict LOS, irrespective of the specific underlying lung pathology. The only limitation is that the illness must affect the lungs, with pneumonia being the most common condition in this dataset.

The proposed ML system accounts for data heterogeneity, both across different data types and within the same data type from varying sources. It also addresses challenges such as dataset imbalance and the varying importance of features. In this work, we describe the design and implementation details of the system.

To sum up, the contribution of this paper can be summarized as follows:

• We illustrate an approach to extract useful features from lungs CT scans and volumes for LOS prediction;
• We provide a reliable solution for LOS prediction powered by CT features, whose results are more efficient than SOTA solutions;
• We expand the solution to be compatible with clinical tabular data, available along with patients’ CT scans, and compared performances given the different types of input;
• We adapt the proposed methodology for the task of post-therapy survival prediction, showing the generalization ability of the approach;
• We illustrate and implement a multi-modal architecture capable of compensating the typical lack of data in the biomedical domain.

The rest of this paper is organized as follows: Section 2 briefly analyses the current literature, with an emphasis on the limits of current proposals; Section 3 introduces the considered datasets; Section 4 describes the implemented methodology; Section 5 reports the obtained results; Section 6 highlights some ethical consideration; finally, Section 7 provides some conclusions.

2. Related Works

Several works have been released and promoted in order to provide solutions to the complex task of hospitalization days, or LOS, prediction, for different kinds of diseases.

Since 2014, several studies have addressed the problem of predicting length of stay (LOS) using ML techniques. The such study by Morton et al. [6] presents a comparative analysis of ML models, including Linear Regressors, Support Vector Machines, Multi-Task Learning Models, and Random Forests, applied to clinical data for LOS prediction. Although this study focuses on diabetic patients, the authors suggest that the approach can be generalized to other diseases. This generalization is explored by Pendharkar et al. [7], who compare three different models, namely Classification and Regression Trees (CARTs), Chi-Square Automatic Interaction Detection (CHAID), and Support Vector Regressors (SVRs), all trained on clinical (tabular) data. Similarly, Turgeman et al. [8] employed a cubist regression tree to predict LOS using static clinical data from a broader patient population. In addition, Lorenzoni et al. [9] applied eight different ML models to predict LOS, focusing on patients with heart disease. Similarly, Daghistani et al. [1] developed a system using various ML classification models to predict LOS in terms of day ranges, specifically for cardiovascular disease patients. Saadatmand et al. [10] propose a system with five different ML algorithms for ICU admission, mortality, and LOS discretized in two classes, while Khajehali et al. [11] applied several ensembling models to detect the LOS discretized in five classes.

While the aforementioned works represent some of the earliest contributions to the field, numerous additional studies have emerged since 2020, including several notable works [12,13,14,15,16,17,18,19,20,21] and systematic reviews [2,3,4,5], all of which leverage ML techniques for LOS prediction. However, regardless of the publication year, few studies attempt to address LOS prediction using diverse data types beyond standard clinical data, or employ deep learning approaches. One example is the work by Zebin et al. [22], which uses an autoencoder combined with a deep neural network trained on the MIMIC-III dataset. It is worth noting that the task in this study was simplified to a binary classification problem, with LOS predictions limited to two possible day ranges.

From 2020 onward, predicting LOS in hospitals became an urgent problem, particularly in response to a unique, unprecedented challenge: COVID-19. The urge to understand how long patients would stay in hospital, before leaving beds for new patients, became particularly compelling during the emergency state the world lived in 2020 and 2021. Therefore, not only was there an immediate spread of works on LOS prediction for COVID-19 patients, but in general a huge increase in these kinds of works. One of the earliest studies addressing LOS prediction for COVID-19 patients was conducted by Roimi et al. [23], who proposed a system that used daily clinical data to predict the required hospital utilization for each patient. Similarly, Ebinger et al. [24] developed a system consisting of three ML models specifically designed to predict LOS for COVID-19 patients. In the following year, Alabbad et al. [25] introduced four ML models to predict the LOS of COVID-19 patients in Saudi Arabia, while Etu et al. [26] presented a similar approach with four ML models for patients at a Detroit hospital.

A common feature in these studies is their reliance almost exclusively on clinical tabular data, despite the frequent use of imaging techniques such as chest CT to monitor disease progression. This underutilization of imaging data represents a potential missed opportunity, as scans could provide valuable insights that might enhance the models’ ability to predict how long patients will require hospital care.

3. Dataset

Through collaborations with the University of Campania Luigi Vanvitelli Hospital (Azienda Ospedaliera Universitaria Vanvitelli) and Cotugno Hospital (Azienda Ospedaliera dei Colli Cotugno), we compiled a heterogeneous dataset of 84 patients. Each patient is represented by tabular data from clinical exams and a chest CT scan. Specifically, the dataset from the Hospital of University Vanvitelli (denoted as Vanvitelli dataset), comprising 34 patients, includes a CT scan session per patient, 71 tabular features, and two outcome variables: patient outcome and hospitalization duration. The dataset from the Hospital Cotugno (denoted as Cotugno dataset) consists of 50 patients, each with CT scan sessions, 75 tabular features, and three outcome variables: patient outcome, mortality status (exitus), and length of stay (LOS).

Besides analysing each dataset separately, we also merged them, analyzing the overall performance on a more variegated collection. Due to differences in data sources, the tabular features were not fully aligned, resulting in a merged dataset with only a subset of the available features. Prior to merging, both datasets underwent pre-processing to ensure they contained significant, independent, normalized features without missing values. As a result, the Vanvitelli dataset was reduced to 36 features, while the Cotugno dataset retained at 49 features. After merging, the final dataset included 14 common features along with two shared outputs (LOS and outcome). Additionally, an extra feature, derived from the LOS output, was created for model training. This feature, termed “range”, represents a discretized version of the LOS, with 7-day intervals (e.g., 01–07 days, 08–14 days). The final category included patients with 50 or more days of hospitalization.

Table 1. Table containing the features for each of the original datasets (Vanvitelli and Cotugno) after pre-processing and filtering. The features that are present in both datasets (9) are marked in bold, resulting in the Complete dataset.

Dataset	Size	Binary Features	Numeric Features	Features Num.
Vanvitelli	34	Gender, Remdesivie, Neutrophils, Ischemic heart disease, COPD, COVID therapy, Ventilation, Anticoagulant therapy, chronic liver disease, Fever, Cough, Diarrhea/vomiting, Arterial hypertension, Atrial fibrillation, Active cancer within the past 5 years, Dyspnea, Ageusia/anosmia, Type II diabetes, Dementia, Lymphocytes	Age, WBC, HB, Platelets, Azotemia, Creatinine, Sodium, Potassium, GOT or GPT, Blood glucose, LDH, INR, Fibrinogen, D dimer, p/F, LUS score	36
Cotugno	50	Gender, Remdesivie, Neutrophils, Ischemic heart disease, COPD, COVID therapy, Ventilation, CDI, Complication, TEP, PNX, PMD, MB, ICU, Pentaglobin, CS, I dose, II dose, III dose, Complete vaccine cycle, Comorbidity, Cardio, DM, Nephro, Pneumo, Emato, K, Tiroid, Stroke, IOP2	Age, WBC, TC score, PF, Complication type, Max ventilation, Charlson, IL6, IL2R, NEU, %, LINF, % (# 1), n/l, ESR, DD, PCR, PCT, Ferritin	49

reports the details of the datasets acquired from Vanvitelli and Cotugno hospitals, respectively, after pre-processing and filtering procedures. The features that are present in both datasets are marked in bold, resulting in the Complete dataset.

Before adopting the listed datasets, the tabular features from clinical exams must be pre-processed in order to be properly balanced and populated. Specifically, features that were empty or nearly empty were removed. Only complete features, or those with minimal missing values that could be reasonably inferred (e.g., missing values in the “COVID therapy” or “Anticoagulant therapy” columns were assumed to indicate no therapy, hence assigned a value of 0), were retained after pre-processing. Furthermore, all non-binary features were normalized based on the range limits derived from the training subsets. Figure 1a,b display the dataset distribution for the two common outcomes, namely patient outcome and length of hospitalization, while Figure 2 illustrates the distribution across the defined hospitalization day ranges. An analysis of the original datasets revealed that, while both are balanced in terms of sex distribution, they are imbalanced with respect to other features, such as age, as well as the output variables. Feature distributions from both datasets are presented in Figure 3, Figure 4 and Figure 5.

Following pre-processing and merging of the datasets, the correlation matrices for the Vanvitelli, Cotugno, and Complete datasets are shown in Figure 6, Figure 7 and Figure 8, respectively. In these matrices, cells colored green represent strong positive correlations between feature pairs, while blue cells indicate negative correlations, meaning the features are still correlated but in an inverse manner. By identifying feature pairs at both extremes—highly correlated and highly anti-correlated—the dataset can be further filtered for model optimization.

4. Proposed Approach

In this paper, we propose an automatic approach to predict the LOS together with the outcome of the patient’s recovery. The proposed methodology is outlined in the following steps, which are described in detail in the subsequent sections:

• Features extraction and collection;
• Features selection;
• LOS and outcome prediction.

The overall system architecture comprises two primary workflows: the dataset generation workflow (Figure 9) and the model training workflow (Figure 10). The dataset generation workflow focuses on feature extraction, as shown in Figure 9. The model training workflow, illustrated in Figure 10, includes steps for feature selection, dataset splitting, normalization, application of balancing algorithms, and the training of machine learning models for specific tasks. These trained models are then used for inference.

Specifically, in the dataset generation workflow, the original datasets are processed to produce an elaborated and filtered version, which contains tables of clinical data, features extracted from CT scans using CNNs, and their concatenation. At this stage, a task can be selected, and the corresponding dataset is prepared for input into the model training workflow.

The training workflow applies pre-processing steps such as normalization and balancing and, based on the selected configuration, determines the approaches to adopt on data, like ensembling or partitioning. In other words, once identified which prediction task to perform, it is possible to choose the dataset to take as input, which features to adopt, whether to apply an ensembling method and a feature selection approach, and the data partitioning approach. The whole pipeline, where each step can be changed according to the preferred configuration, is shown and detailed in Figure 10. In this way, it is possible to have a set of independent combinations, where each combination will provide its performance metrics for a specific task. Once the optimal model and configuration are identified, the system is ready for inference on new data.

4.1. Features Extraction and Collection

The initial intent of the activities for this work was to rely only on deep learning models for the required tasks, in particular by training them on the images’ side of the merged dataset, i.e., the 3D chest CT scans. However, given the low amount of available patients, even after the merging process, performances of any 2D/3D CNN model were unsatisfying for the proposed system, both in the case of fine-tuning and training from scratches. In particular, we tested a 2D model, developed and trained with the PyTorch framework, analyzing the slide wherein the amount of non-zero pixels is the largest, often corresponding to the slice wherein the lungs are at their maximum expansion. On the other hand, 3D models use all the available slices; therefore, no extra processing is needed, apart from the one used to uniform the dataset patients.

Since there were no chances of expanding the dataset with new patients to improve model performances, the focus moved towards the implementation of a machine learning model exploiting pre-trained CNNs as feature extractors for the chest CT scan. In order to accomplish this new task, depending on the original input adopted, which could be either the single slide with the greatest number of non-zero pixels or the whole 3D CT scan, a 2D and a 3D version of ResNet50 were chosen. In particular, the 2D version had pre-trained weights from ImageNet, while the 3D version had pre-trained weights from MedicalNet [27]. Both models were modified with the addition of an extra fully connected layer placed before the output layer, which is useful to fix the number of outputs representing the desired features obtained from a given input scan. A representation of this procedure on the original ResNet50 architecture is shown in Figure 11. These models were fine-tuned for 80 epochs (which resulted in the best amount before resulting in overfitting) and then adopted as feature extractors, thanks to the added fully connected layer. In both cases, the extra fully connected layer has an output vector of length 16, meaning that 16 “features” can be extracted from both single slices and CT volumes. Given these two sets, for each patient of every dataset, there are

• A variable number of features (depending on the source) from clinical exams;
• A total of 16 features from 2D slices;
• A total of 16 features from 3D volumes.

4.2. Features Selection

Focusing on the hospitalization days prediction task, it can be faced as a regression problem where the purpose is, given the available features, to predict the number, in a continuous domain, of days a patient should probably stay in the hospital before being discharged. In order to maximize the performances of adopted regression models, a feature selection workflow was implemented by using the KNIME software (version 5.2.5) [28], which enabled the opportunity to develop a system wherein all the possible attempts and features combinations can be executed effortlessly in a single run. The adopted approaches for feature selection are the following:

• Backward Feature Elimination (BFE): the dimensionality reduction is performed by an algorithm that, for each iteration with n features, removes one feature at a time in order to understand which set of $n-1$ input features results in the minimum increase in the error rate, and then removes the missing feature and continues with the next iteration;
• Forward Feature Selection (FFS): the dimensionality reduction is performed by a dual algorithm that, starting from one feature, at each iteration adds a new feature adopting the same strategy of the BFE.

These approaches can be applied by making inferences with ML models, as they are wrapper methods. For this reason, two ML models were chosen for this purpose: the Linear Regression model and the M5P model. Therefore, for each dataset there are four possible attempts to perform the following:

• BFE with linear regressor (LR);
• FFS with linear regressor (LR);
• BFE with M5P;
• FFS with M5P.

For the outcome prediction task, where the model aims to predict whether a patient will either be discharged or succumb to the disease, a binary classification model is employed. Due to the simpler nature of this task, with only two possible outcomes, classification models generally demonstrate higher performance. As a result, a feature selection process is not necessary for this task.

4.3. LOS and Outcome Prediction

As mentioned before, the LOS prediction consists in determining the length of patients staying in the hospitals. This specific task, given the desired output, has to be treated as a regression problem. However, in order to simplify this task, it is possible to discretize the LOS in a set of days ranges (for instance, the LOS value “18” is equivalent to the days range value “15–21”). In this way, the regression problem becomes a classification problem.

On the other hand, the Outcome prediction consists in determining whether a patient, after the clinical therapy, will survive or not. This specific task is a classification task, and the model in particular shall be a binary predictor.

Depending on the chosen task, there are sets of models which are trained and tested in parallel with the same selected features. In particular, for the LOS prediction, the adopted models are Regression Tree, K-Nearest Neighbor (KNN), Multi-Layer Perceptron (MLP), Linear Regression model, M5P and SMO. For the classification tasks (Days range and Outcome), instead, the adopted models are Decision Tree, Naïve Bayes, Gradient Boosted Tree, Random Forest, MLP, KNN, LibLINEAR and ADABoostM1. All these models are provided by the KNIME environment.

All these models can take as input patients from one of the available datasets, selecting one of the features sets described before and following a specific approach, covering both the ensembling method and the feature selection type. All these possibilities are described and formalized in

Table 2. Table of the combinations performed during the training and inference process. For instance, one possible pipeline is days range classification task, 2D features from CT scans, the Cotugno dataset, and using an ensemble approach.

Task	Features	Dataset	Approach
Outcome (Classification)	Tabular Data, CT Scan Features (2D), CT Volume Features (3D)	Complete, Vanvitelli, Cotugno	Standard, Ensemble
Days range (Classification)	Tabular Data, CT Scan Features (2D), CT Volume Features (3D)	Complete, Vanvitelli, Cotugno	Standard, Ensemble
Length of stay (Regression)	Tabular Data, CT Scan Features (2D), CT Volume Features (3D)	Complete, Vanvitelli, Cotugno	Standard, Ensemble, Feature Selection

:

In particular, regarding the chosen approaches, the standard one does not include any particular condition or configuration for the selected models, while the ensemble one refers to a 4-Bagging Ensemble, where each model is replicated four times and trained on different subsets.

A joint look on Figure 10 and Table 2 makes clear all the possible independent experiments that can be performed with this system. The final user can choose the Prediction Task (Outcome, days range or regression), the Extracted Dataset to adopt (Complete, Vanvitelli or Cotugno) with a specific set of features (Tabular, 2D CT or 3D CT), and a specific Ensembling and Features Selection approach (Standard, 4-Bagging Ensemble, BFE with LR, FFS with LR, BFE with M5P, or FFS with M5P).

For the training process, all datasets are split with both a 80/20 hold-out approach and a 10-fold cross-validation approach [29], where in any case the training set is balanced with SMOTE (Synthetic Minority Oversampling Technique) [30] on the range feature so that the training set could be balanced by LOS, even if not directly, despite its continuous nature.

Figure 12 shows how the SMOTE algorithm, thanks to the oversampling process over the days range feature, can improve the dataset distribution over all its partitions.

5. Results

All the results obtained with the combinations resulting from Figure 10 and Table 2 are reported in this section, along with sets of best features from the selection process. Since the two tasks require different kinds of models, there will be two dedicated subsections covering the obtained results. As mentioned before, for the Outcome prediction task, the feature selection process described before is not required, as results obtained without feature selection are already satisfying. On the other hand, the hospitalization days regression task required this further step in order to retrieve some promising results. In order to understand the difference with the introduction of this extra step, results without feature selection will be provided as well.

Thanks to the design choice of implementing and training several SOTA models in the architecture, it is possible to have a direct and immediate comparison among them for the evaluation process. Therefore, in the next subsections, evaluation metrics from all the available models are reported, and the best ones are chosen for their respective tasks. In all the tables of this section, the hold-out approach is referred as HO (80-20), where 80-20 is the split ratio, while the cross-validation approach is referred to as CV (10), where 10 is the number of folds. Tabular clinical features are referred as Tabular, CT scan features as 2D CT, and CT volume features as 3D CT.

5.1. Outcome Classification Task

As aforementioned, outcome prediction is a binary classification task. The results are most promising regardless of the approach used (standard or bagging ensemble) and the dataset selected (Complete, Vanvitelli, or Cotugno). The following tables display the accuracy results for the chosen models, specifically using the original clinical data (Tabular Data), the 2D features extracted from CT scans (2D CT), and the 3D features extracted from CT volumes (3D CT).

In particular,

Table 3. Accuracy scores of trained models for outcome prediction on complete dataset with the standard approach. Best results for each data type and split approach are in bold.

Model (Outcome, Complete, Standard)	Tabular HO (80-20)	Tabular CV (10)	2D CT HO (80-20)	2D CT CV (10)	3D CT HO (80-20)	3D CT CV (10)
Decision Tree	0.682	0.726	0.545	0.671	0.542	0.701
Naïve Bayes	0.773	0.685	0.545	0.438	0.375	0.416
Gradient Boosted Tree	0.682	0.867	0.545	0.699	0.875	0.701
Random Forest	0.682	0.808	0.591	0.726	0.792	0.753
Multi-Layer Perceptron	0.682	0.740	0.455	0.726	0.792	0.779
LibLINEAR	0.682	0.781	0.636	0.740	0.792	0.753
ADABoostM1	0.682	0.795	0.636	0.726	0.792	0.766
K-Nearest Neighbor 3	0.727	0.726	0.636	0.685	0.708	0.727
K-Nearest Neighbor 5	0.682	0.726	0.455	0.671	0.792	0.753
K-Nearest Neighbor 7	0.682	0.767	0.455	0.699	0.792	0.766

and

Table 4. Accuracy scores of trained models for outcome prediction on complete dataset with the 4-bagging ensemble approach. Best results for each data type and split approach are in bold.

Model (Outcome, Complete, Bagging)	Tabular HO (80-20)	Tabular CV (10)	2D CT HO (80-20)	2D CT CV (10)	3D CT HO (80-20)	3D CT CV (10)
Decision Tree	0.667	0.712	0.667	0.699	0.708	0.714
Naïve Bayes	0.708	0.808	0.708	0.562	0.625	0.662
Gradient Boosted Tree	0.792	0.767	0.750	0.671	0.708	0.649
Random Forest	0.750	0.822	0.667	0.685	0.667	0.623
Multi-Layer Perceptron	0.625	0.836	0.708	0.726	0.750	0.727
LibLINEAR	0.750	0.781	0.667	0.630	0.583	0.662
ADABoostM1	0.792	0.781	0.708	0.630	0.667	0.675
K-Nearest Neighbor 3	0.708	0.740	0.708	0.767	0.625	0.649
K-Nearest Neighbor 5	0.667	0.740	0.667	0.753	0.708	0.688
K-Nearest Neighbor 7	0.667	0.781	0.708	0.753	0.750	0.714

show the results of adopted classification models trained on the complete dataset for the outcome prediction, respectively, with the standard approach and with the 4-bagging ensemble approach, while

Table 5. Accuracy scores of trained models for outcome prediction on Cotugno dataset with the standard approach. Best results for each data type and split approach are in bold.

Model (Outcome, Cotugno, Standard)	Tabular HO (80-20)	Tabular CV (10)	2D CT HO (80-20)	2D CT CV (10)	3D CT HO (80-20)	3D CT CV (10)
Decision Tree	0.923	0.976	0.538	0.619	0.615	0.535
Naïve Bayes	0.923	0.786	0.538	0.643	0.538	0.419
Gradient Boosted Tree	1.000	0.976	0.538	0.667	0.538	0.558
Random Forest	0.923	0.929	0.692	0.643	0.692	0.581
Multi-Layer Perceptron	0.923	0.690	0.538	0.619	0.692	0.674
LibLINEAR	0.846	0.786	0.462	0.548	0.692	0.605
ADABoostM1	1.000	0.976	0.538	0.524	0.692	0.581
K-Nearest Neighbor 3	0.846	0.667	0.615	0.667	0.615	0.605
K-Nearest Neighbor 5	0.846	0.690	0.538	0.571	0.769	0.628
K-Nearest Neighbor 7	0.846	0.690	0.538	0.595	0.692	0.651

and

Table 6. Accuracy scores of trained models for outcome prediction on Cotugno dataset with the 4-bagging ensemble approach. Best results for each data type and split approach are in bold.

Model (Outcome, Cotugno, Bagging)	Tabular HO (80-20)	Tabular CV (10)	2D CT HO (80-20)	2D CT CV (10)	3D CT HO (80-20)	3D CT CV (10)
Decision Tree	0.929	0.905	0.714	0.619	0.533	0.581
Naïve Bayes	0.786	0.667	0.500	0.476	0.533	0.581
Gradient Boosted Tree	0.857	0.857	0.500	0.548	0.533	0.605
Random Forest	0.643	0.667	0.500	0.476	0.467	0.581
Multi-Layer Perceptron	0.714	0.786	0.571	0.619	0.333	0.488
LibLINEAR	0.786	0.667	0.571	0.548	0.667	0.535
ADABoostM1	0.857	0.857	0.429	0.500	0.600	0.581
K-Nearest Neighbor 3	0.571	0.643	0.500	0.405	0.600	0.558
K-Nearest Neighbor 5	0.714	0.690	0.643	0.548	0.667	0.535
K-Nearest Neighbor 7	0.643	0.643	0.643	0.500	0.467	0.581

show the results of models trained on the Cotugno dataset for outcome prediction with the same approaches. There is no result table for the Vanvitelli dataset since, in this dataset, all the patients have an outcome value of 0, meaning that every patient has been dismissed, and consequently a model cannot be trained to detect different outcomes.

As noticeable in these tables, for the Cotugno dataset there are some configurations where the models appear to be almost perfect or perfect, while once the datasets (Vanvitelli and Cotugno) are merged, performances slightly drop. This is expected since new data have only one type of class for the outcome output.

5.2. Days Range Classification Task

On the other hand, the days range prediction task appears to be harder to solve, since it is ascribable to a regression task. Indeed, the results reflect this complexity.

Table 7. Accuracy scores of trained models for days range prediction on complete dataset with the standard approach. Best results for each data type and split approach are in bold.

Model (Days Range, Complete, Standard)	Tabular HO (80-20)	Tabular CV (10)	2D CT HO (80-20)	2D CT CV (10)	3D CT HO (80-20)	3D CT CV (10)
Decision Tree	0.182	0.301	0.364	0.233	0.167	0.143
Naïve Bayes	0.227	0.219	0.091	0.288	0.250	0.169
Gradient Boosted Tree	0.136	0.205	0.409	0.288	0.250	0.182
Random Forest	0.182	0.205	0.409	0.233	0.125	0.104
Multi-Layer Perceptron	0.136	0.151	0.364	0.288	0.250	0.117
LibLINEAR	0.318	0.233	0.318	0.247	0.250	0.130
ADABoostM1	0.273	0.151	0.045	0.123	0.083	0.039
K-Nearest Neighbor 3	0.182	0.164	0.273	0.288	0.083	0.182
K-Nearest Neighbor 5	0.182	0.123	0.273	0.233	0.125	0.130
K-Nearest Neighbor 7	0.227	0.164	0.182	0.192	0.083	0.117

and

Table 8. Accuracy scores of trained models for days range prediction on complete dataset with the 4-bagging ensemble approach. Best models for each data type and split approach are in bold.

Model (Days Range, Complete, Bagging)	Tabular HO (80-20)	Tabular CV (10)	2D CT HO (80-20)	2D CT CV (10)	3D CT HO (80-20)	3D CT CV (10)
Decision Tree	0.292	0.205	0.792	0.781	0.708	0.870
Naïve Bayes	0.125	0.151	0.625	0.507	0.375	0.364
Gradient Boosted Tree	0.208	0.288	0.583	0.753	0.583	0.675
Random Forest	0.125	0.137	0.375	0.384	0.375	0.364
Multi-Layer Perceptron	0.208	0.123	0.750	0.658	0.375	0.455
LibLINEAR	0.208	0.205	0.500	0.507	0.250	0.286
ADABoostM1	0.083	0.082	0.500	0.315	0.500	0.403
K-Nearest Neighbor 3	0.292	0.137	0.250	0.342	0.250	0.260
K-Nearest Neighbor 5	0.292	0.178	0.375	0.301	0.125	0.234
K-Nearest Neighbor 7	0.167	0.164	0.292	0.329	0.208	0.247

show the results of the same models analyzed in the previous section, trained on the complete dataset.

Table 9. Accuracy scores of trained models for days range prediction on Vanvitelli dataset with the standard approach. Best results for each data type and split approach are in bold.

Model (Days Range, Vanvitelli, Standard)	Tabular HO (80-20)	Tabular CV (10)	2D CT HO (80-20)	2D CT CV (10)	3D CT HO (80-20)	3D CT CV (10)
Decision Tree	0.444	0.333	0.444	0.300	0.400	0.212
Naïve Bayes	0.000	0.267	0.444	0.267	0.300	0.273
Gradient Boosted Tree	0.333	0.267	0.444	0.267	0.400	0.273
Random Forest	0.222	0.333	0.333	0.200	0.500	0.242
Multi-Layer Perceptron	0.222	0.133	0.222	0.233	0.300	0.273
LibLINEAR	0.222	0.233	0.111	0.367	0.200	0.273
ADABoostM1	0.222	0.100	0.111	0.100	0.000	0.030
K-Nearest Neighbor 3	0.111	0.233	0.333	0.233	0.200	0.273
K-Nearest Neighbor 5	0.222	0.133	0.222	0.133	0.000	0.152
K-Nearest Neighbor 7	0.222	0.133	0.111	0.267	0.000	0.091

and

Table 10. Accuracy scores of trained models for days range prediction on Vanvitelli dataset with the 4-bagging ensemble approach. Best results for each data type and split approach are in bold.

Model (Days Range, Vanvitelli, Bagging)	Tabular HO (80-20)	Tabular CV (10)	2D CT HO (80-20)	2D CT CV (10)	3D CT HO (80-20)	3D CT CV (10)
Decision Tree	0.417	0.300	0.583	0.667	0.500	0.576
Naïve Bayes	0.167	0.133	0.583	0.367	0.250	0.303
Gradient Boosted Tree	0.250	0.100	0.500	0.467	0.250	0.394
Random Forest	0.250	0.100	0.500	0.367	0.250	0.303
Multi-Layer Perceptron	0.250	0.100	0.500	0.400	0.333	0.394
LibLINEAR	0.333	0.267	0.250	0.200	0.500	0.455
ADABoostM1	0.333	0.233	0.417	0.500	0.333	0.394
K-Nearest Neighbor 3	0.500	0.200	0.417	0.200	0.167	0.333
K-Nearest Neighbor 5	0.333	0.133	0.333	0.367	0.250	0.273
K-Nearest Neighbor 7	0.333	0.133	0.333	0.133	0.250	0.364

show the results with models trained with the Vanvitelli dataset.

Table 11. Accuracy scores of trained models for days range prediction on Cotugno dataset with the standard approach. Best results for each data type and split approach are in bold.

Model (Days Range, Cotugno, Standard)	Tabular HO (80-20)	Tabular CV (10)	2D CT HO (80-20)	2D CT CV (10)	3D CT HO (80-20)	3D CT CV (10)
Decision Tree	0.385	0.190	0.385	0.167	0.231	0.186
Naïve Bayes	0.077	0.119	0.308	0.214	0.231	0.209
Gradient Boosted Tree	0.385	0.143	0.154	0.333	0.231	0.209
Random Forest	0.308	0.238	0.385	0.214	0.154	0.116
Multi-Layer Perceptron	0.077	0.095	0.308	0.286	0.231	0.209
LibLINEAR	0.154	0.143	0.308	0.262	0.154	0.116
ADABoostM1	0.077	0.024	0.154	0.119	0.077	0.093
K-Nearest Neighbor 3	0.000	0.048	0.231	0.214	0.154	0.256
K-Nearest Neighbor 5	0.231	0.143	0.231	0.214	0.154	0.233
K-Nearest Neighbor 7	0.231	0.095	0.308	0.143	0.077	0.093

and

Table 12. Accuracy scores of trained models for days range prediction on Cotugno dataset with the 4-bagging ensemble approach. Best results for each data type and split approach are in bold.

Model (Days Range, Cotugno, Bagging)	Tabular HO (80-20)	Tabular CV (10)	2D CT HO (80-20)	2D CT CV (10)	3D CT HO (80-20)	3D CT CV (10)
Decision Tree	0.286	0.286	0.571	0.571	0.467	0.512
Naïve Bayes	0.429	0.262	0.500	0.286	0.333	0.465
Gradient Boosted Tree	0.571	0.381	0.286	0.476	0.333	0.372
Random Forest	0.286	0.310	0.286	0.333	0.333	0.349
Multi-Layer Perceptron	0.286	0.214	0.286	0.500	0.400	0.465
LibLINEAR	0.214	0.190	0.214	0.262	0.333	0.326
ADABoostM1	0.286	0.214	0.357	0.286	0.333	0.442
K-Nearest Neighbor 3	0.357	0.310	0.286	0.357	0.467	0.419
K-Nearest Neighbor 5	0.357	0.357	0.357	0.429	0.267	0.349
K-Nearest Neighbor 7	0.357	0.405	0.357	0.310	0.267	0.233

show the results with models trained with the Cotugno dataset.

In this case, accuracy scores of the whole workflow have a very low average value, but with few cases where specific models perform way better. Moreover, in any case, the 4-bagging ensemble approach improves the performances of the selected models.

5.3. Length of Stay Regression Task

Once completed the experiments for the classification tasks, some updates on the workflow were implemented for the hospitalization days prediction task. In particular:

• Models were replaced with regression ones;
• SMOTE algorithm with balancing on range feature was added;
• Feature selection algorithms with utility models were added.

Given these differences, the test to be performed expanded significantly. For length reasons, we do not report all the combinations of features selection strategies and used models, but only the best models for each feature selection model and approach combination. Since there are two adopted models (Linear Regressor and M5P) and two approaches (Backward Feature Elimination and Forward Feature Selection), for each data type and for each dataset source, there are four best models.

In order to make proper evaluations of system performances on LOS prediction, the mean absolute error (MAE) will be adopted [31]. In particular, this evaluation metric can be obtained by making the overall sum of differences between predicted values (in this case LOS) and real values divided by the number of occurrences. The formula to obtain this metric is

$$MAE={\displaystyle \frac{\sum |y_i-{\overline{y}}_i|}{n}}$$

(1)

where $y_i$ is the prediction value and ${\overline{y}}_i$ is the real value.

The following tables contain performances, in terms of MAE, of models trained with the feature selection approach, with a comparison with the MAE obtained with the standard approach. In particular, these tables contain the attempts made with the complete dataset (

Table 13. MAE scores of trained models for hospitalization days prediction on complete dataset with feature selection approaches. The best combination is the row in bold.

Split Approach	Features Type	FS Model	FS Approach	Selected Features	Best Model with FS (MAE)	Best Model w/o FS (MAE)
HO (80-20)	Tabular	Linear Reg.	BFE	Ischemic heart dis., COPD, COVID therapy	Reg. Tree (7.528)	SMO (20.357)
CV (10)				COPD, Age	SMO (18.648)	SMO (20.373)
HO (80-20)	Tabular	Linear Reg.	FFS	Gender, Ischemic heart dis., COPD, COVID therapy	Linear Reg. (8.073)	SMO (20.357)
CV (10)				COPD, Age	SMO (18.229)	SMO (20.373)
HO (80-20)	Tabular	M5P	BFE	Ischemic heart dis., COVID therapy	M5P (7.576)	SMO (20.357)
CV (10)				COPD, White blood cells, Gender	SMO (18.567)	SMO (20.373)
HO (80-20)	Tabular	M5P	FFS	COPD, COVID therapy, Ventilation	KNN 7 (7.674)	SMO (20.357)
CV (10)				Neutrophils, Gender	SMO (17.717)	SMO (20.373)
HO (80-20)	2D CT	Linear Reg.	BFE	6 features	Linear Reg. (3.102)	SMO (4.078)
CV (10)				7 features	SMO (5.067)	SMO (5.11)
HO (80-20)	2D CT	Linear Reg.	FFS	9 features	Linear Reg. (3.336)	SMO (4.078)
CV (10)				14 features	SMO (5.106)	SMO (5.11)
HO (80-20)	2D CT	M5P	BFE	6 features	SMO (4.025)	SMO (4.078)
CV (10)				3 features	SMO (5.111)	SMO (5.11)
HO (80-20)	2D CT	M5P	FFS	1 feature	SMO (4.0)	SMO (4.078)
CV (10)				1 feature	SMO (5.001)	SMO (5.11)
HO (80-20)	3D CT	Linear Reg.	BFE	4 features	SMO (12.602)	SMO (12.686)
CV (10)				4 features	SMO (16.796)	SMO (16.898)
HO (80-20)	3D CT	Linear Reg.	FFS	8 features	SMO (12.712)	SMO (12.686)
CV (10)				8 features	SMO (16.6)	SMO (16.898)
HO (80-20)	3D CT	M5P	BFE	3 features	SMO (12.516)	SMO (12.686)
CV (10)				3 features	SMO (17.599)	SMO (16.898)
HO (80-20)	3D CT	M5P	FFS	3 features	SMO (12.472)	SMO (12.686)
CV (10)				3 features	MLP (9.526)	SMO (16.898)

), the Vanvitelli dataset (

Table 14. MAE scores of trained models for hospitalization days prediction on Vanvitelli dataset with feature selection approaches. The best combination is the row in bold.

Split Approach	Features Type	FS Model	FS Approach	Selected Features	Best Model with FS (MAE)	Best Model w/o FS (MAE)
HO (80-20)	Tabular	Linear Reg.	BFE	Sex, Vent., Dementia, Lymph., GOT or GPT, Neutrophils	Linear Reg. (5.319)	GBTree (36.856)
CV (10)				Fever, Cough, Dementia, COPD, Active cancer last 5 years, Chronic liver disease, Neutrophils, Lymphocytes, Ageusia/Anosmia, Type 2 diabetes, p/F	MLP (27.485)	GBTree (46.564)
HO (80-20)	Tabular	Linear Reg.	FFS	Ventilation, Age, Fever, Dyspnea, Fibrinogen, LUS score, Neutrophils	Linear Reg. (4.334)	GBTree (36.856)
CV (10)				21 features	M5P (40.781)	GBTree (46.564)
HO (80-20)	Tabular	M5P	BFE	HB, Platelets, Fibrinogen, LUS score, Cough, Art. hypert.	Reg. Tree (6.833)	GBTree (36.856)
CV (10)				Cough, Dyspnea, Dementia, GOT or GPT, Blood glucose, INR, Ventilation	M5P (31.259)	GBTree (46.564)
HO (80-20)	Tabular	M5P	FFS	COPD, Atrial fibril., HB, Sodium, LUS score, Neutrophils	Reg. Tree (6.667)	GBTree (36.856)
CV (10)				Fever, Cough, Lymphocytes, Blood glucose, Fibrinogen, Ventilation	M5P (26.589)	GBTree (46.564)
HO (80-20)	2D CT	Linear Reg.	BFE	6 features	Linear Reg. (4.866)	MLP (6.639)
CV (10)				6 features	MLP (4.511)	MLP (3.731)
HO (80-20)	2D CT	Linear Reg.	FFS	2 features	M5P (7.281)	MLP (6.639)
CV (10)				2 features	MLP (4.998)	MLP (3.731)
HO (80-20)	2D CT	M5P	BFE	1 feature	GBTree (6.09)	MLP (6.639)
CV (10)				1 feature	MLP (1.43)	MLP (3.731)
HO (80-20)	2D CT	M5P	FFS	1 feature	GBTree (6.09)	MLP (6.639)
CV (10)				1 feature	MLP (4.148)	MLP (3.731)
HO (80-20)	3D CT	Linear Reg.	BFE	5 features	Linear Reg. (28.287)	KNN3 (36.09)
CV (10)				5 features	SMO (23.993)	SMO (28.749)
HO (80-20)	3D CT	Linear Reg.	FFS	6 features	Linear Reg. (34.16)	KNN3 (36.09)
CV (10)				6 features	MLP (14.611)	SMO (28.749)
HO (80-20)	3D CT	M5P	BFE	4 features	GBTree (34.113)	KNN3 (36.09)
CV (10)				4 features	MLP (18.1)	SMO (28.749)
HO (80-20)	3D CT	M5P	FFS	1 feature	GBTree (33.734)	KNN3 (36.09)
CV (10)				1 feature	SMO (26.879)	SMO (28.749)

), and the Cotugno dataset (

Table 15. MAE scores of trained models for hospitalization days prediction on Cotugno dataset with feature selection approaches. The best combination is the row in bold.

Split Approach	Features Type	FS Model	FS Approach	Selected Features	Best Model with FS (MAE)	Best Model w/o FS (MAE)
HO (80-20)	Tabular	Linear Reg.	BFE	Ischemic heart dis., Age, TC score, PF, Compl., MB, ICU, I dose, NEFRO, NEU, LINF, % (#1), n/l	Linear Reg. (3.781)	Simple Reg. (6.951)
CV (10)				Age, Gender, Score TC, Compl. type, TEP, ICU, CS, III dose, CARDIO, THYROID, ictus	Linear Reg. (8.236)	GBTree (8.421)
HO (80-20)	Tabular	Linear Reg.	FFS	TC score, CS, IL6, %, LINF, % (#1), Neutr.	Linear Reg. (2.25)	Simple Reg. (6.951)
CV (10)				Age, Compl. type, TEP, CARDIO, PNEUMO, LINF, % (#1), Ischemic heart dis.	MLP (7.779)	GBTree (8.421)
HO (80-20)	Tabular	M5P	BFE	Age, TC score, Compl., Max vent., CS, I dose, II dose, DM, IL6, NEU, % (#1), VES, DD	KNN 5 (5.225)	Simple Reg. (6.951)
CV (10)				Gender, PF, Compl. type, Max ventilation, PENTAGLOBIN, CARDIO, IL6, %, VES, PCR, COPD, White blood cells, Neutrophils	MLP (7.307)	GBTree (8.421)
HO (80-20)	Tabular	M5P	FFS	COVID therapy, CS, IL6, WBC, %, % (#1), n/l, ictus	Linear Reg. (3.772)	Simple Reg. (6.951)
CV (10)				Age, TEP, PNX, PENTAGLOBIN, PMD, CS, Cardio, PNEUMO, THYROID, NEU, VES, COPD, White blood cells, Neutrophils	Simple Reg. (10.274)	GBTree (8.421)
HO (80-20)	2D CT	Linear Reg.	BFE	1 feature	Linear Reg. (2.659)	Simple Reg. (2.636)
CV (10)				1 feature	MLP (4.701)	MLP (4.633)
HO (80-20)	2D CT	Linear Reg.	FFS	3 features	Linear Reg. (2.455)	Simple Reg. (2.636)
CV (10)				3 features	Linear Reg. (4.845)	MLP (4.633)
HO (80-20)	2D CT	M5P	BFE	1 feature	M5P (2.556)	Simple Reg. (2.636)
CV (10)				1 feature	SMO (4.896)	MLP (4.633)
HO (80-20)	2D CT	M5P	FFS	1 feature	M5P (2.556)	Simple Reg. (2.636)
CV (10)				1 feature	MLP (4.26)	MLP (4.633)
HO (80-20)	3D CT	Linear Reg.	BFE	4 features	Linear Reg. (5.997)	SMO (7.666)
CV (10)				4 features	GBTree (10.479)	GBTree (11.555)
HO (80-20)	3D CT	Linear Reg.	FFS	6 features	GBTree (6.078)	SMO (7.666)
CV (10)				6 features	SMO (11.463)	GBTree (11.555)
HO (80-20)	3D CT	M5P	BFE	5 features	M5P (7.391)	SMO (7.666)
CV (10)				3 features	MLP (6.069)	GBTree (11.555)
HO (80-20)	3D CT	M5P	FFS	4 features	GBTree (5.621)	SMO (7.666)
CV (10)				3 features	MLP (2.947)	GBTree (11.555)

).

As expected, the feature selection approach significantly improves the performances obtained with the standard approach, even with less data from small datasets and simple models. Considering all the feature selection approaches, the most recurrent tabular features that appear to be significant are the COPD, COVID therapy, the ventilation, the LUS score, the TC score and the LINF. Thus, in general, these should be collected and taken into account with priority when new LOS predictions will be performed.

6. Ethical Considerations

While the introduction of AI technologies in the national healthcare system can be a valuable operational ally, it is also necessary to consider the possible risks associated with the use of these tools in a high-risk infrastructure context. Indeed, the use of sensitive data, such as those used in this study, enables the implementation of useful technologies for sensitive infrastructures like healthcare, but their use requires important ethical considerations regarding the nature of the data themselves.

First and foremost, to ensure an ethical and responsible use of these technologies, it is essential to respect patients’ privacy and rights. This implies that, whenever data are collected or reused, the process should be conducted transparently and in compliance with current regulations. Even in the case of retrospective data, such as those used in this study, it is crucial to implement measures that protect data security and confidentiality, in order to maintain public trust in healthcare systems and the prudent use of AI. Secondly, such a model should preferably be designed for an emergency context rather than as a preventive measure. While its use could certainly be beneficial even in non-risk situations (primarily in terms of healthcare service speed), adopting such a system on a daily basis presents potential issues. For instance, the collection and retention of sensitive data pose a greater risk in terms of cybersecurity, for the reasons mentioned above [32]. Even if the protection of these data were guaranteed through advanced security measures to prevent unauthorized access and potential risks, the accumulation of sensitive data still creates a “honeypot” for potential cyber violations, thereby increasing the overall vulnerability of the healthcare system. Although this represents an increased risk rather than a new issue, this consideration is important when evaluating the large-scale implementation of such technologies.

Furthermore, the constant use of such systems might be hindered by habitual reliance on these models, potentially leading to a lack of responsibility among medical personnel towards individual patients [33]. As a retrospective study, this cannot happen in our analysis. Nonetheless, this risk must be taken into account for future studies. Indeed, over-reliance on machine predictions could result in medical personnel treating patients considered by the machine to be at lower risk with less precision, relying on algorithms rather than comprehensive, ongoing medical evaluations. The risk of stigmatization and approximate medical decisions, inherent in any application of the model, must be mitigated by maintaining human supervision and integrating the system’s predictions with clinical judgment. Therefore, such a tool should be used exclusively for bed management and not as a measure of patient severity.

For these reasons, regardless of the use case, it is strongly recommended that healthcare facilities adopting these technologies also establish an ethical committee. This committee, preferably multidisciplinary in nature regarding the involved parties, should be able to manage any ethical and practical issues arising from the use of such tools, basing its decisions also on the establishment of predefined protocols and knowledge of relevant regulations (e.g., ISO/IEC, IEEE, etc.). The ethical committee should monitor the use of the system, evaluate the ethical implications of the decisions made, and ensure that the adoption of these technologies occurs responsibly and respectfully of patients’ rights, considering individual cases.

In addition to the previous considerations, it is also crucial to point out the risk of potential bias in an artificial intelligence model. In particular, in the present case, it is important to emphasize that possible biases could arise if the model is trained on datasets that are limited in size or unrepresentative. Indeed, if the data used for training do not adequately reflect the diversity of the patient population, there is a risk that the model will produce biased or inaccurate predictions, with potentially negative consequences for certain groups of patients. For example, if the model is trained only on data from patients of the same age, gender, or ethnicity, it may not be able to generalize correctly to patients with different characteristics. In addition to an imperfect model, this could lead to a reduction in the quality of service for certain categories of patients or, in extreme cases, perpetuate existing inequalities in access to care and health outcomes, with consequences that would make the system both ethically problematic and unhelpful. While the risk of bias in a model cannot be completely eliminated, it can certainly be reduced. To minimize this risk, it seems essential to ensure that the datasets used are as large and representative as possible and that validation studies are carried out in different populations. In order to have greater control over performance, it may also be useful to provide for the possibility of periodically reviewing the model and improving it by incorporating new data to correct any biases that may emerge over time.

7. Discussion and Conclusions

The aim of this work was to identify the best approach, starting from a multimodal dataset collected thanks to the effort of the Hospital of University of Campania Luigi Vanvitelli and of the Cotugno Hospital, for the prediction of pneumonia patients’ hospital length of stay, so that healthcare infrastructures can schedule and manage in advance beds and doctors’ availability and prevent emergencies. In particular, the task was splitted into three subtasks, which are the outcome prediction, the LOS range prediction, and the LOS prediction (in order of increasing complexity).

The results obtained from the three proposed subtasks highlight their varying levels of difficulty and underscore the importance of feature filtering and selection. While the outcomes for the outcome prediction and days range prediction tasks are consistent with those reported in the literature, a significant improvement was achieved in the Length of Stay (LOS) prediction task. This improvement can be attributed to the specific considerations made in this study, despite the challenges posed by the limited number of observations—especially when the datasets were analyzed separately—and the reduced number of features in the merged dataset. To summarize, the best results for each task across the three datasets (Complete, Vanvitelli, Cotugno) are as follows:

• The best outcome prediction accuracy is of 0.875 on the Complete dataset by adopting the standard approach and 3D features, not available on the Vanvitelli dataset since all the patients have the same outcome, and 1.0 on the Cotugno dataset by adopting the standard approach and clinical data.
• The best days range prediction accuracy is of 0.870 on the Complete dataset by adopting the bagging ensemble approach and 3D features, of 0.667 on the Vanvitelli dataset by adopting the bagging ensemble approach and 2D features, and of 0.571 on the Cotugno dataset by adopting the bagging ensemble approach and 2D features.
• The best length of stay prediction MAE is 3.102 on the Complete dataset by adopting the feature selection approach and 2D features, 1.43 on the Vanvitelli dataset by adopting the feature qelection approach and 2D features, and 2.25 on the Cotugno dataset by adopting the feature selection approach and clinical data.

It is remarkable to notice several different aspects. First of all, the outcome prediction task is easy enough to adopt the standard approach and, if other methods are adopted, the system will overfit. Unlike the outcome task, the days range prediction task performs better with an ensemble approach, meaning that it is difficult enough not to make the system overfit during the training process. On the other hand, the regression task relies on feature selection, providing results which are way better than the ones obtained with the simplified version of the task; that is, the days range prediction, considering that the average error is of only 3 days. Moreover, even though, for the regression task, the best results are obtained with a selection of tabular features, the most common features adopted by the most performing systems are the 2D scan features retrieved from the ResNet, meaning that the convolutional neural network (CNN) was capable of extracting some extremely useful information from the images. An interesting aspect is that, when 2D features are adopted, models require only few features (sometimes just one) to perform at their best.

Further experiments were performed in order to understand whether the system behaves better with different fold values for cross-validation (2, 5, 20). However, experiments performed with $k=10$ revealed to be the best among the other attempts. Moreover, despite the fact that there is a single case wherein cross-validation performed better than any other with hold-out (Vanvitelli dataset, CT scan features), where the MAE reached a value of 1.43, when the Complete dataset was adopted, it is still preferable to adopt a hold-out approach, since the best cross-validation model is about two days worse than the best hold-out model.

Finally, it is worth mentioning that the feature selection process, despite selecting different features based on the task and on the considered dataset, highlighted that some features (the COPD, COVID therapy, ventilation, the LUS score, the TC score, and the LINF) tend to be prominent in the evaluation of LOS. This is interesting, as it suggests that these metrics should always be considered and acquired during the early stages.

After evaluating the system’s performance on the available datasets, a comparison with existing methods can be conducted to assess whether the proposed approach provides a significant contribution to the research on LOS prediction for hospital patients. To this aim,

Table 16. Comparison table containing some aspects of the proposed work and of three relevant works in the literature for hospitalization prediction tasks.

Work	Prediction Tasks	Disease	Data Type	Multiple Datasets	Missing Values Handling	Dataset Balancing
Proposed work	Outcome (Classification) Days Range (Classification) LOS (Regression)	Pneumonia COVID-19	Clinical data CT scans CT volumes	Yes	Yes	SMOTE
Saadatmand et al. [10]	Outcome (Classification) Days Range (Classification)	COVID-19	Clinical data	Yes	Yes	ROSE
Khajehali  et al. [11]	Days Range (Classification)	Pneumonia	Clinical data	No	Yes	No
Mekhaldi  et al. [4]	LOS (Regression)	Generic	Clinical data	No	No	SMOTE

shows similarities and differences between the proposed solution and some other significant work in the literature [4,10,11].

We follow the details provided by the authors in [4,10,11] to implement the state-of-the-art solutions, and we report in

Table 17. Performances of best models of compared works in all their available tasks with the Complete dataset. The best results for each task are in bold.

Task	Model	Data Type	Metric	Value
Outcome (Classification)	Gradient Boosted Tree (Our)	CT volumes	Accuracy	0.875
	XGBoost Tree Ensemble [10]	Clinical data	Accuracy	0.765
Days Range (Classification)	4-Bagging Ensemble of Decision Trees (Our)	CT volumes	Accuracy	0.870
	Bagged CART [10]	Clinical data	Accuracy	0.294
	Voting Classifier [11]	Clinical data	Accuracy	0.235
LOS (Regression)	BFE + Linear Regression (Our)	CT scans	MAE	3.102
	Multiple Linear Regression [4]	Clinical data	MAE	24.45

the results of the comparison in 10-fold CV with the workflow described in our work.

All the studies from the literature exhibit an average performance decrease of 0.3 in accuracy for classification tasks and 21 days for the regression task. This can primarily be attributed to two factors, namely the absence of key pre-processing steps, such as handling missing values or balancing the dataset, and the use of clinical data alone, rather than data extracted from CT scans and volumes.

Indeed, these values are compatible with the ones obtained with the proposed work itself, but with the adoption of standard clinical data.

Starting from these observations, the consequent future works would focus on the expansion of the dataset, trying to collect from the new ones’ features which are already present in the available datasets, on the adoption of other CNNs for features extraction from images, with consequent XAI analysis to understand where the different models focus more, and a study of the best features from clinical tests which doctors and researchers should provide to retrieve the best results possible.

This first result, obtained with the precious collaboration with the University of Campania Luigi Vanvitelli and the Cotugno Hospital, provides an extremely useful tool in hospital structures management, both for emergencies and for the daily routine, and a clear vision of how powerful machine learning models, if careful considerations and precise analysis on data are made, can be and how impactful they are in the biomedical context, even with a limited amount of data.