A comparative study of neural network architectures for vital signs monitoring based on the national early warning systems (NEWS)

Abstract

Objective: The study aims to assess the efficacy of various neural network architectures in predicting the National Early Warning Systems (NEWS) score, using vital signs, to enhance early warning and monitoring in clinical settings. Methods: A comparative evaluation of 29 neural network architectures, including Discriminant Analysis, Support Vector Machines, Logistic Regression, Decision Trees, Neural Networks, and Ensemble methods, was performed. These architectures were assessed based on accuracy, sensitivity, processing speed, model size, and execution time, using synthetically generated data representing 9000 clinical scenarios. Results: The analysis revealed that Linear Discriminant Analysis, narrow and medium Neural Networks, and specific Support Vector Machine (SVM) configurations, particularly Linear SVM, Quadratic SVM, and Coarse Gaussian SVM, achieved 100% accuracy and efficiency in predicting NEWS scores, making them suitable for real-time monitoring. Other architectures exhibited varying performance, with many failing to meet the required accuracy for clinical applications. Conclusion: The study identified Linear Discriminant Analysis and narrow and medium Neural Networks, along with Linear, Quadratic, and Coarse Gaussian SVMs, as optimal for integrating machine learning with NEWS, due to their precision, speed, and suitability for deployment in healthcare environments, particularly in Intensive Care Units.

Keywords

neural networks vital signs early warning systems

Introduction

Vital signs including heart rate (HR), body temperature (T), respiration rate (RR), blood pressure (BP), and blood oxygen saturation (SPO2) are important parameters that reflect the health condition and provide critical and valuable information necessary for the accurate assessment and diagnosis of patients.¹ Monitoring these parameters on a regular basis is essential, especially in Intensive Care Units (ICUs) where timely signs deterioration detection can improve the prompt and proper intervention. Many Early Warning Systems (EWS) are available to determine patients at risk based on their monitored vital signs. Typically, these systems give scores to each vital sign range, with a high score indicating a greater level of deterioration.²

EWS’s are very valuable for patient deterioration monitoring due to their simplicity and effectiveness in most medical settings.³ The most common EWS’s which incorporate vital signs are NEWS, the Modified Early Warning Score (MEWS), Hamilton Early Warning Score (HEWS), Acute Physiology Score (RAPS), Simple Early Warning Score (SEWS) and Rapid Emergency Medicine Score (REMS).^4,5 Out of the available EWS’s, numerous research publications have shown that NEWS is the most accurate, with superior sensitivity and specificity in predicting deterioration in emergency departments and various healthcare setting.^6,7

While NEWS has been shown to be useful in healthcare monitoring, its implementation without automation or machine learning (ML) can present significant disadvantages.^8,9 These disadvantages include, and not limited to, higher frequency of false alarms, increased human error, limitations in quickly analyzing large datasets, and a slower response time to emerging health situations.¹⁰ ML can automate and improve the accuracy and responsiveness of EWS like NEWS. ML models can generalize complex patterns and trends within vital signs data and accurately and quickly predict the NEWS score which leads to earlier and more precise classification of patient deterioration.

This study aims to compare the performance of different neural network (NN) architectures in predicting NEWS scores from five key vital signs. The aim is to identify the most effective approach for integrating ML with NEWS and enhancing its predictive capabilities, in terms of precision speed, and potential for deployment to hardware implementation. Precisely, the objective of this paper is to compare the available NN architectures and evaluate their performance in determining the NEWS score from the five vital signs. The performance will be evaluated in terms of accuracy, sensitivity, and complexity.

Literature review

The adequacy of EWS’s, particularly NEWS, has encouraged researchers to seek the standardization of patient monitoring across various clinical settings. The main goal was to establish advanced techniques that can overcome NEWS limitations caused by the threshold-based decision, causing problems of sensitivity and specificity. Various ML approaches were explored including Fuzzy Logic (FL) and NN.

An approach of using fuzzy logic to design and implement an equivalent warning system for patients’ status classification was compared to MEWS.¹¹ The approach uses a large number of fuzzy rules (1800 rules) that are manually generated. Despite the relative complexity of the system design and deployment, the authors claim that the results are acceptable despite the absence of evaluation metrics.

A study, based on fuzzy logic, compared 16 different algorithms and verified their performance with simulation on 12 datasets.¹² The evaluation performance indicates that some fuzzy approaches tend to have a large rule base, some others are less accurate than needed. However, a classification system using fuzzy logic and gene expression programming (GPR), was found to have good classification performance with a small number of explainable rules.

Of the ML approaches applied to EWS, NN received the greatest attention due to their generalization capabilities and their power in dealing with complex data sets. A brief guide to deep learning in healthcare summarizes various applications like computer vision, generalization, and reinforcement learning and it shows how these could be beneficial to important medical applications such as disease risk and genetic traits.¹³

The advantages and disadvantages of ML in predicting patients’ health deterioration in medical settings are discussed in a published systematic review.¹⁴ The review examined 29 research papers and found that the various ML models had an area under the curve (AUC) ranging from 55% to 99%. Many models are used to automate the health deterioration risk and there is still a need for further improvement, especially in real-world situations and in areas related to patient clinical deterioration.

Another scoping review,¹⁵ reached the same conclusion that ML-based EWS models are promising but still need further research to be successfully applied in clinical practice. The article presented many ML models including Kernel-based, tree-based, and regression-based for risk deterioration prediction with AUC ranging from 57% to 97%. A simple feedforward NN to predict the deterioration of five vital signs (HR, T, BP, RR, and SPO2) was able to achieve 95% precision.¹⁶

Many publications investigated the time series approach to vital signs monitoring using various ML algorithms. For example, a hybrid KNN-LS-SVM learning algorithm was used to predict future vital signs values with less than 5% mean absolute percentage error.¹⁷ A Recurrent NN (RNN) with three layers achieved an AUROC of 87%.¹⁸ Another time-series work using six statistical ML algorithms (Naive Bayes, Gradient Boosting, Decision Trees, Ensemble Methods, Logistic Regression, and Random Forest) achieved an AUC ranging from 84% to 96%.¹⁹ Similar approaches are applied for respiratory deterioration, where EWS’s need improvement.²⁰ The algorithm performance in predicting the 24 h ahead deterioration achieved an AUROC of 94% with a 70% accuracy. Other publications with similar NN architectures (logistic regression, Naive Bayes classifier, decision trees, support vector classifier, K-Neighbors Classifier, and gradient boosting classifier) achieved results ranging from 84% to 89% accuracy and AUROC ranging from 68% to 94%.^21–23 ML was also applied to EWS for specific objectives such as mortality predictions and cardiac arrest risks and showed a potential for fast identification of high-risk patients, with reasonable accuracy an capability of reducing daily alarm rates by over 20%.^24–27

The review of the previous work indicates good potential for applying ML with EWS. Many of the evaluated approaches provided reasonable accuracy. However, the study of the complexity of these approaches and their potential deployment is not thoroughly discussed. In addition, with the advancement of methodologies and tools, higher precision and low complexity are more possible now, which will be investigated in this paper.

Methodology

Data collection and processing

The NEWS model has five inputs representing the five vital signs (HR, BP, T, RR, and SPO2) and a single output representing the NEWS score, which is the aggregation of the scores of the vital signs. Vital signs are scored according to Table 1. The aggregate output will be a set of 16 classes ranging from 0 to 15 and is typically interpreted according to Table 2.²⁸

Table 1.

NEWS scoring system.

	Score
Physiological parameter	3	2	1	0	1	2	3
Respiration rate (per minute)	≤8		9–11	12–20		21-24	≥25
SpO2 scale 1 (%)	≤91	92–93	94–95	≥96
Systolic BP (mmHg)	≤90	91–100	101–110	111–219			≥220
Pulse (per minute)	≤40		41–50	51–90	91-110	111-130	≥131
Temperature (°C)	≤35.0		35.1–36.0	36.1–38.0	38.1-39.0	≥39.1

Table 2.

NEWS clinical risk level classification.

NEWS score	Clinical risk	Response
Aggregate score 0-4	Low	Ward-based response
Aggregate score 5-6	Medium	Key threshold for urgent response
Aggregate score 7 or more	High	Urgent or emergency response

Data points are synthetically generated to represent various situations. Synthetic data is preferred for the evaluation objective because real data tends to represent normal and close to normal situations with little or no representation of extreme clinical values. A set of 9000 cases is divided into training and testing sets (80% of the data for training and 20% for testing). It was ensured that the training and testing sets are well-balanced and that all the classes are equally represented in both.

NN architectures

All possible and relevant NN architectures included in the Classification Learner App of MATLAB’s ML Toolbox v. 2024 are used for evaluation. A total of 29 NN architectures representing nine different classifier types are used and they are presented in Table 3.

Table 3.

List of NN Architectures evaluated in this study.

Classifier type	Architecture	Parameters
Discriminant analysis	Linear discriminant analysis (LDA)	Gaussian distribution of classes equal covariance matrices
Discriminant analysis	Quadratic discriminant analysis (QDA	Different covariance matrices for each class
SVM	Linear SVM	Linear Kernel, auto SCLAE
	Fine Gaussian	Gaussian Kernel SCLAE = 0.56
	Coarse Gaussian	Gaussian Kernel SCLAE = 8.9
	Medium Gaussian	Gaussian Kernel SCLAE = 2.2
	Quadratic	Quadratic Kernel SCLAE = auto
	Cubic	Cubic Kernel SCLAE = auto
Logistic regression	Multinomial logistic regression	Auto, one-vs-one
Decision trees	Coarse tree	4 splits
	Medium tree	20 splits
	Fine tree	100 splits
NN	Narrow NN	1 layer with 10 neurons
	Medium NN	1 layer with 25 neurons
	Bilayered NN	10 × 10
	Trilayered NN	10 × 10 × 10
Ensemble methods	AdaBoost (boosted tree)	Maximum number of splits = 20, number of learners = 30
	Subspace K-Nearest Neighbors (SKNN)	Subspace dimension = 3, number of learners = 30
	Random subspace: Random Undersampling Boosting	Maximum number of splits = 20, number of learners = 30
	Bagged tree	Maximum number of splits = 7679, number of learners = 30
K-Nearest Neighbors (KNN)	Coarse KNN	Euclidean distance, equal weightl, 10 neighbors
	Medium KNN	Euclidean distance, equal weightl, 100 neighbors
	Fine KNN	Euclidean distance, equal weightl, 1 neighbor
	Weighted KNN	Euclidean distance, squared inverse weight, 10 neighbors
	Cosine KNN	Cosine distance, equal weightl, 10 neighbors
	Cubin KNN	Minkowski (cubic) distance equal weight, 10 neighbors
Naive Bayes	Kernel (Gaussian) Naive Bayes	Gaussian distribution
Kernel	SVM Kernel	Auto, one-vs-one
Kernel	Logistic regression Kernel	Auto, one-vs-one

Evaluation metrics

All the architectures will be evaluated using the following metrics:

• Testing Accuracy (%), knowing that the data used is balanced.

• Total Testing Cost, assuming equal penalties for all misclassifications during testing.

• Prediction Speed in number of observations per second (obs/sec), representing the number of data samples processed by the model per second during testing.

• Model Size in kB or MB, useful for deployment study.

• Precision.

• Recall.

• F1Score.

• Mean AUROC.

• Execution Time (using: i7-6500U, 2.50 GHz, 8.00 GB, 64-bit OS).

Results

After training and evaluating all the 29 architectures, three performance groups are identified and will be labelled as: low performers, average performers, and top performers.

Low-performing architectures

Based on the testing results, 11 architectures performed poorly, especially in terms of accuracy, and they are namely most of the ensemble classifiers (three out of four), all the decision trees classifiers, most of the KNN classifiers (four out of six), and the Naïve Bayes classifier. A tabulation of their performance is summarized in Table 4. A Sample confusion matrix and ROC plot of these results is given in Figures 1 and 2.

Table 4.

Testing Results for the low performing architectures.

Model type	Accuracy % (Test)	Total cost (Test)	Precision	Recall	F1 Score	Mean AUROC	Process. Time (sec)
Cubic_KNN	87.50	472	0.880	0.881	0.880	0.987	2.085
Medium_KNN	87.50	472	0.880	0.881	0.880	0.987	0.164
Cosine_KNN	86.65	504	0.875	0.876	0.874	0.990	0.521
Fine_Tree	64.78	1330	0.627	0.633	0.623	0.962	0.434
Coarse_KNN	47.85	1969	0.436	0.460	0.442	0.946	0.094
Boosted_Tree_Ensemble	41.08	2225	NaN	0.425	NaN	0.921	0.160
RUSBoosted_Tree_Ensemble	39.35	2290	NaN	0.403	NaN	0.897	0.060
Medium_Tree	39.35	2290	NaN	0.403	NaN	0.897	0.020
Kernel_Naive Bayes	32.57	2546	NaN	0.329	NaN	0.873	2.890
Subspace_KNN_Ensemble	29.32	2669	NaN	0.305	NaN	0.783	0.539
Coarse_Tree	23.76	2879	NaN	0.234	NaN	0.783	0.007

Figure 1.

Confusion Matrix for one of the low-performing architectures (Ensemble subspace KNN).

Figure 2.

ROC plot for one of the low-performing architectures (Coarse Tree).

Average-performing architectures

Here also 11 architectures are considered average performers based on their testing results. These architectures and their testing results are summarized in Table 5. A Sample confusion matrix and ROC plot of these results is given in Figures 3 and 4. Due the nature of the classification problem and its importance in clinical monitoring, and given the average complexity of the data, any architecture achieving excellent testing accuracy and excellent testing precision (very close to 100%) but not achieving exactly 100% is considered an average performer.

Table 5.

Testing Results for the average performing architectures.

Model type	Accuracy % (Test)	Total cost (Test)	Precision	Recall	F1 Score	Mean AUROC	Process. Time (sec)
Trilayered_Neural_Network	99.97	1	1.000	1.000	1.000	1.000	0.304
Bilayered_Neural_Network	99.97	1	0.999	0.999	0.999	1.000	0.015
Wide_Neural_Network	99.97	1	0.999	0.999	0.999	1.000	0.017
Cubic_SVM	99.92	3	1.000	1.000	1.000	1.000	0.933
Medium_Gaussian_SVM	99.66	13	0.998	0.998	0.998	1.000	0.236
Fine_KNN	97.30	102	0.971	0.971	0.971	0.985	0.038
Weighted_KNN	97.27	103	0.972	0.972	0.972	0.991	0.030
Fine_Gaussian_SVM	97.25	104	0.972	0.971	0.971	0.996	0.485
Ensemble_Bagged_Tree	97.09	110	0.971	0.971	0.970	0.991	0.122
SVM_Kernel	96.72	124	0.963	0.963	0.963	0.997	1.208
Logistic_Regression_Kernel	93.46	247	0.944	0.944	0.944	0.995	1.028

Figure 3.

Confusion Matrix for one of the average-performing architectures (Logistic Regression Kernel).

Figure 4.

ROC plot for one of the average-performing architectures (SVM Kernel).

Top-performing architectures

Out of the 29 tested architectures, seven performed exceptionally well and achieved 100% testing accuracy and testing precision, as shown in Table 6. A Sample confusion matrix and ROC plot of these results is given in Figures 5 and 6. The performance of these models will be presented in more details.

Table 6.

Testing Results for the top performing architectures.

Model type	Accuracy % (Test)	Precision	Recall	F1 Score	Mean AUROC	Procssing time (sec)
Efficient_Logistic_Regression	100	1	1	1	1	0.247
Linear_Discriminant	100	1	1	1	1	0.046
Medium_Neural_Network	100	1	1	1	1	0.024
Narrow_Neural_Network	100	1	1	1	1	0.025
Coarse_Gaussian_SVM	100	1	1	1	1	0.315
Quadratic_SVM	100	1	1	1	1	0.196
Linear_SVM	100	1	1	1	1	0.162

Figure 5.

Confusion Matrix for one of the top-performing architectures (Narrow NN).

Figure 6.

ROC plot for one of the top-performing architectures (Linear Discriminant).

A summary of the relative complexity of the top-performing architectures is shown in Table 7. The complexity can be observed from the prediction speed, the model size, and with less importance the training time.

Table 7.

Complexity summary of the top performing architectures.

Group label	Model type	Training time (sec)	Prediction Speed (obs/sec)	Model Size (kB)
Linear discriminant	Linear_Discriminant	10.847	∼140000	8
NN	Narrow_Neural_Network	146.11	∼140000	8
NN	Medium_Neural_Network	146.73	∼120000	10
SVM	Linear_SVM	43.351	∼13000	564
SVM	Quadratic_SVM	59.339	∼6500	784
Logistic regression	Efficient_Logistic_Regression	30.528	∼8900	1000
SVM	Coarse_Gaussian_SVM	118.16	∼2500	1000

Discussion

The comparative analysis of 29 neural network architectures for NEWS score prediction reveals a broad spectrum of performance. Most of the models—primarily those based on KNN, decision trees, and ensemble methods—showed insufficient accuracy and precision for reliable clinical applications, underscoring the critical importance of careful model selection in high-stakes medical contexts. While several models showed acceptable performance, only seven (Linear Discriminant Analysis, narrow and medium neural networks, and some SVM configurations) achieved perfect accuracy and precision. The superior performance and efficiency of Linear Discriminant Analysis and simpler neural networks make them particularly promising candidates for real-time deployment in clinical settings.

However, there is still a crucial need for explainable AI (XAI) in healthcare. High accuracy is essential, but the ability to understand why a model generates a specific prediction is paramount for building clinician trust and ensuring responsible AI use. The need for transparency and interpretable results, despite the challenges, is evident.²⁹ In high-stakes medical applications like early warning systems, understanding the model’s reasoning is essential for ensuring accurate and prompt clinical intervention. Therefore, future research should prioritize the development of XAI methods, such as SHAP or LIME, to enhance the interpretability and trustworthiness of these models. In conclusion, this study identifies promising architectures for NEWS score prediction. However, further research is needed to fully realize the potential of machine learning in this critical area.

Limitations

Few limitations of this study warrant consideration. The use of synthetic data, while allowing for important insights, may not fully capture the complexity and variability inherent in real-world clinical data, potentially limiting the generalizability of the findings. The study also did not explicitly assess model performance under conditions of noisy or incomplete data, a common occurrence in clinical practice. Furthermore, the integration of these models into existing clinical workflows and their impact on human-computer interaction requires further investigation for practical implementation. These factors need further research to improve model robustness and address practical challenges in clinical settings.

Conclusions

This comparative analysis reveals a significant disparity in performance among the 29 NN architectures investigated. While several architectures, notably those based on KNN, ensemble methods, and decision trees, showed inferior performance, a group of 11 showed acceptable accuracy, suggesting a possibility for potential real-world applications. However, only seven architectures achieved exceptional performance, reaching 100% accuracy and precision. These include Linear Discriminant, NN (narrow and medium), and SVM (Linear, Quadratic, and Coarse Gaussian), along with the Efficient Logistic Regression. Among these top performers, Linear Discriminant, narrow NN, and medium NN stand out due to their exceptional speed (over 120,000 observations per second), minimal model size (less than 10 kB), and real-time monitoring potential, making them ideal candidates for deployment in real-world clinical environments like ICUs. This research underscores the critical need for comprehensive analysis and thorough selection of proper NN architectures for improving the performance and feasibility of ML-based EWS in healthcare.

Footnotes

ORCID iD

Adel BenAbdennour

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Declaration of conflicting interests

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

References

Goldberg

. Vital signs. Practical Guide to Clinical Medicine . https://meded.ucsd.edu/clinicalmed/vital.html (accessed 1 September 2024).

Sapra

Malik

Bhandari

. Vital sign assessment. StatPearls Publishing LLC (internet), 2024. https://www.ncbi.nlm.nih.gov/books/NBK553213/

Downey

Tahir

Randell

, et al. Strengths and limitations of early warning scores: a systematic review and narrative synthesis. Int J Nurs Stud 2017; 76: 106–119.

Gerry

Bonnici

Birks

, et al. Early warning scores for detecting deterioration in adult hospital patients: systematic review and critical appraisal of methodology. BMJ 2020; 369: m1501.

Covino

Sandroni

Della Polla

, et al. Predicting ICU admission and death in the emergency department: a comparison of six early warning scores. Resuscitation. 2023; 190: 1–10.

Wei

Xiong

Wang

, et al. The accuracy of the National Early Warning Score 2 in predicting early death in prehospital and emergency department settings: a systematic review and meta-analysis. Ann Transl Med 2023; 11: 1–18.

Royal College of Physicians . National early warning score (NEWS) 2: standardising the assessment of acute-illness severity in the NHS. London: Updated report of a working party. RCP.

Paulson

Dummett

Green

, et al. What do we do after the pilot is done? Implementation of a hospital early warning system at scale. Jt Comm J Qual Patient Saf 2020; 46: 207–216.

Arora

Balasubramanian

Stranieri

, et al. Classification of methods to reduce clinical alarm signals for remote patient monitoring: a critical review. In Cloud Computing in Medical Imaging. Auerbach Publications, 2023, pp. 173–194.

10.

Ahmed

Spooner

Isherwood

, et al. A systematic review of the barriers to the implementation of artificial intelligence in healthcare. Cureus; 15. DOI: 10.7759/cureus.46454.

11.

Al-Dmour

Sagahyroon

Al-Ali

, et al. A fuzzy logic–based warning system for patients classification. Health Informatics J 2019; 25: 1004–1024.

12.

Czmil

. Comparative study of fuzzy rule-based classifiers for medical applications. Sensors 2023; 23: 1–20.

13.

Esteva

Robicquet

Ramsundar

, et al. A guide to deep learning in healthcare. Nat Med 2019; 25: 24–29.

14.

Jahandideh

Ozavci

Sahle

, et al. Evaluation of machine learning-based models for prediction of clinical deterioration: a systematic literature review. Int J Med Inf 2023; 175: 1–14.

15.

Muralitharan

Nelson

, et al. Machine learning–based early warning systems for clinical deterioration: systematic scoping review. J Med Internet Res 2021; 23: 1–22.

16.

Al-Shwaheen

Wen

. A new model for tracking and detection of deterioration of vital signs based on artificial neural network. J Theor Appl Inf Technol 2019; 97: 3809–3818.

17.

Amer

AYA

Wouters

Vranken

, et al. Vital signs prediction and early warning score calculation based on continuous monitoring of hospitalised patients using wearable technology. Sensors 2020; 20: 1–21.

18.

C-F

Chiu

S-I

Jang

J-SR

, et al. Improved inpatient deterioration detection in general wards by using time-series vital signs. Sci Rep 2022; 12: 11901–11911.

19.

Kobylarz Ribeiro

dos Santos

HDP

Barletta

, et al. A machine learning early warning system: multicenter validation in Brazilian hospitals. 2020 IEEE 33rd International Symposium on Computer-Based Medical Systems (CBMS), Rochester, MN, USA, 2020, pp. 321- 326, DOI: 10.1109/CBMS49503.2020.00067.

20.

Youssef

Kouchaki

Shamout

, et al. Development and validation of early warning score systems for COVID-19 patients. Healthc Technol Lett 2021; 8: 105–117.

21.

Aldhoayan

Aljubran

. Prediction of ICU patients’ deterioration using machine learning techniques. Cureus 2023; 15: 1–7.

22.

Rangan

Pathinarupothi

Anand

KJS

, et al. Performance effectiveness of vital parameter combinations for early warning of sepsis—an exhaustive study using machine learning. JAMIA Open 2022; 5: 1–11.

23.

Lashen

St John

Almallah

, et al. Machine learning models versus the national early warning score system for predicting deterioration: retrospective cohort study in the United Arab Emirates. JMIR AI 2023; 2: 1–16.

24.

Logothetis

Green

Holland

, et al. Predicting acute clinical deterioration with interpretable machine learning to support emergency care decision making. Sci Rep 2023; 13: 1–16.

25.

Lee

Cho

K-J

Kwon

, et al. A multicentre validation study of the deep learning-based early warning score for predicting in-hospital cardiac arrest in patients admitted to general wards. Resuscitation 2021; 163: 78–85.

26.

Baker

Xiang

Atkinson

. Continuous and automatic mortality risk prediction using vital signs in the intensive care unit: a hybrid neural network approach. Sci Rep 2020; 10: 1–12.

27.

Salehinejad

Meehan

Rahman

, et al. Novel machine learning model to improve performance of an early warning system in hospitalized patients: a retrospective multisite cross-validation study. eClinicalMedicine 2023; 66: 1–9.

28.

McMorran

Crowther

Henderson

, et al. National early warning score (NEWS). gpnotebook . https://gpnotebook.com/en-GB/pages/infectious-disease/national-early-warning-score-news (2024, accessed 1 September 2024).

29.

Singh

Merdivan

Kropf

, et al. Class imbalance in multi-resident activity recognition: an evaluative study on explainability of deep learning approaches. Univers Access Inf Soc. 2024; 1–19.