Continuous vital sign monitoring for predicting hospital length of stay: a feasibility study in chronic obstructive pulmonary disease and chronic heart failure patients

Abstract

Background

Vital signs monitoring provides clinicians with real-time information regarding patients’ current medical condition. We hypothesize that applying comprehensive analytical methods to underutilized, routinely collected vital signs data can yield valuable insights to support clinical decision-making. In this study, we present a novel approach for vital signs time series analysis applied to hospitalization length of stay (LOS) prediction in chronic obstructive pulmonary disease (COPD) and chronic heart failure (CHF) patients.

Methods

Heart rate (HR), respiratory rate (RR) and peripheral oxygen saturation (SpO₂) were continuously monitored during the first 24 h of hospital admission in COPD and CHF patients admitted to general, non-ICU hospital wards. The resulting time series were submitted to a comprehensive analysis through a highly comparative, massive feature extraction. We identified key patterns associated with hospitalization length of stay (LOS). Finally, we developed a predictive model for hospitalization LOS combining predictive features from the three vital signs time series.

Results

A total of 101 patients were enrolled in the study, 74 of whom were eligible for analysis (39 COPD and 35 CHF patients). Periodicity and self-correlation in HR and RR time series were associated to hospitalization LOS. In SpO₂ time series, short-term fluctuations and local dynamics were associated to hospitalization LOS. The predictive model for hospitalization LOS was built using nineteen predictive features and achieved an area under the curve (AUC) of 0.975, an accuracy of 0.944, a sensitivity of 0.979, and a specificity of 0.900 in 10-fold cross-validation.

Conclusion

Through a comprehensive feature-based analysis, we identified key patterns in HR, RR, and SpO₂ time series associated with hospitalization LOS in COPD and CHF patients and a compact set of features that can accurately predict LOS in COPD and CHF patients using only routinely collected data from the first 24 h of admission.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13049-025-01458-4.

Background

Vital signs have been routinely monitored in hospital settings for over a century [1], as they provide clinicians with real-time information of their patients’ current medical condition [2]. Abnormal vital signs are known to be indicative of clinical deterioration and often precede adverse events [3, 4], enabling timely medical intervention [5]. Consequently, vital signs monitoring is part of the usual practice in high-risk environments like intensive care units (ICUs) [6]. In these settings, vital sign monitoring has been utilized for patient risk stratification [7], forecasting vital signs trends [8, 9] and predicting outcomes like sepsis [10, 11] and mortality [12–14].

Although continuous monitoring improves vital signs control and patient management, its adoption in general wards has been traditionally hindered by technological limitations, substantial economic costs and negative implications in patient mobility and recovery [15, 16]. Nevertheless, the development of novel devices that overcome these challenges has allowed for transferring the tailored care of the ICUs to the general ward. In this regard, previous research has shown that continuous vital signs monitoring in general wards is not only feasible [17] but also improves patient outcomes [18]. Several studies have examined the relationship between continuous vital sign monitoring in general wards and various patient outcomes, with outcomes differing from those in ICUs due to the typically lower-risk profile of ward patients. Beyond in-hospital mortality [19] and clinical deterioration [20], unplanned ICU admissions [19, 21], length of stay (LOS) [22], rapid response team activations [21] and non-elective readmissions [23] have been assessed.

Considering the abovementioned evidence, we hypothesize that applying analytical procedures to already existing monitoring protocols may yield clinically relevant insights. In this regard, the capacity to predict hospitalization LOS during the initial hours following hospital admission would be of significant contribution in two key areas: hospital management and patient care. From a hospital management perspective, timely LOS prediction would allow for targeted interventions to streamline care and reduce unwarranted extended LOS [24], therefore optimizing resource allocation. From a patient care standpoint, short hospitalization LOS has been associated with favourable post-discharge outcomes such as lower early readmission and mortality rates [25], whereas prolonged LOS has been associated with increased risk of adverse events in ICUs [26], acute care wards [27] and general wards [28].

This work presents a protocol for the implementation of non-invasive, continuous vital signs monitoring in general wards, together with a comprehensive pipeline for massive feature extraction and analysis. Through this pipeline, our aim was (i) to collect and identify key features able to profile vital signs time series through their association with hospitalization LOS in COPD and CHF patients; and (ii) to build a predictive model capable of predicting hospitalization LOS in our cohort of COPD and CHF patients.

Materials and methods

Study design

This was a prospective observational study designed to implement an analysis protocol for clinical outcome prediction through vital signs time series in general wards.

Study population

The study population comprised all patients diagnosed with chronic obstructive pulmonary disease (COPD) and chronic heart failure (CHF) who had hospital admissions to non-ICU hospital wards due to COPD exacerbations or CHF decompensations and were submitted to continuous vital signs monitoring in accordance with standard clinical practice at Hospital Universitario Arnau de Vilanova and Hospital Universitario Santa Maria in Lleida, Spain, between January 2023 and February 2024.

Sample size

Our study was as designed a pilot study to assess the feasibility of employing vital signs continuous monitoring to predict hospitalization LOS. Given that COPD and CHF are known to be affected by seasonality, with a clear winter predominance [29, 30], we decided to recruit our cohort during an estimated one-year period. Moreover, we ensured that the resulting sample would suffice according to feasibility study guidelines [31].

Inclusion criteria

Inclusion criteria for COPD patients included being aged ≥ 40 years at the time of admission, having a confirmed COPD diagnosis based on Global Initiative for Chronic Obstructive Lung Disease (GOLD) criteria as per medical records [32], undergoing continuous monitoring of vital signs during their hospital stay according to standard clinical practice, having no active cancer, tuberculosis, pneumoconiosis, pneumonectomy, moderate or severe cognitive decay (Global Deterioration Scale (GDS) ≥ 5) and accepting participation in the study by signing an informed consent. Inclusion criteria for CHF patients included being aged ≥ 45 years at the time of admission, having a confirmed CHF diagnosis based on European Society of Cardiology (ESC) guidelines criteria as per medical records [33], undergoing continuous monitoring of vital signs during their hospital stay according to standard clinical practice and clinician decision, having no decompensation attributable to an identifiable cause unrelated to CHF, having no planned invasive cardiac medical intervention in the following three months, having no active cancer, advanced renal failure (Stage 5: glomerular filtration rate < 15 ml/min) or moderate or severe cognitive decay (GDS ≥ 5) and accepting participation in the study by signing an informed consent. These criteria ensured clinical relevance and population homogeneity. Cognitive impairment exclusion criteria were set to uphold ethical standards.

Patient recruitment

During their hospital stay, potentially eligible patients were visited by a member of the research team and a nurse, who informed them of the study and requested their participation. No specific or predefined criteria for patient monitoring were set before the study, as monitoring depended on monitoring devices’ availability. As a result, patient recruitment was non-consecutive. Vital signs monitoring started before the research team could contact eligible patients, and monitoring data was automatically transferred to hospital servers. Patients who were not monitored were neither contacted by the study team nor registered in any form. In cases where eligible patients signed the informed consent, their monitoring data were transferred to the research team. In the absence of consent, the monitoring data were deleted after hospital discharge. The project was approved by the Ethics Committee of Hospital Arnau de Vilanova of Lleida (CEIC-2731).

Data collection

Continuously-monitored vital signs of eligible patients — specifically heart rate (HR), respiratory rate (RR) and peripheral oxygen saturation (SpO₂) — were extracted from EDAN IM60 multiparametric monitors (Edan Instruments, China). HR was monitored through 5-lead electrocardiogram, RR was monitored through impedance pneumography and SpO₂ was monitored through SpO₂ pulse oximetry. Vital signs values were recorded every thirty seconds during the first 24 h of hospital admission and stored in the local monitoring centre of each nursery wing. The monitoring window was set at the first 24 h of hospital admission to ensure the capture of both daytime and nighttime physiological patterns and to enable early outcome prediction. The information stored in each monitoring centre was then processed using a healthcare-specific integration engine that transmitted all monitoring data through Health Level Seven messaging to the Hospital Arnau de Vilanova cloud server, operated by the hospital’s Information Technology department. This data was then integrated into a database containing eligible patient’s medical record number and their time-annotated vital sign measurements.

Vital signs analysis

Time series processing

Time series data corresponding to HR, RR and SpO₂ were extracted from the continuous monitoring records of the study population. Patients whose monitoring did not commence within the first 24 h of hospital admission were excluded from further analyses. Subsequently, patients’ time series were trimmed to contain only the first 24 h of monitoring. Time series with less than twelve hours of recorded data or less than 75% of data completeness within this 24-hour period were also excluded.

Following the selection of the time series that met the required quality standards, we conducted an imputation process to account for any potential gap in our time series. In this regard, previous research has pointed out that vital signs’ trends are found to correlate [34, 35], especially during acute deterioration episodes, where vital signs can be used as adverse event predictors [36]. Thus, missing data in the selected patients were imputed using Self-Attention-based Imputation for Time Series (SAITS), a Python package specialised in multivariate time series imputation whose self-attention mechanism enabled the identification of potential relationships between the recorded vital signs [37].

Before training the imputation model, we individually scaled the three vital sign time series (RR, HR, and SpO₂). The SAITS model utilizes artificially masked values as training data, employing a diagonally-masked self-attention mechanism to evaluate model performance on the masked data. Consequently, the complete scaled time series dataset was used for model training. Subsequently, the trained model was used to impute any individual time series that met the aforementioned criteria: timely monitoring, more than 12 h of recorded vital signs and over 75% data completeness. This procedure was performed using Python 3.12.6.

Feature extraction

Following imputation, each of the three vital signs time series extracted from each patient was subject to a comprehensive analysis. The selected framework for this procedure was highly comparative time-series analysis (hctsa), a MATLAB package designed for conducting highly comparative time-series analysis by massive feature extraction, computing over 7700 features for each time series [38, 39]. This method has been previously employed to discover signatures of fatal neonatal illness from vital signs [40] or predict post-cardiac arrest outcomes [41]. The analysis of the univariate time series from each of the three vital signs resulted in a single value for each of the 7700 features. This procedure was performed using MATLAB R2024b and hctsa 1.9.0.0.

Vital signs profiling

The extensive feature set derived from each vital signs time series was then processed through an individual feature selection procedure aimed at identifying the most interpretative features, that is, those most strongly associated with hospitalization LOS. This characterization of each vital sign’s time series was achieved using VSURF, an R package specifically designed for feature selection in high-dimensional settings [42]. The features were then ranked according to their permutation-based importance metric. For each of the three vital signs, the top fifty features with the highest value for importance were selected as interpretative variables.

Length of stay prediction

Previous research has demonstrated that dichotomized predictions of hospitalization LOS have more practical implications in clinical decision-making systems than continuous predictions [43, 44]. In this regard, Quintana et al. established a threshold of nine days to distinguish between short and prolonged stays when evaluating the performance of various LOS predictors in COPD exacerbations [45]. Accordingly, we adopted the same nine-day threshold for LOS, categorizing hospitalizations under nine days as “Short LOS” and hospitalizations equal to or exceeding nine days as “Prolonged LOS”.

The predictive features obtained from each of the three VSURF analyses were extracted and combined to create a prediction feature subset, merging the features derived from the three vital signs time series. The dichotomized LOS variable was then used as the response variable in a logistic regression model built using the predictive subset. The performance of the predictive model was assessed using 10-fold cross-validation.

Results

Characteristics of the cohort

Out of all potentially eligible patients, a total of 101 patients fulfilled inclusion criteria and were therefore enrolled in the study, and 74 were eligible for analysis as per signal quality requirements. The patient recruitment workflow can be seen in Fig. 1.

Fig. 1.

Patient recruitment pipeline. The recruited patients refer to eligible patients who met inclusion criteria and were therefore enrolled in the study. As the study was conducted under real-life clinical conditions, the total number of potentially eligible patients cannot be calculated

The baseline characteristics of the Short Stay and Prolonged Stay groups can be seen in Table 1. Briefly, the study cohort had a median [p25;p75] age of 82.5 [75.0;89.0] years. 31 (41.9%) patients were female and 43 (58.1%) were male. 39 (52.7%) patients were admitted due to a COPD exacerbation and 35 (47.3%) were admitted due to a CHF decompensation.

Table 1.

Sociodemographic and anthropometric characteristics of the cohort

	Short Stay	Prolonged Stay	p-value
	N = 34	N = 40
	n (%) or median [p25;p75]	n (%) or median [p25;p75]
Sociodemographic and Anthropometric Characteristics
Age (years)	79.5 [67.0;88.8]	85.0 [77.0;89.2]	0.114
Sex			1.000
Female	14 (45.2%)	17 (54.8%)
Male	20 (46.5%)	23 (53.5%)
Weight (kg)	73.2 [60.0;90.9]	73.1 [64.5;88.0]	0.741
Height (m)	1.62 [1.56;1.70]	1.63 [1.55;1.68]	0.983
BMI (kg/m2)	27.5 [22.7;32.1]	28.7 [25.6;31.6]	0.501
Habits
Smoking Status			0.245
Active Smoker	7 (70.0%)	3 (30.0%)
Former Smoker	17 (47.2%)	19 (52.8%)
Non smoker/Occasional Smoker	10 (37.0%)	17 (63.0%)
Pack-Year Index	42.0 [23.0;64.0]	33.0 [21.8;54.5]	0.584
Alcohol Consumption			0.716
Active Drinker	8 (53.3%)	7 (46.7%)
Former Drinker	6 (50.0%)	6 (50.0%)
Never Drinker/Occasional Drinker	18 (40.9%)	26 (59.1%)
Comorbidities
Diabetes	11 (44.0%)	14 (56.0%)	0.409
Renal Disease	12 (54.5%)	10 (45.5%)	0.522
Charlson Comorbidity Index	6.00 [4.00;7.00]	6.00 [6.00;8.00]	0.126
Arterial Hypertension	24 (41.4%)	34 (58.6%)	0.223
Atrial Fibrillation	11 (32.4%)	23 (67.6%)	0.041
Hypercholesterolemia	7 (63.6%)	4 (36.4%)	0.367
Asthma	2 (40.0%)	3 (60.0%)	1.000
Medical Consultations in the Previous 12 Months
Total Medical Consultations	16.0 [8.00;28.0]	16.0 [11.0;24.0]	0.782
Consultations for COPD or CHF	3.00 [2.00;9.00]	4.50 [2.00;9.00]	0.737
Laboratory Parameters at Hospital Admission
Glucose (mg/dL)	115 [92.2;143]	122 [99.0;162]	0.318
Creatinine (mg/dL)	0.89 [0.76;1.31]	1.12 [0.74;1.41]	0.337
Sodium (mmol/L)	140 [137;141]	140 [138;141]	0.914
Potassium (mmol/L)	3.95 [3.76;4.31]	4.16 [3.95;4.38]	0.409
Haemoglobin (g/dL)	13.1 [11.5;14.3]	12.1 [10.6;14.0]	0.170
Leukocyte count (x109/L)	8.82 [6.96;11.6]	9.22 [7.12;11.3]	0.733
Platelet count (x109/L)	232 [175;301]	231 [172;298]	0.778
Neutrophil count (x109/L)	6.60 [4.73;9.67]	7.49 [5.58;9.81]	0.432
Lymphocyte count (x109/L)	1.07 [0.67;1.76]	0.86 [0.68;1.21]	0.206
Monocyte count (x109/L)	0.56 [0.30;0.81]	0.63 [0.28;0.80]	0.931
Eosinophil count (x109/L)	0.08 [0.00;0.16]	0.03 [0.00;0.09]	0.228
Basophil count (x109/L)	0.03 [0.01;0.06]	0.03 [0.01;0.03]	0.263
Exacerbation Type			0.228
COPD	21 (53.8%)	18 (46.2%)
CHF	13 (37.1%)	22 (62.9%)

In continuous variables, p−values have been obtained through the the Kruskall−Wallis test. In categorical variables, p−values were calculated using the chi−square test, or the exact Fisher test when the expected frequencies were less than five

Vital signs analysis

Time series processing

A total of 1604.63 h of monitoring data were obtained from the 74 patients. The median number of monitoring hours per patient was 22.46 h [21.01;23.15].

Feature extraction

The hctsa feature generation process rendered a total of 7755 features per vital sign. After filtering erroneous or non-calculated features, 6404 features with adequate values in the three vital signs were scaled and selected for the following feature selection process.

Vital signs profiling

The HR time series (Fig. 2) is associated with hospitalization LOS through features capturing the auto-mutual information of its distribution (CO_HistogramAMI) and its self-correlation structure (CO_fzcglscf). Furthermore, the analysis of interpretative features showed a cluster of features indicating robust recurring patterns (SP_Summaries_pgram_hamm) and a cluster of features related to scaling properties of fluctuations in time series (SC_FluctAnal), which are commonly linked to self-similarity and long-range correlations of signals. Overall, the interpretative features found in HR suggest a strong degree of association between self-correlation and periodicity along the HR time series and hospitalization LOS. A summary of the top fifty interpretative features of the HR time series can be found in Table 2.

Fig. 2.

Variable importance (top) and Pearson correlation matrix (down) of the fifty selected interpretative features for HR

Table 2.

Top Fifty interpretative features of heart rate time series

Feature Type	Feature Name (as per HCTSA)	Nº of occurrences	Feature Description
Correlation	CO_fzcglscf	2	The first zero-crossing of the generalized self-correlation function
	CO_HistogramAMI	5	The automutual information of the distribution using histograms
	IN_AutoMutualInfoStats	1	Statistics on automutual information function of a time series
Distribution	DN_RemovePoints	1	How time-series properties change as points are removed
	DN_SimpleFit	1	Fit distributions or simple time-series models to the data
	EN_DistributionEntropy	2	Estimates of entropy from the distribution of a data vector
Extreme Values	EX_MovingThreshold	1	Moving threshold model for extreme events in a time series
Model Fitting	MF_CompareTestSets	1	Robustness of test-set goodness of fit
	MF_GP_LocalPrediction	1	Gaussian Process time-series model for local prediction
Non-Linear Analysis	NL_TSTL_FractalDimensions	3	Fractal dimension spectrum of a time series
	NL_TSTL_LargestLyap	1	Largest Lyapunov exponent of a time series
	SC_FluctAnal	9	Implements fluctuation analysis by a variety of methods
Pre-Processing	PP_Compare	5	Compare how time-series properties change after pre-processing
Symbolic Transformations	SB_MotifThree	1	Motifs in a coarse-graining of a time series to a 3-letter alphabet
	SB_TransitionMatrix	2	Transition probabilities between time-series states
Wavelet Transforms	SP_Summaries	8	Statistics of the power spectrum of a time series
Basic Statistics	ST_LocalExtrema	1	How local maximums and minimums vary across the time series
Stationarity and Step Detection	SY_LocalGlobal	2	Compares local statistics to global statistics of a time series
	SY_SpreadRandomLocal	1	Bootstrap-based stationarity measure
	CP_l1pwc_sweep_lambda	1	Dependence of step detection on regularization parameter

The analysis of the interpretative features of RR associated with hospitalization LOS (Fig. 3) shows a more widespread distribution of features. The correlation matrix in Fig. 3 shows a small cluster of features related to autoregressive model fitting using autocovariance (MF_AR_arcov), as well as clusters relating to predictability (MF_steps_ahead) or fluctuation analysis (SC_FluctAnal). Overall, the interpretative features found in RR indicate a heterogeneous distribution of variance, with highlighted clusters suggesting an association between hospitalization LOS and a self-correlated, persistent temporal pattern in the RR signal. A summary of the top fifty interpretative features of the RR time series can be found in Table 3.

Fig. 3.

Variable importance (top) and Pearson correlation matrix (down) of the fifty selected interpretative features for RR

Table 3.

Top Fifty interpretative features of respiratory rate time series

Feature Type	Feature Name (as per HCTSA)	Nº of occurrences	Feature Description
Correlation	CO_Embed2_Basic	1	Point density statistics in a 2-d embedding space
	CO_StickAngles	1	Analysis of line-of-sight angles between time-series data points
	IN_AutoMutualInfoStats	1	Statistics on automutual information function for a time series.
	ST_MomentCorr	1	Correlations between simple statistics in local windows of a time series.
Distribution	DN_RemovePoints	1	How time-series properties change as points are removed
Entropy	EN_DistributionEntropy	1	Estimates of entropy from the distribution of a data vector
	EN_Randomize	4	How time-series properties change with increasing randomization
Model Fitting	FC_LocalSimple	3	Simple local time-series forecasting
	FC_Surprise	1	Quantifies measures of surprise given recent memory
	MF_AR_arcov	3	Fits an AR model of a given order, p
	MF_StateSpace_n4sid	3	State space time-series model fitting
	MF_steps_ahead	5	Goodness of model predictions across prediction lengths
Non-Linear Analysis	NL_crptool	1	Analyzes the false-nearest neighbors statistic
	NL_MS_nlpe	1	Normalized drop-one-out constant interpolation nonlinear prediction error
	NL_TSTL_LargestLyap	1	Largest Lyapunov exponent of a time series
	SC_FluctAnal	6	Implements fluctuation analysis by a variety of methods
Visibility Graph	NW_VisibilityGraph	3	Visibility graph analysis of a time series
Dynamical Modeling	PH_ForcePotential	1	Couples the values of the time series to a dynamical system
Pre-processing	PP_Compare	5	Compare how time-series properties change after pre-processing
	PP_Iterate	2	How time-series properties change in response to iterative pre-processing
Wavelet Transforms	SP_Summaries	1	Statistics of the power spectrum of a time series
Basic Statistics	ST_LocalExtrema	1	How local maximums and minimums vary across the time series
Stationarity	SY_DriftingMean	2	Mean and variance in local time-series subsegments
	SY_SlidingWindow	1	Sliding window measures of stationarity

The analysis of the interpretative features of SpO₂ associated with hospitalization LOS (Fig. 4) reveals a small cluster of wavelet-based transformation features (WL_cwt), related to long-term trends and short-term fluctuations. In addition, the analysis renders two clusters indicative of local patterns and anomalies as well as short term dynamics and fluctuations (SB_MotifThree). Overall, the highlighted clusters suggest that hospitalization LOS is associated to both stable trends and unpredictable local dynamics in the SpO₂ signal. A summary of the top fifty interpretative features of the HR time series can be found in Table 4.

Fig. 4.

Variable importance (top) and Pearson Correlation Matrix (down) of the fifty selected interpretative features for SpO₂

Table 4.

Top Fifty interpretative features of oxygen saturation time series

Feature Type	Feature Name (as per HCTSA)	Nº of occurrences	Feature Description
Correlation	CO_AddNoise	1	Changes in the automutual information with the addition of noise
	IN_AutoMutualInfoStats	1	Statistics on automutual information function of a time series
Distribution	DN_OutlierInclude	4	How statistics depend on distributional outliers
	DN_RemovePoints	4	How time-series properties change as points are removed
	DN_SimpleFit	1	Fit distributions or simple time-series models to the data.
Entropy	EN_Randomize	4	How time-series properties change with increasing randomization
Model Fitting	MF_arfit	1	Statistics of a fitted AR model to a time series.
	MF_ExpSmoothing	3	Exponential smoothing time-series prediction model
	MF_GP_hyperparameters	2	Gaussian Process time-series model parameters and goodness of fit.
	MF_hmm	1	Hidden Markov Model (HMM) fitting to a time series
Dynamical Modeling	PH_ForcePotential	1	Couples the values of the time series to a dynamical system
Pre-processing	PP_Compare	4	Compare how time-series properties change after pre-processing
Symbolic Transformations	SB_MotifThree	8	Motifs in a coarse-graining of a time series to a 3-letter alphabet
	SB_TransitionMatrix	1	Transition probabilities between time-series states
Stationarity	SY_RangeEvolve	2	How the time-series range changes across time
	CP_ML_StepDetect	2	Analysis of discrete steps in a time series.
Non-Linear Analysis	TSTL_localdensity	1	Local density estimates in the time-delay embedding space.
Wavelet Transforms	WL_cwt	4	Continuous wavelet transform of a time series
	SP_Summaries	5	Statistics of the power spectrum of a time series

Length of stay prediction

The complementary information captured by the predictive features of the three vital signs time series was integrated into a model to predict hospitalization LOS. The model was built with six predictive features from the HR time series, six predictive features from the RR time series and seven predictive features from the SpO₂ time series. The selected predictive features, along with a brief description of their characteristics and functioning, are presented in Table 5.

Table 5.

Predictive features for hospitalization LOS included in the predictive model

	Feature Name (as per hctsa)	Feature Description
Heart Rate	CO_HistogramAMI_even_2bin.ami1	Automutual information of the distribution using histograms
	SC_FluctAnal_2_dfa_50_0_logi.ssr	Fluctuation analysis by polynomial trends
	SC_FluctAnal_2_range_50_logi.ssr	Fluctuation analysis by range
	NL_TSTL_FractalDimensions_2_100_02_1 _5_10_32_1_5.linfit_a	Fractal dimension spectrum of a time series
	ST_LocalExtrema_n100.minmaxonminabsmin	Variation of local maximums and minimums
	SP_Summaries_pgram_hamm.logarea_4_4	Power spectrum statistics
Respiratory Rate	PH_ForcePotential_dblwell_1_02_01.finaldev	Couples the values of the time series to a dynamical system
	SC_FluctAnal_2_range_50_logi.logtausplit	Fluctuation analysis by range
	SC_FluctAnal_sign_2_dfa_50_2_logi.resac1	Fluctuation analysis by polynomial trends
	CO_StickAngles_y.q10_n	Analysis of line-of-sight angles between time-series data points
	EN_Randomize_dyndist.ac2fexprmse	How time-series properties change with increasing randomization
	MF_StateSpace_n4sid_3_05_1.A_9	State space time-series model fitting
Oxygen Saturation	EN_Randomize_dyndist.statav5fexprmse	How time-series properties change with increasing randomization
	SP_Summaries_pgram_hamm.linfitloglog_hf_sigrat	Power spectrum statistics
	SB_MotifThree_quantile.abcc	Motifs in a coarse-graining of a time series to a 3-letter alphabet
	PP_Compare_sin1.swss5_1	Compare how time-series properties change after pre-processing
	MF_ExpSmoothing_05_best.dwts	Exponential smoothing time-series prediction model
	WL_cwt_sym2_32.maxonmeanSC	Continuous wavelet transform of a time series
	PP_Compare_diff2.statav6	Compare how time-series properties change after pre-processing

The predictive model built with the combination of predictive features of the three vital signs achieved an AUC of 0.975, an accuracy of 0.944, a sensitivity of 0.979, and a specificity of 0.900 in 10-fold cross validation.

Discussion

Significance of results

The present study tried to assess the feasibility of using vital signs time series obtained in the first 24 h of hospital admission in the general ward to predict hospitalization LOS. Through a massive, feature-based comparative analysis, we were able to profile the variance found within the three vital signs time series and identify the most relevant patterns associated with hospitalization LOS in COPD and CHF patients. According to the interpretative feature clusters found in the time-series profiling, hospitalization LOS in COPD and CHF patients was associated to periodic and self-correlating patterns in HR and RR time series. In the case of SpO₂ time-series, hospitalization LOS was associated to fluctuant, wavelet-based patterns. These findings are consistent with previous studies that have demonstrated an association between abnormalities or instability in vital signs and adverse clinical outcomes [3, 4].

Our findings indicate that a compact set of features derived from vital signs time series can accurately predict LOS in COPD and CHF patients using only routinely collected data from the first 24 h of admission. The predictive model achieved an AUC of 0.975 and an accuracy of 0.944 in 10-fold cross-validation, utilizing nineteen features. These results highlight the potential of early vital sign dynamics to inform LOS prediction and suggest the broader applicability of our protocol for anticipating different clinical outcomes.

Limitations

Our study has several limitations that should be acknowledged. First, it is important to note that our results are only applicable to COPD and CHF patients meeting inclusion criteria. In this regard, further studies should evaluate the generalizability of our results to other populations. However, it must be noticed that this is a feasibility study which purpose is not to export the obtained algorithms to other settings. Secondly, patient recruitment was not consecutive, as the study was conducted under real-life clinical conditions where vital signs monitoring depended on device availability, and this could cause some degree of selection bias.

Related works

The potential of vital signs monitoring in clinical practice, together with the need of further research, has long been acknowledged [46, 47]. Continuous monitoring of vital signs of inpatients is a common practice in ICUs and medium care settings, in contrast to general wards [5, 48], where cumbersome equipment constrained patient mobility and recovery [49]. However, the appearance of novel and easy-to-wear devices has increased the feasibility of a widespread, continuous monitoring of vital signs in general wards [50, 51]. In this sense, Leenen et al. provided a review of existing studies regarding wireless wearable devices for continuous vital signs measurement [5]. Although continuous monitoring intuitively appears more informative than intermittent monitoring, current evidence has shown mixed results. Bowles et al. found no significant improvements in clinical outcomes for patients in randomised controlled trials comparing continuous monitoring and usual care in general wards [52]. Similarly, Areia et al. reported no strong evidence suggesting that continuous monitoring through wearables is superior to standard care, though trends towards decreased ICU transfers, hospitalization LOS and hospital mortality were found [53]. These reviews, however, are limited to studies comparing patient outcomes between usual care and alarm-based continuous vital sign monitoring, with no further analysis whatsoever of the resulting monitoring data. This research gap underscores the importance of new studies to obtain novel insights from continuous vital sign monitoring data. This is, to our knowledge, the first study to conduct a comprehensive analysis of the characteristic features of vital signs time series in relation to clinical outcomes, in this case hospitalization LOS. In addition, to the best of our knowledge, studies addressing clinical outcome prediction through continuous vital signs monitoring are mainly restricted to ICUs [14] or high-risk settings [54]. Our study is, therefore, the first study to assess hospitalization LOS prediction in general ward patients using continuous vital signs monitoring data.

Lessons learned

From a technological perspective, the use of stationary monitoring devices limited the continuous collection of vital signs data. Whenever a patient left the hospital bed, for instance to go to the bathroom, they became disconnected, requiring nursing staff to manually reconnect them. This translated in monitoring gaps that needed to be addressed after data recollection. Consequently, the use of portable monitoring devices for further monitoring studies is heavily recommended. In this regard, Weenk et al. showcased the capabilities of continuous vital signs monitoring in the general ward through wearable technology [50, 55]. Continuous, non-invasive, wearable, wireless monitoring technology could streamline and improve the collection of vital signs data, while minimizing discomfort for patients.

From a managerial perspective, patient recruitment was largely influenced by the emergency room’s workload, as eligible patients were identified at this stage before being transferred to wards in Hospital Arnau de Vilanova or Santa Maria. Since the emergency room is shared between both hospitals, transfer times were often delayed, resulting in monitoring initiation deferral. Additionally, monitoring data was linked to the bed to which the monitoring device was associated, meaning that patients’ ID codes had to be manually linked to the monitored data.

Conclussions

Through a comprehensive feature-based analysis, we identified key patterns in HR, RR, and SpO₂ time series associated with hospitalization LOS in COPD and CHF patients. Our findings also indicate that a compact set of features derived from vital signs time series can accurately predict LOS in COPD and CHF patients using only routinely collected data from the first 24 h of admission.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AUC

Area Under the Curve

CHF

Chronic Heart Failure

COPD

Chronic Obstructive Pulmonary Disease

ESC

European Society of Cardiology

GDS

Global Deterioration Scale

GOLD

Global Initiative for Chronic Obstructive Lung Disease

hctsa

Highly comparative time-series analysis

HR

Heart Rate

ICU

Intensive Care Unit

LOS

Length Of Stay

RR

Respiratory Rate

SAITS

Self-Attention-based Imputation for Time Series

SpO₂

Peripheral Oxygen Saturation

Author contributions

Ivan Juez-Garcia and Iván D Benítez contributed equally and are co-first authors.I.J.-G.: Conceptualization, Methodology, Software, Formal analysis, Writing - Original Draft, Writing - Review & Editing. I.D.B.: Conceptualization, Methodology, Software, Formal analysis, Writing - Review & Editing. G.T.: Conceptualization, Data Curation. J.G.: Supervision, Project administration. L.U.: Resources, Data Curation. A.P.: Resources, Data Curation. N.V.: Resources, Data Curation. I. C.: Resources, Data Curation. F.B.: Supervision, Project administration. J.dB.: Conceptualization, Methodology, Writing - Review & Editing, Funding acquisition.

Funding

IJ-G was supported by Instituto de Salud Carlos III (ISCIII) through a predoctoral fellowship (PFIS: FI24/00084), co-funded by the European Union. JdB was supported by Instituto de Salud Carlos III (ISCIII; Miguel Servet 2019: CP19/00108), co-funded by the European Social Fund (ESF). FB is supported by the ICREA Academia programme. This study has been funded by Instituto de Salud Carlos III (ISCIII) through the project “PI21/00924”, co-funded by the European Union; beca SEPAR 2022 (proyecto 1253); and beca SOCAP 2023.

Data availability

The data underlying this article cannot be shared publicly to protect patient privacy. The data will be shared on reasonable request to the corresponding author.

Declarations

Competing interests

The authors declare no competing interests.