Physical exercises to prevent and rehabilitate hospital-associated disability in hospitalised older people: A protocol for a living systematic review with network meta-analysis

Current state of research in the field

A systematic review 15 studies from 12 RCTs including 2618 patients reported that supervised exercise was effective in improving or attenuating the decline in functional independence and physical performance in older adults hospitalised for acute care (Valenzuela et al., 2020). More recently, Hartley et al. (2022) conducted a systematic review and evaluated the benefits and harms of exercise interventions in hospitalised older adults on functional ability. The same researchers observed statistically significant effects of exercise interventions on the ability to perform ADLs, in comparison to usual care. However, the differences were not clinically important (Hartley et al., 2022).

The published systematic reviews are limited, and several questions remain unanswered when it comes to preventing HAD in older patients hospitalised in an acute setting. Hartley et al. (2022) excluded studies investigating management of HAD in the rehabilitation setting (Parker et al., 2015; Stevens-Lapsley et al., 2016). While it is well-established that HAD can have a long-term impact on ADL disability, previews systematic reviews have not examined the effects of physical interventions for rehabilitating HAD. The previously mentioned systematic reviews (Bachmann et al., 2010; Hartley et al., 2022; Valenzuela et al., 2020) considered studies with control usual care or a sham-control intervention and these researchers excluded studies with a control group receiving additional exercise. It is important to include such studies in a review to have an overview of the effectiveness of physical exercise interventions in hospitalised older adults in an acute setting. Since the Hartley et al. (2022) study was published (Hartley et al., 2022), new RCTs have been published (Carrasco Paniagua et al., 2022; Godino et al., 2022) adding new evidence to the status quo.

A recent network meta-analysis (NMA) investigated the optimal dose and type of physical exercise activity to improve functional capacity and decrease adverse events in acutely hospitalised people over 50 years of age (Gallardo-Gómez et al., 2023). Our project differs from theirs in several respects: i) our study will include articles with older population (over 65 years of age), which does represent the typical geriatric population, ii) our study will investigate a broad range of types of physical exercise interventions, iii) the long-term effects on functional capacity will be studied and iv) our NMA will be based on a living systematic review (LSR) which will be updated over three years.

We plan to carry out an NMA because it offers several advantages which are relevant in the context of the present review questions: i) NMA allows the integration of multiple comparisons within one analytical framework. For the outcome HAD various physical exercise modalities such as resistance training, aerobic exercises or cycling have been proposed. Unlike pair-wise (head-to-head) meta-analyses (e.g. resistance training vs. aerobic exercises), NMA enables us to compare all modalities simultaneously; ii) NMA enables us to assess the comparative effectiveness of each exercise modality, thus allowing us to propose a hierarchy of physical exercise interventions; iii) NMA increases precision of the effect estimates compared to traditional meta-analyses (Chaimani et al., 2019). The latter advantage of NMA is crucial for outcomes like functional decline in older patients after a hospital stay.

Objectives

In the first year a systematic review on the effectiveness of physical exercise intervention to prevent and rehabilitate HAD in patients aged 65 years and more hospitalised in acute setting will be performed. In the consecutive three following years the systematic review will regularly be updated using a LSR approach.

We will also assess factors (training intensity, volume, and frequency) that potentially influence the treatment effectiveness on HAD in such patients through meta-regression.

Regarding the distinction between the prevention and rehabilitation we will use the following classification: depending on the setting of the interventions we will analyse the effect of preventive physical exercise on HAD in older, hospitalised patients (i.e. the exercises are performed within the acute care setting in order to minimise a potential HAD). If physical exercise interventions are administered in a rehabilitation setting after a hospital stay, we will analyse the effect of rehabilitation on HAD (i.e. the exercises are used to reduce HAD).

The primary outcome of this review will be short term basic functional ability in performing ADL tasks following hospitalisation discharge or rehabilitation discharge (between 0 and 14 days after hospital or rehabilitation discharge).

Secondary outcomes will include i) long term basic functional ability in performing ADL tasks in older patients following hospitalisation or rehabilitation discharge (at least 15 days after discharge). Furthermore, we will analyse the following secondary outcomes: ii) balance; iii) strength; iv) endurance; v) walking abilities; vi) length of stay, vii) re-hospitalisation rate; viii) adverse events and ix) mortality rate.

We will compare the active interventions (i.e. different types of physical exercise) whit each other in the NMA but we will also include non-active comparators such as usual care or sham interventions into the analysis.

Detailed research plan

We will conduct a living systematic review of RCTs with network meta-analysis. Guidelines from the Cochrane Handbook for a systematic review of interventions will be followed (Higgins et al., 2023).

Eligibility criteria

Types of studies that will be included

We will include only published RCTs and cross-over studies if they fulfil the condition of randomised allocation of study participants. Other designs, such as non-randomised clinical trials, retrospective studies, cohort studies, single case reports and letters to editors, will be excluded. Articles in English, French, German, and Italian will be included. Other languages will be excluded if the automatic translation with applications such as Google translate (https://translate.google.com/) or Deepl (https://www.deepl.com/) is not sufficient to understand the content of the study. No restrictions will be placed on the date of publication.

Types of participants and condition to be studied

We will include studies that investigated older adults, aged 65 years and above, who were hospitalised in acute care setting (e.g. hospitalised for an acute illness that responds to usual medical management (Covinsky et al., 2011)) and at risk of developing HAD. HAD refers to disabilities acquired during hospitalisation or the worsening of a pre-existing disability due to hospitalisation (Covinsky et al., 2011). Moreover, we will also include studies investigating older people discharged from rehabilitation centre (after a hospital stay) with a reduced ability to perform ADLs, which were likely caused by hospitalisation. All acute pathologies will be included, except those with a high risk of permanent loss of function, function (e.g., stroke, post COVID-19 condition, fractures, cancer), which are directly disabling.

Types of interventions

In this project, we will include all studies that have assessed physical exercise interventions targeting the prevention and rehabilitation of HAD in older, hospitalised patients. Combined exercises will also be eligible (e.g., resistance training and balance training). Reports analysing exercise programmes including strengthening and mobility exercises compared to usual care in such patients will also be eligible (Jones et al., 2006). Studies implementing cognitive exercise without physical exercises to prevent or rehabilitate HAD in such patients will be excluded. For example, studies investigating cognitive training in fully immersive virtual reality compared to pharmacology to prevent or rehabilitate HAD in older, hospitalised patients will be excluded (Kang et al., 2021).

Types of comparators

Studies included in this systematic review with NMA must investigate at least two groups and one of them must be active (i.e., performing physical exercises). The control group can be a non-active intervention (such as a control or sham intervention).

Types of outcomes

We will include studies that evaluated the immediate and long-term effectiveness of physical exercise on the ability to perform basic ADLs in older patients with HAD following acute hospital care. Additionally, we will consider studies that investigated the effectiveness of these interventions on balance, strength, endurance, walking abilities, length of hospital stays, rehospitalisation rate, adverse events and mortality rates.

Information sources

We will perform the literature search on four electronic bibliographic databases: PubMed (MEDLINE), Embase (via Embase.com), CENTRAL (Cochrane Central Register of Controlled Trials, The Cochrane Library) and CINAHL Ultimate (via EBSCO). The literature search algorithm will first be developed for Medline (PubMed) and then adapted for implementation on the other databases.

All four electronic databases will be screened from inception until present. The search algorithm will be updated every two months for a period of three years from August 2025 on.

We will use the recommended highly sensitive filters for clinical trials suggested by the Cochrane Collaboration (Higgins et al., 2023). Grey literature will be searched by consulting Google Scholar (RRID: SCR_008878), the reference list of important publications on the topic, WHO International Clinical Trials Registry Platform (RRID: SCR_004475), and ClinicalTrials.gov (RRID: SCR_002309). Articles will be searched from inception to the present.

Search strategy

The search strategy will be developed using the litsearchR package (Grames et al., 2019). The following steps will be performed: First, we will perform a naïve search on PubMed using the search terms “hospitalised older people”, “physical exercise” and “functional decline”. The identified records will be downloaded. We will extract potentially relevant search terms from title, abstract and keywords section using Rapid Automatic Keyword Extraction (Rose et al., 2010). In a follow up-step a term and keyword co-occurrence network will be established. The parameters will be set as follows: i) a term had to occur at least in three studies to be included; ii) the minimum total number of times a term have to occur will set to three. We used the ggraph package (Pedersen, 2023) to visualise the network. To identify the most relevant terms we will rank the terms with regard to their strength in the network (i.e. the number of times they co-occur with other terms). We will set the cut-off point for inclusion of search terms to 80% strength and remove the remaining terms. The identified terms will be classified into four search concepts (i.e. terms related to: i) population such as older people; ii) setting (i.e. acute hospital); iii) interventions such as physical exercise and iv) outcome such as functional decline. Terms within each concept will be combined using the Boolean operator “OR”. All four concepts will be combined using “AND”. The search string will be further refined on the individual database such as adding specific subject headings and terms which will not pass the threshold of but will be deemed relevant by the review team. In addition, we will add the database specific methodological filter for clinical trials to the search string (Higgins et al., 2023).

To validate the search string we will identify six studies, which fulfil our selection criteria (Blanc-Bisson et al., 2008; Brown et al., 2016; Hu et al., 2020; Martínez-Velilla et al., 2019; McCullagh et al., 2020; Ortiz-Alonso et al., 2020). We will select the reference standard from the systematic reviews of Valenzuela (2020) and Gallardo-Gómez et al. (2023). A search on Embase and PubMed will be identified all studies of our reference standard. The McCullagh study (2020) is not indexed in Embase but can be identified with the PubMed search. The search string for PubMed is presented in Table 2.

Study records

Data management

Bibliographic data will be downloaded from the databases and stored on a secured cloud platform with servers located in Switzerland. We will use the AsysD package (Hair et al., 2023) to de-duplicate records. The Covidence tool (Covidence systematic review software) will be used for screening and data extraction.

Selection process

For the first screening phase, two human reviewers (SW, KG, MS or SG) will independently screen all references by reading the titles and abstracts. In the pilot phase, in parallel with the evaluation of the human evaluators, a selection of articles by artificial intelligence (AI) will also be carried out. The results of the relevant articles included by the two human reviewers supported by AI-software (such as https://asreview.nl/) will be compared. In case of different results, another human reviewer (SG or MS) will be involved in the decision-making process and the two people will find a consensus through discussion. If no consensus is found, the article will automatically be included in the full-text reading. In the living systematic review study phase, if the results of the articles selection obtained by the AI-software match the results of the human assessors, we will be making the selection using only a human reviewer supported by AI.

At the full-text selection stage, articles will be independently reviewed by two human (SW, KG or MS) reviewers according to the inclusion criteria. If there is any disagreement, a third person (SG) will make the final decision. Articles included in the systematic review will be incorporated into the NMA if they provide sufficient information about outcome data. If data is missing, we will reach out to the original authors for further information.

Data collection process

One reviewer will extract data manually (SW or MS) and a second reviewer (KG or SG) will control them independently. Disagreements and errors will be discussed and resolved through a joint agreement. Data extraction will be performed though forms designed in Power App (RRID: SCR_021865) and the functionality of the forms will be tested by on person (SW) by integrating different studies.

Data items

The following information will be extracted for the study characteristics: authors of the publication, year of publication, country of study, age of participants, number of participants randomised, gender proportion, pathologies, type of physical exercise, training modalities (frequency of interventions, duration, and intensities), type of control intervention, length of intervention. For the following variables under investigation, we will extract the central tendencies and measures of variation: score of functional ability assessment tools (e.g. Katz Index, Barthel Index, etc.), physical performance parameters (e.g. balance, strength, endurance, and walking abilities). We will also extract hospital length of stay, the rehospitalisation rate, adverse events and mortality rate.

Outcomes and prioritisation

Primary outcomes

The short-term effectiveness of physical exercise interventions that target prevention of HAD and the rehabilitation of HAD in older hospitalised patients will be assessed in the following ways:

All tools designed to measure basic ability in everyday activities will be considered for inclusion. Examples of potential relevant tools are the Barthel Index (Mahoney & Barthel, 1965), the Functional Independence Measure (Keitll et al., 1987) or the 5-items Katz Index (1963).

If a study uses different measures of ADLs, the order of priority for extraction of the measures is based on the recommendations of Hopman-Rock et al. (2019): the Functional Autonomy Measurement System (SMAF) (Boissy et al., 2007), 5-items Katz Index (1963), Functional Independence and Difficulty Scale (FIDS) (Saito et al., 2016), and the Barthel Index (Mahoney & Barthel, 1965).

Secondary outcomes

The long-term effect of physical exercise interventions (i.e. for prevention and rehabilitation of HAD in older, hospitalised patients in an acute care setting) on ADLs will be investigated as secondary outcome. ADLs will be assessed at the last available time point, provided that other follow-up measures are available. We have defined long-term disability after hospital discharge with a period of interest from 15 to 90 days after hospital discharge.

Additional outcomes will include the investigation of other health parameters affected by physical inactivity due to bed rest such as balance, strength, endurance, and walking abilities will be assessed. We will also assess length of stays, rehospitalisation rate, adverse events and mortality rate.

Risk of bias in individual studies

For risk of bias (RoB) assessment, we will use the Cochrane Group's RoB 2.0 tool (Sterne et al., 2019). RoB 2.0 will be evaluated independently by two reviewers (SW, KG or SG). The differences in evaluation will be discussed between the two reviewers until a consensus is reached. RoB will be analysed for the primary outcome, i.e. dependence on ADLs. In a sensitive analysis, we will remove articles that have been identified as having a high RoB. A high risk of bias refers to situations where there are serious concerns about the methodological quality of the study that are likely to lead to misleading or unreliable results. The RoB 2.0 (Risk of Bias 2.0) tool, assess five domains: i) bias arising from the randomisation process; ii) bias due to deviations from the intended interventions; iii) bias due to missing outcome data; iv) bias in measurement of the outcome; v) bias in selection of the reported result (Sterne et al., 2019).

A low risk of bias is considered when all five domains of the RoB 2.0 are rated with minimal to no concerns that the results could be affected by systematic errors (Sterne et al., 2019).

Data synthesis

Pair-wise meta-analyses

The statistical analyses will be performed using the statistical software package R (Team, 2023). In the first step, we will use the “meta” package (Schwarzer et al., 2015) to analyse the direct evidence with pair-wise meta-analyses (i.e., the same interventions will be analysed together). For example, all available trials reporting aerobic exercise versus a control intervention will be analysed together. This “head-to-head” approach will be repeated for all identified comparisons and will provide direct evidence of the effectiveness of the physical exercise interventions on HAD in older adults. A head-to-head comparison in meta-analyses is defined as the direct comparison of two interventions in the same study. This type of comparison is straightforward because it is based on the analysis of data in which the interventions have been directly compared against each other in RCTs.

Summary measures for pair-wise meta-analyses

For the analysis of continuous outcomes, number of study participants, means and standard deviations (SDs) of the included studies will be used. If the study means are expressed in the same units (e.g., all studies use the Barthel Index), we will use the between-group raw mean difference as effect size. In cases where the mean is expressed in different units across studies, the standardised between-group mean difference (SMD) will be selected as the effect size (Takeshima et al., 2014). SMD will be interpreted according to Cohen (1988). This means that an effect size around 0.2 will be considered as small, 0.5 as medium and 0.8 as large. Dichotomous outcomes will be analysed using Peto odds ratios (Yusuf et al., 1985). We will employ the Peto method for effect size calculation. The between study variance (Tau²) will be estimated using the Paule-Mandel estimator.

Method of analysis for pair-wise meta-analyses

As we assume that the included studies have some differences in terms of interventions and study populations, a random-effects model will be used for the analyses. The Sidik-Jonkman approach will be used to determine the between-study heterogeneity (Sidik & Jonkman, 2005). In addition, the Hartung-Knapp adjustment for confidence intervals will be used (Hartung, 1999; Hartung & Knapp, 2001).

Assessment of inconsistency for pair-wise meta-analyses

We will assess statistical heterogeneity using the Chi-squared test with its corresponding degrees of freedom and p-value. Higgins’ I² statistic will be calculated as a measure of heterogeneity following the Cochrane Handbook recommendations (Higgins et al., 2023). In addition, we will calculate and use tau-squared to quantify the variance of the true effect sizes and tau for the standard deviation of the true effect sizes (Borenstein et al., 2021).

Analysis of moderator variables for pair-wise meta-analyses

The second step will be to perform a meta-regression in R using the “meta” package (Schwarzer et al., 2015). A mixed effects model will be used, and a restricted maximum likelihood estimator will be employed for estimating between-study variance (Tau²). We will explore the effect of the following potential moderator variables on the estimated effect of the HAD interventions: i) intervention volume (the total dose of received intervention in minutes); ii) average hospitalisation duration and iii) average age of the study participants. According to Paterson and Warburton (2010), there seems to be a dose-response relationship between physical activity and functional outcomes. However, the adequate treatment ‘dose’ in older adults in terms of recommendation is unclear. There is evidence that ‘higher’ levels of physical activity are advantageous to be independent in ADL. Therefore, it is necessary to establish if there is a minimum threshold of physical activity in order to achieve benefits in functional independence in older adults (Paterson & Warburton, 2010).

Network meta-analysis

In a second step a frequentist NMA will be performed using the “netmeta” package of the R software (Rücker et al., 2015). The transitivity assumption will be examined for all studies included in the NMA (Salanti et al., 2014). Firstly, we will investigate whether potential effect modifiers are sufficiently similar between studies. For this, we will investigate important clinical and methodological characteristics such as the severity of illness at baseline (e.g. proportion of participants hospitalised for acute conditions, the pathologies, the age of participants and reasons for admission), sample size and study quality (or RoB). This evaluation will be based on a clinical understanding of the disease (i.e. the research team will review this characteristics). If sufficient studies are available, we will statistically assess the distribution of the effect modifiers (Chaimani et al., 2017). Secondly, we will check whether all interventions are meaningful to all participants. This means that any participant could theoretically receive any treatment in the network of interventions and that the included trials must be sufficiently similar, except for the treatments being compared (Chaimani et al., 2017). If the requirement for a NMA is not met, we will analyse results with a standard pair-wise meta-analysis.

Geometry of the network

The network of treatments will be visualised using network plots using the “netmeta” package of the R software (Rücker et al., 2015). Each intervention will be represented as a circle (i.e. the network nodes) and circles will be connected with lines (i.e. edges) if a direct comparison (i.e. head-to-head comparison) between two interventions is available. The network will be examined for the presence of open and closed loops. A closed loop refers to comparisons with multiple pathways, incorporating both direct and indirect evidence. An open loop refers to comparisons lacking direct evidence (Antoniou et al., 2019). To facilitate interpretation of the network the comparisons will be colour coded.

Summary measures for NMA

We will use the same summary measures in the NMA as proposed for the pair-wise meta-analyses above. That is, standardised mean differences will be used for continuous outcomes and Peto odds ratios (Yusuf et al., 1985) will be used for dichotomous outcomes. A treatment ranking will be produced using the P-score, which is the frequentist equivalent to the SUCRA score (Rücker & Schwarzer, 2015). The SUCRA score (is a statistical measure employed in network meta-analysis to rank treatments based on their effectiveness. It provides a single number that reflects the likelihood that a treatment is the best option among all alternatives. A higher SUCRA score indicates a higher likelihood of the treatment being the most effective (Rücker & Schwarzer, 2015).

Planned method of analysis for NMA

Assuming a relatively high degree of clinical and methodological heterogeneity between the included studies, we will perform a random-effects model. As treatment effects in multi-arm studies are not independent, comparisons of each multi-arm study will be reweighted following Rücker and Schwarzer (2014).

Assessment of inconsistency

Heterogeneity and inconsistency will be quantified with Higgins’ I² statistics (Higgins et al., 2019) using the “netmeta” package of the R software (Rücker et al., 2015). The value of heterogeneity and inconsistency will be interpreted in accordance with the Cochrane Handbook: 0% to 40% may be considered as not important; 30% to 60% may represent moderate heterogeneity; 50% to 90% may represent substantial heterogeneity; 75% to 100% considerable heterogeneity (Higgins et al., 2019). In addition, heterogeneity (within designs) and inconsistency (between designs) will be explored. A comparison of direct treatment estimates and indirect estimates will be performed to check the robustness of the indirect treatment estimates. This comparison will be supported by i) net league tables contrasting direct effect estimates and network effect estimates and ii) net split forest plots showing direct, indirect and network effect estimates. Furthermore, we will visualise inconsistency with net heat plots.

Additional analyses

Planned separated sensitivity analyses will be performed by including only studies with an overall low RoB rating (i.e. the network metanalysis will be performed with the same parameters described above and considering only studies with a low overall RoB rating), or on specific domains of the RoB tool 2.0 (Sterne et al., 2019) (randomisation process, deviation from intended interventions, missing outcome data, measurement of the outcome, and selection of reported results). Additional post hoc sensitivity analyses will be considered to evaluate if the results are robust to the choices that were made during the review process such as analysis of cross-over studies or imputed standard deviation values. Missing values such as means and standard deviation will be imputed following the Cochrane Handbook (ref ) and if we are in doubt about an imputed value sensitivity analyses will be performed.

Meta-bias(es)

We will use comparison adjusted funnel plots to investigate a possible publication bias (Chaimani & Salanti, 2012) using the “netmeta” package of the R software (Rücker et al., 2015). In order to use this statistical approach, we will check if the studies can be ordered in a meaningful way after inclusion of all studies (Salanti et al., 2014). A possible order could be the degree of intensity ranging from low intensity exercises to high intensity exercises. Other possible order variables could be: i) the novelty of the exercise type ranging from traditional exercise over gamified exercises to virtual reality supported exercises; or ii) meaningful functional exercises vs analytical exercises.

Grading of the level of evidence

We will use the Confidence in NMA (CINeMA, RRID: SCR_023193) framework (Salanti et al., 2014) (https://cinema.ispm.unibe.ch), which is based on the Grading of Recommendations Assessment, Development and Evaluation (GRADE), to rate the quality level of the evidence. If a NMA cannot be performed, we will use the standard grading system from the GRADE working group (Schünemann et al., 2013).

Dissemination

Once the initial systematic review is completed, the article will be published in an open access peer-reviewed journal. Afterwards the bi-monthly search update of the results will be presented on the project website (https://had-prevention.netlify.app). In case of relevant changes of the findings an update of the LSR will be submitted to a peer-reviewed journal.

Study status

Adapting the search string in the different databases.