Algorithm‐based pain management for people with dementia in nursing homes

Abstract

Background

People with dementia in nursing homes often experience pain, but often do not receive adequate pain therapy. The experience of pain has a significant impact on quality of life in people with dementia, and is associated with negative health outcomes. Untreated pain is also considered to be one of the causes of challenging behaviour, such as agitation or aggression, in this population. One approach to reducing pain in people with dementia in nursing homes is an algorithm‐based pain management strategy, i.e. the use of a structured protocol that involves pain assessment and a series of predefined treatment steps consisting of various non‐pharmacological and pharmacological pain management interventions.

Objectives

To assess the effects of algorithm‐based pain management interventions to reduce pain and challenging behaviour in people with dementia living in nursing homes.

To describe the components of the interventions and the content of the algorithms.

Search methods

We searched ALOIS, the Cochrane Dementia and Cognitive Improvement Group's register, MEDLINE, Embase, PsycINFO, CINAHL (Cumulative Index to Nursing and Allied Health Literature), Web of Science Core Collection (ISI Web of Science), LILACS (Latin American and Caribbean Health Science Information database), ClinicalTrials.gov and the World Health Organization's meta‐register the International Clinical Trials Registry Portal on 30 June 2021.

Selection criteria

We included randomised controlled trials investigating the effects of algorithm‐based pain management interventions for people with dementia living in nursing homes. All interventions had to include an initial pain assessment, a treatment algorithm (a treatment plan consisting of at least two different non‐pharmacological or pharmacological treatment steps to reduce pain), and criteria to assess the success of each treatment step. The control groups could receive usual care or an active control intervention. Primary outcomes for this review were pain‐related outcomes, e.g. the number of participants with pain (self‐ or proxy‐rated), challenging behaviour (we used a broad definition that could also include agitation or behavioural and psychological symptoms assessed with any validated instrument), and serious adverse events.

Data collection and analysis

Two authors independently selected the articles for inclusion, extracted data and assessed the risk of bias of all included studies. We reported results narratively as there were too few studies for a meta‐analysis. We used GRADE methods to rate the certainty of the results.

Main results

We included three cluster‐randomised controlled trials with a total of 808 participants (mean age 82 to 89 years). In two studies, participants had severe cognitive impairment and in one study mild to moderate impairment. The algorithms used in the studies varied in the number of treatment steps. The comparator was pain education for nursing staff in two studies and usual care in one study.

We judged the risk of detection bias to be high in one study. The risk of selection bias and performance bias was unclear in all studies.

Self‐rated pain (i.e. pain rated by participants themselves) was reported in two studies. In one study, all residents in the nursing homes were included, but fewer than half of the participants experienced pain at baseline, and the mean values of self‐rated and proxy‐rated pain at baseline and follow‐up in both study groups were below the threshold of pain that may require treatment. We considered the evidence from this study to be very low‐certainty and therefore are uncertain whether the algorithm‐based pain management intervention had an effect on self‐rated pain intensity compared with pain education (MD ‐0.27, 95% CI ‐0.49 to ‐0.05, 170 participants; Verbal Descriptor Scale, range 0 to 3). In the other study, all participants had mild to moderate pain at baseline. Here, we found low‐certainty evidence that an algorithm‐based pain management intervention may have little to no effect on self‐rated pain intensity compared with pain education (MD 0.4, 95% CI ‐0.58 to 1.38, 246 participants; Iowa Pain Thermometer, range 0 to 12).

Pain was rated by proxy in all three studies. Again, we considered the evidence from the study in which mean pain scores indicated no pain, or almost no pain, at baseline to be very low‐certainty and were uncertain whether the algorithm‐based pain management intervention had an effect on proxy‐rated pain intensity compared with pain education. For participants with mild to moderate pain at baseline, we found low‐certainty evidence that an algorithm‐based pain management intervention may reduce proxy‐rated pain intensity in comparison with usual care (MD ‐1.49, 95% CI ‐2.11 to ‐0.87, 1 study, 128 participants; Pain Assessment in Advanced Dementia Scale‐Chinese version, range 0 to 10), but may not be more effective than pain education (MD ‐0.2, 95% CI ‐0.79 to 0.39, 1 study, 383 participants; Iowa Pain Thermometer, range 0 to 12).

For challenging behaviour, we found very low‐certainty evidence from one study in which mean pain scores indicated no pain, or almost no pain, at baseline. We were uncertain whether the algorithm‐based pain management intervention had any more effect than education for nursing staff on challenging behaviour of participants (MD ‐0.21, 95% CI ‐1.88 to 1.46, 1 study, 170 participants; Cohen‐Mansfield Agitation Inventory‐Chinese version, range 7 to 203).

None of the studies systematically assessed adverse effects or serious adverse effects and no study reported information about the occurrence of any adverse effect. None of the studies assessed any of the other outcomes of this review.

Authors' conclusions

There is no clear evidence for a benefit of an algorithm‐based pain management intervention in comparison with pain education for reducing pain intensity or challenging behaviour in people with dementia in nursing homes. We found that the intervention may reduce proxy‐rated pain compared with usual care. However, the certainty of evidence is low because of the small number of studies, small sample sizes, methodological limitations, and the clinical heterogeneity of the study populations (e.g. pain level and cognitive status). The results should be interpreted with caution. Future studies should also focus on the implementation of algorithms and their impact in clinical practice.

Plain language summary

Step‐by‐step (algorithm‐based) pain management for people with dementia living in nursing homes

What is the aim of this review?

We were interested in how nurses can best manage pain in people with dementia living in nursing homes. Pain management involves measuring pain and providing pain treatment if necessary. We aimed to find out whether step‐by‐step guidance (an algorithm) for nurses on how to manage pain can reduce pain or behaviours that may indicate someone is in distress (such as hitting, shouting or wandering).

What was studied in the review?

People with dementia in nursing homes often experience pain. However, they cannot always tell their caregivers if they are in pain, so it can be difficult to recognise, and we know that nursing home residents with dementia receive less pain medication than those without dementia. Untreated pain can have a negative impact on well‐being and health, and can also be one reason for challenging behaviour, such as aggression. The use of detailed step‐by‐step guidance for nursing staff, in this review called an algorithm, is designed to improve pain management. Algorithms start with a structured pain assessment and then set out different treatment steps, which can be non‐medication or medication treatments for reducing pain. If pain is detected, the treatment described in the first step is applied. If this treatment does not reduce pain, the treatment from the next step is applied, and so on.

Studies included in the review

In June 2021 we searched for trials that investigated pain management based on the use of an algorithm. We found three studies including 808 participants. Two of these studies compared algorithm‐based pain management with education for the nursing staff on pain and dementia, and one study compared algorithm‐based pain management with usual care.

The level of pain and the severity of the participants' dementia differed in the three studies. One study included all the residents in the nursing homes, most of whom had no pain, or almost no pain, at the start of the study (fewer than half of the included people experienced pain), and two studies included only people with mild to moderate pain. In one study the participants' dementia was of mild or moderate severity and in two studies the participants had severe dementia.

In two studies, those people with dementia who were able to do so reported on their own pain and the nursing staff also judged whether the participants showed signs of pain. In the third study, pain was rated by members of the research team, but not by the participants themselves. The nurses and the researchers used observations of things like facial expressions, gestures and breathing to judge whether someone was in pain.

What are the key findings?

When we looked at the study in which people started out on average with no pain, or almost no pain, we could not be certain whether algorithm‐based pain management had an effect on the intensity of pain they experienced during the study. This was true whether the study participants reported on their own pain or whether nurses judged pain intensity. We also could not tell from this study whether algorithm‐based pain management reduced challenging behaviour.

For people who started out with mild to moderate pain, we found that, compared to education for nursing staff, algorithm‐based pain management may have little or no effect on pain intensity reported by the people themselves (based on the results from one study). When the pain was rated by somebody else (a 'proxy', who was a nurse or research assistant), we found that algorithm‐based pain management may be better than usual care, although it may not be more effective than pain education. However, it is difficult to be sure about the accuracy of pain ratings made by other people.

Our confidence in the results was limited because of the small number of included studies, the variation in the intensity of pain and in the severity of the participants' dementia at the start of the trials, and the quality of the studies.

No study looked for harmful effects, and no study described that any harmful effects occurred.

What is the conclusion?

We found no good evidence that introducing an algorithm to guide pain management for people with dementia in nursing homes is any better than education for nursing staff for reducing pain or challenging behaviours, but it may be better than usual care at reducing pain (rated by observers). The amount of evidence was small, and we could not be certain of the results. More research in this area would be valuable.

Summary of findings

Summary of findings 1. Summary of findings table ‐ Algorithm‐based pain management for people with dementia and mild to moderate pain.

Algorithm‐based pain management for people with dementia and mild to moderate pain
Patient or population: people with dementia and mild to moderate pain Setting: nursing home Intervention: algorithm‐based pain management Comparison: pain education for nursing staff or usual care
Outcomes	Anticipated absolute effects* (95% CI)		Relative effect (95% CI)	№ of participants (studies)	Certainty of the evidence (GRADE)	Comments
	Risk with pain education for nursing staff or usual care	Risk with algorithm‐based pain management
Pain intensity (self‐rated) assessed with: Iowa Pain Thermometer; lower scores indicate lower pain intensity Scale from: 0 to 12 follow‐up: mean 6 months	The mean pain intensity (self‐rated) was 4.6	MD 0.4 higher (0.58 lower to 1.38 higher)	‐	246 (1 RCT)	⊕⊕⊝⊝ Lowa,b	Comparator was pain education for nursing staff.
Pain intensity (proxy‐rated) assessed with: different instruments; range: see text; lower scores indicate lower pain intensity follow‐up: range 12 weeks to 6 months	One study found no evidence of a difference in pain intensity between the intervention and control groups where the comparator was pain education for nursing staff (MD ‐0.2, 95% CI ‐0.79 to 0.39; range 0 to 12; 383 participants) and one study found a reduction of pain in the intervention group compared with the control group receiving usual care (MD ‐1.49, 95% CI ‐2.11 to ‐0.87; range 0 to 10; 128 participants).			510 (2 RCTs)	⊕⊕⊝⊝ Lowa,c
Challenging behaviour	No study assessed this outcome.			(0 studies)	‐
Number of people with adverse events	No study assessed this outcome.			(0 studies)	‐
Number of people with serious adverse events	No study assessed this outcome.			(0 studies)	‐
Quality of life	No study assessed this outcome.			(0 studies)	‐
Performance of activities of daily living	No study assessed this outcome.			(0 studies)	‐
*The risk in the intervention group (and its 95% confidence interval) is based on the assumed risk in the comparison group and the relative effect of the intervention (and its 95% CI). CI: confidence interval; MD: mean difference
GRADE Working Group grades of evidence High certainty: we are very confident that the true effect lies close to that of the estimate of the effect. Moderate certainty: we are moderately confident in the effect estimate: the true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different. Low certainty: our confidence in the effect estimate is limited: the true effect may be substantially different from the estimate of the effect. Very low certainty: we have very little confidence in the effect estimate: the true effect is likely to be substantially different from the estimate of effect.
See interactive version of this table: https://gdt.gradepro.org/presentations/#/isof/isof_question_revman_web_422034393400007384.

^a Downgraded one level for risk of bias: high risk of detection bias and unclear risk of selection bias, performance bias and selective reporting.
^b Downgraded one level for imprecision: CI indicates an effect both in the direction of the intervention and the control group.
^c Downgraded one level for inconsistency: non‐overlapping CI of two studies

Background

Description of the condition

Dementia is a syndrome characterised by progressive cognitive and functional decline. About 24 million people are affected by dementia worldwide, among them approximately six million people in Europe, and the number is expected to rise due to demographic change (Prince 2013; Prince 2015; Wittchen 2011). In long‐term care facilities, the prevalence of dementia ranges from 50% to 80% (Hoffmann 2014a; Seitz 2010; Stewart 2014).

People with dementia in nursing homes often experience acute or chronic pain (Rajkumar 2017; Takai 2010). An epidemiological study in seven European countries and one non‐European country (Services and Health for the Elderly in Long Term Care (SHELTER) Study) found an overall prevalence of pain in nursing home residents of 48.4% (1900 residents with pain out of 3926 residents) and approximately 46% of the sample had a diagnosis of dementia. Chronic pain was found in 12% of all residents with pain (Lukas 2013a).

Although guidelines for caring for people with dementia recommend adequate pain treatment (e.g. NICE 2016; RNAO 2013), people with dementia are often under‐treated for pain (Hemmingsson 2018). In the SHELTER study, 24% of the total sample of residents experiencing pain did not receive any pain medication; however, a great variation between different countries was found, ranging from 3.4% of residents without any pain medication in Finland to 69.5% in Italy (Lukas 2013b). Other studies reported differences in pain treatment between people with comparable diagnoses with and without cognitive impairment (Monroe 2014; Tan 2015).

Assessing pain in people with cognitive impairment is challenging. Self‐report is considered to be the gold standard for assessing pain since pain is a subjective experience (RNAO 2013). People with mild and moderate dementia are often able to express their pain verbally or by using simple visual or numerical pain intensity scales. In people with severe dementia, the use of self‐rating scales is often not feasible due to cognitive decline and the loss of communication skills. Proxy‐rating instruments have been developed to assess pain based on observation of verbal and nonverbal (such as facial) expressions, and physical behavior (Achterberg 2020; Lichtner 2014). However, there is limited evidence about the psychometric properties of these instruments, the level of pain cannot be assessed accurately by proxy‐rating, and it is hardly possible to distinguish between acute and chronic pain (Achterberg 2020; Lichtner 2014). Further barriers of a comprehensive pain assessment in people with dementia are a lack of knowledge and skills of healthcare professionals performing the assessment (Knopp‐Sihota 2019). Therefore, a structured pain assessment is often not well implemented in nursing homes and nursing staff, hampering pain management (Pringle 2021).

The experience of pain has a significant impact on well‐being and peoples’ quality of life (Helvik 2021), and is associated with negative health outcomes (Hurt 2008; Rajkumar 2017). Pain might also be one cause of challenging behaviour (Rajkumar 2017; van Dalen‐Kok 2015). Challenging behaviour leads to further stress and reduced quality of life for both people with dementia and their caregivers, and increases the burden of care (Feast 2016; Hurt 2008). While there is increasing evidence about an overuse of analgesics, especially opioids, in people with dementia living in nursing homes, undertreatment of pain is still prevalent (Hemmingsson 2018; Jensen‐Dahm 2020; La Frenais 2018; La Frenais 2021). An adequate treatment of pain is recommended to be standard of care (Schofield 2018), and might also be one approach to reduce challenging behaviour (NICE 2016).

Description of the intervention

There are various non‐pharmacological and pharmacological approaches to the treatment of acute and chronic pain. Pain treatment is complex in people with cognitive impairment living in nursing homes. Their experience of pain is often not recognised by health care professionals because no regular pain assessment is performed, and because of the residents' difficulties in articulating clearly whether or not they experience pain. Therefore, people with dementia receive less pain medication than people without dementia.

Algorithm‐based pain management aims to overcome these challenges. It usually starts with pain assessment and, if pain is present, offers pain treatment in the form of a structured, stepwise intervention. Each step defines a specific pain treatment and dose (or dose range). If the treatment is not effective, the treatment defined in the next step is applied. Algorithms often comprise both non‐pharmacological and pharmacological pain treatments. If non‐pharmacological pain treatments are included, they are often administered prior to pharmacological treatments — that is, in earlier steps of the algorithm.

How the intervention might work

People with dementia often exhibit specific behaviours, such as aggression or agitation. In this review, we use a broad definition of challenging behaviour, also including behaviours commonly referred to as behavioral and psychological symptoms of dementia. Untreated acute and chronic pain has been identified as one possible reason for challenging behaviour (Rajkumar 2017; van Dalen‐Kok 2015); other possible causes are brain changes and unmet needs (Cohen‐Mansfield 2010). Discriminating between challenging behaviour caused by pain and by other factors is often difficult.

A structured pain management plan, for example based on an algorithm, could be one approach to reducing pain and challenging behaviour caused by pain. Algorithm‐based interventions represent a pragmatic approach, addressing the complexity of pain management in people with dementia by using a stepwise approach with different treatments that can effectively reduce pain.

Algorithms comprise several steps, and for each step a specific pain treatment is defined. Some algorithms include both non‐pharmacological and pharmacological treatments (e.g. Chen 2016); others may consist of pharmacological treatments only. Guidelines recommend non‐pharmacological approaches as the first choice for managing pain in people with dementia (NICE 2016; RNAO 2013). Pharmacological treatment steps often start with non‐opioid analgesics, such as non‐steroidal anti‐inflammatory drugs (NSAIDs), followed by opioids, either alone or in combination with non‐opioids to maximise the efficacy of the treatment and to minimise adverse effects of the medication (RNAO 2013). However, algorithms might also specify different sequences of pain medications. In the study by Chen 2016, the pain medications defined for the different steps of the algorithm followed the recommendations of the American Geriatrics Society (American Geriatric Society Panel 2002). Since non‐pharmacological approaches are often used as the first choice, an algorithm‐based pain management might also reduce the overuse of pain medication and opioids.

To implement an algorithm, the treatment defined in the first step is applied to the person with dementia. The effect of the treatment is regularly evaluated, for example two or three times per day. If there is no treatment effect, the treatment defined in the next step is applied and evaluated. This procedure continues until the treatment is successful or until the last step of the algorithm is applied.

Given the high prevalence of undetected or under‐treated pain in nursing homes, using such a pragmatic approach in residents with challenging behaviour might reduce both pain and challenging behaviour.

Why it is important to do this review

Two systematic reviews have investigated the effects of pain treatment on challenging behaviour: Husebo 2011a included three studies with several methodological limitations and found inconsistent effects; Pieper 2013 included 16 publications evaluating a wide range of interventions and using different study designs (experimental and observational studies). Most of the included studies did not distinguish between acute and chronic pain. The results of the systematic review by Pieper 2013 suggest that pain treatment might be effective in reducing challenging behaviour, but tailored pharmacological approaches seem to be more effective than fixed treatment regimens. Due to the methodological and clinical heterogeneity of the studies and interventions, the internal validity and the generalisability of the results are limited. Both reviews recommended conducting further studies using rigorous study designs (Husebo 2011a; Pieper 2013).

Since these reviews were published, several randomised controlled trials (RCTs) investigating algorithm‐based interventions have been published (e.g. Chen 2016). A systematic review is warranted to describe these interventions and to summarise their effects on acute and chronic pain and challenging behaviour.

Algorithm‐based interventions are complex, including different components and treatment options (Craig 2008; Skivington 2021). To assess the effects of complex interventions, a description of the interventions' characteristics, i.e. aim and theoretical basis, and their components is required to ensure the comparability of the interventions and to draw clear conclusions for clinical practice (Burford 2013). In this review, the description of the interventions' characteristics was guided by the Criteria for Reporting the Development and Evaluation of Complex Interventions in healthcare 2 (CReDECI 2; Möhler 2015) and the Template for Intervention Description and Replication (TiDieR) guideline (Hoffmann 2014b). The effects of complex interventions are also influenced by the fidelity of implementation and process‐related outcomes (e.g. degree of implementation and response of the target group). Therefore, we also included implementation‐related information in this review.

The results of this review can help to improve the quality of care and pain treatment in people with dementia, and reveal knowledge gaps which can inform further research in the field.

Objectives

• To assess the effects of algorithm‐based pain management interventions to reduce pain and challenging behaviour in people with dementia living in nursing homes.

• To describe the components of the interventions and the content of the algorithms.

Methods

Criteria for considering studies for this review

Types of studies

As described in the review protocol (Labonté 2019), we included randomised controlled trials, individually or cluster‐randomised, investigating the effects of algorithm‐based pain management interventions for reducing pain and challenging behaviour in people with dementia. These designs were chosen since they are considered the 'gold standard' to assess the effects of interventions. There were no language restrictions.

Types of participants

We included people with dementia or cognitive impairment living in long‐term care facilities. There were no restrictions regarding the stage of dementia or cognitive impairment. We also included studies that had participants without dementia if the proportion of these participants was less than 20% of the total study population.

Types of interventions

We included all interventions offering pain treatment based on an algorithm, that is a set of unambiguous instructions explaining how to apply a pre‐determined, stepped approach to pain management. All interventions had to include the following elements.

• An initial pain assessment

• A treatment algorithm: a stepwise treatment plan that comprises at least two treatment steps to reduce pain. Algorithms might include different pharmacological and non‐pharmacological treatments, i.e. different types of pain medications, acupuncture or other non‐pharmacological interventions, to reduce pain. For all treatments, dose or dose range and frequency of application or delivery had to be defined.

• Criteria to assess the success of each treatment step, including an assessment method and a predefined response threshold

We included interventions using terms other than 'algorithm' – such as 'decision tree' or 'clinical pathway' – if they met the above‐defined criteria.

We excluded interventions offering only one specific non‐pharmacological or pharmacological treatment.

Comparison: we included control groups receiving usual care (standard pain assessment and treatment in the participants' care setting) or an active control intervention (i.e. other non‐pharmacological or pharmacological treatments for reducing pain or challenging behaviour not based on an algorithm).

Types of outcome measures

All included studies should have reported outcomes assessed by validated and reliable instruments.

Primary outcomes

• Pain‐related outcomes, e.g. the number of participants with pain, mean change in pain intensity, number of participants with at least 50% improvement in pain intensity

• Challenging behaviour

• Number of people with serious adverse events

To assess the presence of pain, we included both self‐ and proxy‐rating scales. We considered results assessed by self‐rating, for example verbal rating scales (VRS), as the most valid rating. We also included results assessed by proxy‐rating in this review, since people with dementia living in long‐term care institutions often have moderate to severe cognitive impairment and may not be able to use self‐assessment instruments.

Challenging behaviour could include agitation or, if no agitation scale is used, overall behavioural and psychological symptoms. It might be assessed with different validated instruments, for example the Cohen‐Mansfield Agitation Inventory (CMAI) or the Neuropsychiatric Inventory (NPI) (Cohen‐Mansfield 1989; Cummings 1994).

We expected a variety of instruments to be used and did not define a minimal clinically important difference for pain or challenging behaviour (see Data synthesis).

We defined serious adverse events as life‐threatening events or events requiring hospitalisation.

Secondary outcomes

• Quality of life, assessed by e.g. EuroQol (EQ‐5D) or Dementia‐Related Quality of Life (DEMQOL).

• Performance of activities of daily living, including mobility, assessed by appropriate validated instruments.

• Depression, assessed by appropriate validated instruments, e.g. the Cornell Scale for Depression in Dementia (CSDD).

• Number of people experiencing adverse events, e.g. sedation, constipation, nausea

• Mortality

• Effect on the caregivers, including caregivers' distress (assessed by e.g. Neuropsychiatric Inventory Caregiver Distress Scale (NPI‐D)), burden (assessed by e.g. the Zarit Burden Interview) or quality of life (assessed by e.g. EQ‐5D).

• Intervention costs

• Implementation‐related outcomes, e.g. implementation fidelity

Search methods for identification of studies

Electronic searches

We searched ALOIS (alois.medsci.ox.ac.uk), which is the Cochrane Dementia and Cognitive Improvement Group’s (CDCIG) specialised register. The last search was performed on 30 June 2021. Further information about the literature searches can be seen in Appendix 1.

ALOIS is maintained by the Information Specialists for the CDCIG, and contains studies that fall within the areas of dementia prevention, dementia treatment and management, and cognitive enhancement in healthy elderly populations. The studies are identified through:

• searching a number of major healthcare databases — MEDLINE, Embase, CINAHL (Cumulative Index to Nursing and Allied Health Literature) and PsycINFO;

• searching a number of trial registers — ClinicalTrials.gov and the World Health Organization’s International Clinical Trials Registry Platform (ICTRP) which covers ISRCTN; the Chinese Clinical Trials Register; the German Clinical Trials Register; the Iranian Registry of Clinical Trials; and the Netherlands National Trials Register, plus others;

• searching the Cochrane Library's Central Register of Controlled Trials (CENTRAL);

• searching grey literature sources: ISI Web of Science Core Collection.

To view a list of all sources searched for ALOIS, please follow this link to the ALOIS website (alois.medsci.ox.ac.uk).

Details of the search strategies run in healthcare bibliographic databases and used for the retrieval of reports of dementia, cognitive improvement and cognitive enhancement trials can be viewed on the CDCIG’s website: dementia.cochrane.org/searches

We ran additional searches in MEDLINE, Embase, PsycINFO, CINAHL, LILACS (Latin American and Caribbean Health Science Information database), ClinicalTrials.gov and the WHO Portal/ICTRP to ensure that the searches for this review were as comprehensive and as up to date as possible. The search strategy that we used can be seen in Appendix 1.

Searching other resources

We checked the reference lists of included studies and relevant reviews, and we performed forward citation tracking for all included studies (using Google Scholar). Additionally, we contacted study authors and experts in the field to identify unpublished and additional ongoing studies.

Data collection and analysis

Selection of studies

We used Cochrane's Screen4Me workflow to help assess the search results. Screen4Me comprises three components: known assessments – a service that matches records in the search results to records that have already been screened in Cochrane Crowd and been labelled as an RCT or as Not an RCT; the RCT classifier – a machine learning model that distinguishes RCTs from non‐RCTs, and if appropriate, Cochrane Crowd – Cochrane’s citizen science platform where the Crowd helps to identify and describe health evidence.

For more information about Screen4Me and the evaluations that have been done, please go to the Screen4Me webpage on the Cochrane Information Specialist’s portal (community.cochrane.org/organizational-info/resources/resources-groups/information-specialists-portal). In addition, more detailed information regarding evaluations of the Screen4Me components can be found in the following publications: Marshall 2018; McDonald 2017; Noel‐Storr 2018; Thomas 2017.

After the results had been through the Screen4Me workflow, we independently assessed the remainder. We used Covidence for study selection (Covidence). Two reviewers (CM, VL) independently screened all titles, abstracts or full publications against the inclusion criteria to identify potentially relevant studies. We resolved disagreement by discussion or, if necessary, by consulting a third reviewer (RM).

Data extraction and management

Two reviewers (CM, VL) extracted all relevant data independently using Covidence (Covidence). We checked extracted data for accuracy. In case of disagreement, we consulted a third reviewer (RM) to reach consensus.

For each study, we extracted the following data: information about prospective trial registration or a published study protocol or both; study design; characteristics of participants; baseline data; length of follow‐up; outcome measures; study results; and adverse effects. For cluster‐randomised trials, we also extracted estimates of the intra‐cluster correlation coefficient (ICC) if possible.

For each intervention, we extracted the relevant information for complex interventions based on the CReDECI 2 (Möhler 2015) and TiDieR checklists (Hoffmann 2014b): theoretical base and characteristics of the algorithm (i.e. method of pain assessment, number of steps, treatment options defined in each step, rules for switching to the next step of the algorithm, procedures to evaluate the treatment effect), implementation strategy, characteristics of the control group, information about implementation fidelity (e.g. the dose delivered and received, adherence to the algorithm, deviations from the protocol, and the frequency and dose of each treatment option used with the participants) and barriers and facilitators to the implementation of the interventions.

If any of the above information was missing, we sought it from study authors.

Assessment of risk of bias in included studies

We assessed risk of bias using the Cochrane risk of bias tool, as described in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2017). Two reviewers (CM, VL) independently assessed the methodological quality of the included studies in order to identify any potential sources of bias. We addressed the following domains, including design‐related criteria for cluster‐randomised trials: selection bias (generation of the randomisation sequence, concealed allocation of participants or clusters), performance bias (blinding of personnel and participants), detection bias (blinding of outcome assessment), attrition bias (attrition of clusters, incomplete reporting of outcome data); reporting bias (selective reporting) and other bias. We resolved any disagreements through reaching a consensus with a third reviewer (RM).

Measures of treatment effect

For dichotomous data, we calculated risk ratios (RR) with 95% confidence intervals (CI), if possible. For continuous outcome data assessed with the same rating scale, we calculated the mean difference (MD) with 95% CI. One study reported the change from baseline for pain outcomes (Ersek 2016a) and we used these data in the analysis. We performed all statistical analyses using RevMan Web (Review Manager 2022).

We did not define a minimal clinically important difference for pain or challenging behaviour because we could not identify generally accepted thresholds for clinical importance in the literature.

We planned to describe the implementation fidelity in each study as a percentage of the extent to which the participants received the treatments defined in the algorithm. Since these data were not available for the induced studies we described the implementation fidelity narratively, i.e. by comparing the available study data (e.g. MD with standard deviation (SD)) at baseline and follow‐up narratively (Campbell 2020).

Unit of analysis issues

For the included cluster‐randomised trials, we checked for unit of analysis issues. All included studies randomised participants in clusters but performed the analysis at the individual level; this might have led to a unit of analysis error. None of the studies reported an ICC, and none of the study authors provided information about the ICC. We did not use an external estimate of the ICC since we did not identify sufficiently similar studies with an ICC, and we did not perform any recalculation of the effect estimates.

Dealing with missing data

We described the amount of, and reasons for, missing data due to participant dropout in the Characteristics of included studies table. Where other information was missing, we contacted the study authors and asked for additional information.

Assessment of heterogeneity

We assessed clinical heterogeneity of interventions, i.e. differences in the algorithms used (see Types of interventions). In case of any clinical differences, two authors discussed whether the interventions were sufficiently similar in clinical and methodological characteristics to be included in a meta‐analysis. We planned to assess statistical heterogeneity by calculating the I² and Chi² statistics, but we did not perform any meta‐analyses in the review.

Assessment of reporting biases

We assessed selective reporting bias by comparing information about planned trials (e.g. identified in study registers and from conference abstracts) against the publications of included studies. We did not explore reporting bias with a funnel plot because of the small number of included studies.

Data synthesis

We planned to perform meta‐analyses using a random‐effects model if possible. We did not perform meta‐analyses for self‐rated pain since we included only two studies for this outcome and pain intensity at baseline was heterogeneous. For proxy‐rated pain, we did not perform a meta‐analysis because of the differences in pain intensity at baseline between the studies and the way in which the study results were reported. The two studies used different instruments and one study reported change from baseline (Ersek 2016a), whilst the other study reported means and SDs at follow‐up (Liu 2017). It was not possible to calculate the standardised mean difference (SMD) using these data. Only one study assessed challenging behaviour. We presented the results of these studies in narrative form, i.e. MD and RR if available, or the study means with SD (Campbell 2020).

Subgroup analysis and investigation of heterogeneity

We did not conduct subgroup analyses due to the small number of included studies.

Sensitivity analysis

We did not conduct sensitivity analyses due to the small number of included studies.

Summary of findings and assessment of the certainty of the evidence

We assessed the certainty of evidence for the most important outcomes using the GRADE method, judging study limitations (RoB as described above), inconsistency of the study results, indirectness of the evidence, imprecision of the results, and risk of publication bias (Guyatt 2011). Two reviewers (RM, CM) made GRADE ratings independently and resolved any disagreements by discussion or, if necessary, by consulting a third reviewer (VL).

We used GRADEpro GDT software to prepare a summary of findings table for the effects of algorithm‐based pain management for people with dementia and mild to moderate pain. The table included the following outcomes: pain intensity (self‐ and proxy‐rated), challenging behaviour, number of people with serious adverse events, number of people with adverse events, quality of life, and performance of activities of daily living.

Results

Description of studies

Results of the search

We identified a total of 4886 records (see Figure 1). After deduplication and an initial assessment by Screen4Me, two reviewers screened title and abstracts of 4011 records against the inclusion and exclusion criteria. Of these, 47 records were potentially eligible and were screened in full‐text. Three studies (reported in seven publications) met the inclusion criteria and were included in the review (Chen 2016; Ersek 2016a; Liu 2017).

1.

Figure 1: Study flow diagram

Included studies

We contacted the corresponding authors of all studies and asked for additional information on methodological details that were not reported in the publications (we sent one reminder to all non‐responding authors). We contacted one study author twice (the first request was sent during the study selection process). All study authors responded and offered additional information, but the author contacted twice did not respond to the second request.

All studies were cluster‐randomised controlled trials with nursing homes (Ersek 2016a; Liu 2017) or special care units (Chen 2016) as the unit of allocation. The duration of follow‐up ranged from three months (Chen 2016) to six months (Ersek 2016a).

For further information about the included studies see Characteristics of included studies.

Setting and Participants

One study was conducted in the USA (Ersek 2016a), one in Hong Kong (Liu 2017), and one in Taiwan (Chen 2016). The number of nursing homes per study was 17 (Liu 2017) and 27 (Ersek 2016a), respectively; the study by Chen 2016 included six special dementia care units.

The number of participants included in the studies ranged from 128 (Liu 2017) to 485 (Ersek 2016a), with a total of 808 participants in all studies. Of these, 673 participants completed the studies, with a range from 119 (Liu 2017) to 384 (Ersek 2016a) per study. The mean age of the participants was between 82 and 89 years. Most participants in the studies were female (56.6% to 83.6%).

Dementia was an inclusion criterion in two of the studies (Chen 2016; Liu 2017), and the participants had severe cognitive impairment (assessed by the Mini‐Mental State Examination (MMSE)) with mean scores of 8.3 and 6.7 (Chen 2016), and 4.5 and 2.3 (Liu 2017) in the intervention and control groups, respectively. There was no cognitive inclusion criterion in the Ersek 2016a study, but the participants had mild to moderate impairment based on assessment with the Cognitive Performance Scale, with mean scores of 2.29 (intervention group) and 2.71 (control group), respectively (range 0 to 6, with higher scores indicating more severe cognitive impairment).

Two studies assessed care dependency using the Barthel Index. In the study by Liu 2017, the participants had a mean score of 17.63 points in the control group and 25.0 points in the intervention group (range 0 to 100, higher scores indicating greater independence). In the study by Chen 2016, the mean score of the participants was 44.47 points.

Two studies defined pain‐specific inclusion criteria: Ersek 2016a intended to include participants with moderate to severe pain, and the participants in the study by Liu 2017 had to have at least one pain‐related diagnosis or pain condition.

The level of pain in the study samples varied. In the Chen 2016 study, all residents were included; less than half of the participants experienced pain and the mean pain intensity score in both study groups (self‐ and proxy rating) at baseline indicated no, or almost no, pain. In the other two studies, the participants had mild to moderate pain (Ersek 2016a, self‐ and proxy‐rated, despite the specified inclusion criterion of moderate to severe pain; and Liu 2017, proxy‐rated).

Description of interventions

The description of the included interventions uses categories relevant for complex interventions (Hoffmann 2014a; Möhler 2015).

Theoretical basis and components of the interventions

The algorithms were developed based on various recommendations, such as clinical guidelines (e.g. from the American Medical Directors Association, American Geriatric Society or British Pain Society and the British Geriatric Society) (Chen 2016; Ersek 2016a; Liu 2017), or those of an Interdisciplinary Expert Consensus Statement (Chen 2016; Liu 2017). In addition, two studies included previous literature reviews in the development of the algorithms (Chen 2016; Ersek 2016a). All algorithms were reviewed and revised several times by a panel of experts in pain management, gerontology and dementia (Chen 2016; Ersek 2016a; Liu 2017).

The interventions differed in structure and in the number of steps and decision trees. We describe the three algorithms in detail below.

Pain Recognition and Treatment Protocol

Chen 2016 used the Pain Recognition and Treatment Protocol (PRT) and offered pain education to registered nurses. The PRT consisted of the four steps listed below.

1. Primary pain assessment: used for the detection and recognition of pain, including a self‐report of pain from the residents, the observation of residents' nonverbal expressions of pain (according to the recommendations of the American Geriatrics Society panel), a physical examination to identify potential causes of pain and information from other healthcare providers or family members.

2. Secondary pain assessment: a comprehensive assessment of the characteristics of pain (presence, location, type, intensity and frequency of pain, frequency of unusual behaviours, "things" that improve pain), psychosocial comorbidities (e.g. psychological well‐being, interpersonal interaction) and a summary of the characteristics and causes of the residents’ pain and its impact on their lives.

3. Pain treatment: including the treatment of underlying causes of pain using pharmacological and non‐pharmacological treatments as well as consultations and referrals with other healthcare professionals, if needed. No further information about the non‐pharmacological and pharmacological treatments applied was reported.

4. Reassessment: re‐evaluation and monitoring the treatment effects in regular intervals, observation of side effects of pain treatment, and adjustment of the treatment, if necessary.

Pain management algorithm

A pain management algorithm (ALG) was used in the study by Ersek 2016a. This ALG consisted of 11 evidence‐based decision trees, some of them linked with other decision trees (Ersek 2016b).

1. General (initial) pain assessment: focussing on location, intensity, possible causes, duration, character of pain, effects on e.g. physical function, mood, social activities, factors that influence the pain, current treatments. Depending on the character and intensity of pain, the algorithm refers to different decision trees. The intervals of the reassessments are based on pain history (e.g. weekly reassessments for recently controlled pain, monthly reassessments for pain controlled for more than four weeks), pain treatment (e.g. 20 to 30 minutes after short‐acting medication) and pain intensity (e.g. daily reassessment for moderate to severe pain).

2. Managing pain in nonverbal residents (for residents not able to give self‐report): including a pain assessment based on residents' history (e.g. diagnosis), current health status and pain‐related behaviours as well as treatments affecting pain and behaviours (e.g. addressing participants' basic needs).

3. Pain treatment: five decision‐trees addressing the treatment of pain with different pharmacological treatment options each (such as acetaminophen, non‐steroidal anti‐inflammatory drugs and opioids).

4. Management of medication side effects: three decision trees addressing the management of medication side effects, such as constipation, sedation and delirium.

Observational pain management protocol

An observational pain management protocol was applied in Liu 2017. It consisted of five steps:

1. Pain assessment: at least once per day during potentially pain‐inducing care procedures or exercises using the Chinese version of Pain Assessment in Advanced Dementia Scale (PAINAD‐C) (Peng 2007).

2. Verifying scores: including the assessment of possible causes of pain (e.g. injury or a pain‐related diagnosis) by asking simple yes/no questions or obtaining information from the nursing staff or others; if no possible causes of pain are identified, the aim is to interpret the meaning of the behaviour (by consultation with caregivers familiar with the residents, if needed) and to ensure that basic needs are met.

3. Interpreting scores of the PAINAD‐C: a score lower than 2 indicates no pain, 2 to 3 indicates mild pain and scores of 4 or higher indicate moderate pain or severe pain.

4. Pain treatment based on the PAINAD‐C score: for pain scores of 2 to 4, strategies in the delivery of care are recommended to reduce pain (e.g. timely warnings before movement, recommendations for painless mobilisation in bed, transferring as well as seating) (Talerico 2006). If the PAINAD‐C score is not reduced or if the score is higher than 4, non‐pharmacological (e.g. heat therapy, cold therapy, transcutaneous electrical nerve stimulation (TENS), massage etc. tailored to the needs of each participant) or pharmacological treatments are used.

5. Evaluation and monitoring to assess the effectiveness of the pain treatment using the PAINAD‐C; if the pain score decreases, monitoring is continued, and if pain‐related behaviour persists, the interventions are modified.

Liu 2017 tested the intervention in a pilot study. The other studies did not provide information on feasibility or pilot tests (Chen 2016; Ersek 2016a).

Delivery of the intervention

All studies included educational components to train the users and to facilitate the implementation of the algorithms.

In the study by Chen 2016, registered nurses received six hours of pain education. This consisted of three two‐hour educational sessions with an introduction to pain assessment and treatment strategies, and the barriers to pain management in older adults. In addition, registered nurses received three hours of instruction and one hour of practical training on the application of the algorithm. This training included case examples to demonstrate the use of the algorithm, and real cases were also discussed. All educational sessions were carried out by the same researcher based on a protocol (Chen 2016). A manual with information about the algorithms was available to the nursing staff and could also be used by staff who did not attend the educational sessions. Chen 2016 also offered a telephone consulting service from the research team, if necessary, to discuss challenging cases in pain assessment and treatment. In order to improve the use of the algorithm, the registered nurses additionally received 30 minutes on‐site teaching every month for four months, starting four weeks after the initial intervention (Chen 2016).

In the study by Ersek 2016a, nursing staff received four educational sessions with information covering the different decision trees included in the algorithm (as described above), with discussions of case examples and real cases. A manual with information about the algorithm was available for nursing staff. The educational sessions were also recorded on video, e.g. for nursing staff who did not attend the educational sessions or for future viewing. Additional written materials were available for the participating nursing homes (e.g. with additional pain‐related information to support nursing staff, administrators and primary care providers). Ersek 2016a used Rogers' Diffusion of Innovations Theory to support adoption of the algorithm and evidence‐based pain practice (Rogers 2003). The implementation strategy comprised feedback, implementation of interdisciplinary pain management teams and clinical champions and chart forms, as well as policies to implement the algorithms into clinical practice, and four bi‐weekly booster activities that started eight weeks after the educational session. These activities included items (e.g. pens and magnets) with pain‐related logos, posters and sheets with a short description of the educational sessions' content. The research team also offered telephone counselling for the members of the pain team, discussing challenges of specific cases or implementation barriers (Ersek 2012; Ersek 2016a).

Liu 2017 offered four one‐hour educational sessions to the nursing staff addressing pain assessment, management of older people with dementia and the use and implementation of the algorithm. Separate educational sessions were offered for nursing staff and non‐professional caregivers (not specified) in order to adapt the content to the different educational backgrounds (Liu 2014). In the study by Liu 2017 a steering committee was established to supervise the implementation of the algorithm, consisting of members of the research team and a physiotherapist or a nurse‐in‐charge from each facility. To support the implementation of the algorithm, daily to weekly visits were made by a trained research assistant during the implementation period in each cluster and quality‐control meetings were held monthly to bi‐monthly to evaluate the implementation of the algorithm (Liu 2017).

Characteristics of the control conditions

Two studies offered pain education as an active control condition (Chen 2016; Ersek 2016a).

Chen 2016 offered the same pain education in the control group as in the intervention group (performed by the same researcher using the same manual); 83% of the nursing staff attended the educational sessions. The telephone consulting service from the research team was also available to the control group. Ersek 2016a offered four one‐hour educational sessions on the basic principles of pain assessment and management for older adults to licensed nursing staff in each control facility. No further information was reported.

In the study by Liu 2017, no intervention was offered in the control group and the nursing homes continued with their usual pain management, comprising a pain assessment without a specific assessment tool and, in case of pain, a treatment based on nurses' experience and knowledge. No further information was reported.

Outcomes and methods of data collection

Primary outcomes

Two studies assessed pain intensity via self‐rating by the study participants (Chen 2016; Ersek 2016a). Chen 2016 used the Verbal Descriptor Scale (VDS) (Closs 2004), with scores ranging from 0 to 3 (higher scores indicate more severe pain). The VDS scores were collected on seven consecutive days. Ersek 2016a assessed the usual and worst pain using the Iowa Pain Thermometer (Herr 2007). The Iowa Pain Thermometer is a modified verbal descriptor scale with scores ranging from 0 (no pain) to 12 (most intense pain imaginable).

All studies used proxy‐ratings to assess pain intensity. Two studies (Chen 2016; Liu 2017) used the PAINAD‐C, comprising five key observational indicators of pain behaviour: breathing, negative vocalisation, facial expression, body language and consolability (Peng 2007). The severity for each behaviour is rated from 0 to 2 and a total score is calculated from all behaviours (range: 0 to 10 points; higher scores indicate more severe pain). In the study by Chen 2016 research assistants assessed the PAINAD‐C on seven consecutive days, observing the participants from 8:00 a.m. to 5:30 p.m. Ersek 2016a used the Iowa Pain Thermometer (Herr 2007) for the proxy‐rating of usual and worst pain based on the observations of the nursing staff during the previous week (range 0 to 12 points, with higher scores indicating more severe pain). Liu 2017 used multiple imputation of missing data for the pain proxy‐rating.

Only one study assessed challenging behaviour (Chen 2016), and measured it using the Cohen‐Mansfield Agitation Inventory, Chinese version (CMAI‐C) (Lin 2007). CMAI‐C is a proxy‐rating instrument to assess the frequency of 29 possible agitated behaviours in older adults with dementia. Each behaviour is rated on a 7‐point scale (1 = never to 7 = several times an hour). The research assistant assessed the behaviours based on observations on seven consecutive days.

Secondary outcomes

All studies assessed data about the implementation of the interventions, but none of the studies assessed the other secondary outcomes of this review.

We included the following implementation‐related data in the review:

We used data about the number of pain assessments and the delivery of non‐pharmacological and pharmacological pain treatments to describe implementation fidelity, since all interventions aimed to improve pain assessment and the delivery such treatments. We also used data assessing nurses' attendance at the educational sessions and the correct use of the algorithms to describe implementation fidelity.

Two studies assessed the success of the educational session (Chen 2016; Liu 2017). In the study by Chen 2016, all nurses had to complete two real cases to check whether the steps of the algorithm were performed correctly. At least 90% agreement between the researcher and the nurse was required, otherwise the training and the test had to be repeated (Chen 2016). Liu 2017 used a checklist to verify that the facilities had followed the algorithm.

Chen 2016 and Liu 2017 assessed the use of pharmacological and non‐pharmacological treatments. Liu 2017 used multiple imputation of missing data for these outcomes. In the study by Chen 2016, the total number of pharmacological and non‐pharmacological treatments and the total number of referrals with other healthcare professionals per week (step 3 of the intervention) were collected by research assistants based on the documentation of the previous week. Liu 2017 assessed pharmacological treatments (pain medication and psychotropic agent) using the Medication Quantification Scale version III (MQS III, Harden 2005). Based on the MQS III, the residents' pain medications were quantified according to the dosage, pharmacological classification, and detriment weight (i.e. the potential of each medication to produce side effects). Liu 2017 also assessed the type and frequency of non‐pharmacological treatments from the documentation of the past seven days.

Ersek 2016a did not distinguish between the different types of treatment but between the fidelity of the assessments and total treatments offered. The study evaluated adherence to the procedures recommended in the evidence‐based decision‐trees in a 30‐day period using a self‐developed Pain Management Chart Audit Tool with 17 items, scored between 0 and 2 (0 = no adherence, 1 = partial adherence, 2 = full adherence, or n.a. for not applicable). A total score was calculated as a sum of all items divided by the total possible score (excluding items with not applicable ratings) multiplied by 100 (range from 0 to 100 points, higher scores indicate greater implementation fidelity).

Other implementation process‐related data

Ersek 2016a conducted focus groups in four out of 13 nursing homes in the intervention group in order to identify barriers and facilitators to the adoption of the algorithm (six months after the data collection of all study outcomes). A convenience sample of 24 nurses was recruited (17 registered nurses, three licensed practical nurses, one advanced practice registered nurse and two certified nursing assistants). A directed content analysis was conducted. Categories were developed by one researcher and validated by a second researcher (Ersek 2014; Ersek 2016a).

Excluded studies

We excluded 40 records after screening the full texts because the study design or the interventions did not meet the inclusion criteria, e.g. the algorithms did not include at least two different treatment steps or did not specifically focus on pain (for example, the serial trial intervention used by Kovach 2006c). See Characteristics of excluded studies.

Risk of bias in included studies

The methodological quality of the included studies was low. We judged the risk of bias to be unclear in three domains for all included studies. In addition, we judged one study to be at high risk of bias in one domain (see Characteristics of included studies; Figure 2; Figure 3).

2.

Figure 2: Risk of bias graph

3.

Figure 3: Risk of bias summary

Allocation

The Chen 2016 study adequately generated a random sequence. The study by Liu 2017 generated the sequence adequately, but the researcher allocated one of the 17 nursing homes to the control group (not at random). Accordingly, we judged the risk of bias for this domain to be unclear. The study by Ersek 2016a did not report any information about the method of sequence generation, and we considered the risk of bias in this domain to be unclear.

Group allocation was adequately concealed in one study (Ersek 2016a) and unclear in two studies (Chen 2016; Liu 2017).

One study identified participants in each cluster before randomisation (Ersek 2016a). In one study, participants of most clusters were identified before randomisation but probably not in the single cluster allocated by the researchers (Liu 2017), and in one study no information was reported (Chen 2016); we rated the risk of bias to be unclear.

Blinding

Blinding of personnel was not possible in all studies due to the nature of the intervention; however, we have insufficient information to judge whether this led to a high risk of bias.

Participants were blinded to group allocation in one study (Ersek 2016a), in one study blinding of participants was unclear (Chen 2016), and in one study participants were not blinded to group allocation (Liu 2017). However, the intervention was delivered to the nursing staff and not directly to the participants in all studies. We have insufficient information to judge whether the risk of bias was low or high and judged the risk of performance bias as unclear for all studies.

Outcomes were rated by the participants (self‐rated pain) in two studies. Participants in the Ersek 2016a study were blinded to group allocation, so we judged this to have a low risk of bias. We also judged the risk of bias to be low in the Chen 2016 study, since the control group also received pain education and the participants were not the target group of the intervention.

In all studies, some outcomes were also assessed by nursing staff or research assistants. Two studies blinded outcome assessors to group allocation (Chen 2016; Liu 2017) and we judged the risk of detection bias for proxy‐rated outcomes to be low. In the study by Ersek 2016a, the proxy‐rated outcomes were collected by nursing staff not blinded to group allocation, and we judged the risk of detection bias to be high.

Incomplete outcome data

No clusters were lost to follow up in any of the studies (Chen 2016; Ersek 2016a; Liu 2017). All studies reported the reasons for attrition; attrition rates were low and comparable between the study groups (Chen 2016; Ersek 2016a; Liu 2017).

Selective reporting

Two studies were registered (Ersek 2016a; Liu 2017), but only one prospectively (Liu 2017). A study protocol was published for only one study (Liu 2017).

In the study by Ersek 2016a, the pre‐planned primary outcome self‐rated pain was changed to proxy‐rated pain because of the low rate of participants who were able to complete the self‐rating, due to the progression of dementia. However, this might have not introduced a reporting bias, and we judged this to present an unclear risk of bias.

Liu 2017 mentioned the primary and secondary outcomes in the trial register, but in neither the study protocol nor the main publication. However, this might have not introduced a reporting bias. We judged this to have a low risk of selection bias.

Other potential sources of bias

We judged all studies to be at low risk of other bias.

Effects of interventions

See: Table 1

See Table 1 for people with mild to moderate pain. We did not prepare a summary of findings table for participants with no or almost no pain.

Primary outcomes

Pain‐related outcomes

Pain was assessed by self‐rating in two of the included studies (Chen 2016; Ersek 2016a), and by proxy‐rating in all studies (Chen 2016; Ersek 2016a; Liu 2017). All three studies reported data on pain intensity before and after the intervention or on change in pain intensity.

None of the studies reported outcome data on our other two pain‐related primary outcomes, namely the number of participants with pain, or the number of participants with at least 50% improvement in pain intensity scores. None of the studies distinguished between acute and chronic pain.

Pain intensity: self‐rating

Two studies reported pain intensity rated by the participants using different assessment instruments (Chen 2016; Ersek 2016a). In the study by Ersek 2016a only about two thirds of the participants were able to perform the self‐rating at baseline and follow‐up. In both studies, the control condition was pain education for nursing staff.

In the study by Chen 2016, less than half of the participants experienced pain at baseline; the mean pain intensity in both study groups indicated no pain or almost no pain at baseline (assessed by VDS, mean score: intervention group 0.42 ± 0.61, control group 0.63 ± 0.77; range 0 to 3). We found no evidence of a difference in pain intensity between the intervention and the control group at follow‐up (MD ‐0.27, 95% CI ‐0.49 to ‐0.05; range 0 to 3; 170 participants; Analysis 1.1). We found very low‐certainty evidence (downgraded one level for risk of bias and two levels for imprecision) and we are uncertain whether an algorithm‐based pain management intervention has an effect on self‐rated pain intensity in people with no pain or almost no pain compared with pain education.

1.1. Analysis.

Comparison 1: Algorithm‐based pain management vs control, Outcome 1: Pain (self‐rating)

In the study by Ersek 2016a, the participants had mild to moderate pain at baseline assessed by the Iowa Pain Thermometer. There was no evidence of a difference in pain intensity between the intervention and control groups for self‐rated pain after six months (MD 0.4, 95% CI ‐0.58 to 1.38; range 0 to 12; 246 participants; Analysis 1.1). We judged this to be low‐certainty evidence (downgraded one level for risk of bias and one level for imprecision) that an algorithm‐based pain management intervention may have little to no effect on self‐rated pain compared with pain education.

Ersek 2016a also reported residents' worst pain. This was reduced slightly in the control group that received pain education for nursing staff compared to the algorithm‐based pain management group (MD 0.9, 95% CI 0.31 to 1.49; 259 participants; Analysis 1.1).

Pain intensity: proxy‐rating

A pain intensity proxy‐rating was performed in all three studies. In two studies pain assessment was performed by members of the research team (Chen 2016; Liu 2017), and in Ersek 2016a by the nursing staff.

As described above, in the study by Chen 2016 less than half of the participants experienced pain at baseline and the mean pain intensity in both study groups indicated no pain, or almost no pain, at baseline. In the intervention group, pain was reduced slightly (baseline 1.31 ± 1.31, 3‐month follow‐up 0.48 ± 0.83; 88 participants) and was nearly unchanged in the control group (baseline 0.60 ± 0.90, 3‐month follow‐up 0.82 ± 1.09; PAINAD‐C, a score lower than 2 indicates no pain or almost no pain; 82 participants). However, the baseline values differed slightly between the study groups and the mean values in both study groups were below the threshold of pain requiring any treatment (< 2 points, PAINAD‐C) at baseline and follow‐up. We found very low‐certainty evidence (downgraded one level for risk of bias and two levels for imprecision) and we are uncertain whether an algorithm‐based pain management intervention has an effect on proxy‐rated pain intensity in people with no pain, or almost no pain, compared with pain education.

The participants in Ersek 2016a and Liu 2017 had mild to moderate pain at baseline (proxy‐rated). In the study by Ersek 2016a, there was no evidence of a difference in pain intensity between the intervention group and control group that received pain education after six months (MD ‐0.2, 95% CI ‐0.79 to 0.39; range 0 to 12; 382 participants; Analysis 1.2). Liu 2017 found a reduction of pain in the intervention group compared with the control group receiving usual care after 12 weeks (MD ‐1.49, 95% CI ‐2.11 to ‐0.87; PAINAD‐C, range 0 to 10; 128 participants; Analysis 1.2). We considered this to be low‐certainty evidence (downgraded one level for risk of bias and one level for inconsistency) that an algorithm‐based pain management intervention may reduce proxy‐rated pain in comparison with usual care but may not be more effective than pain education.

1.2. Analysis.

Comparison 1: Algorithm‐based pain management vs control, Outcome 2: Pain (proxy‐rating)

In the study by Ersek 2016a, there was also no evidence of a difference in the intensity of residents' worst pain (rated by proxy) between the intervention and control groups (MD ‐0.1, 95% CI ‐0.88 to 0.68; 300 participants; Analysis 1.2).

Challenging behaviour

Challenging behaviour was investigated in one study (Chen 2016) using the CMAI‐C. In this study, less than half of the participants experienced pain at baseline, and it remains unclear to which extent pain was a relevant cause of challenging behaviour. We found very low‐certainty evidence (downgraded one level each for risk of bias, indirectness and imprecision) and we are uncertain whether an algorithm‐based pain management intervention reduces challenging behaviour in participants with no pain, or almost no pain, at baseline compared with pain education after three months (MD ‐0.21, 95% CI ‐1.88 to 1.46; 1 study; 170 participants; range 7 to 203; Analysis 1.3).

1.3. Analysis.

Comparison 1: Algorithm‐based pain management vs control, Outcome 3: Challenging behaviour

Serious adverse effects

No study assessed serious adverse effects and no study provided information about serious adverse effects.

Secondary outcomes

No study assessed adverse effects, such as medication side effects, and no study described such events. No study assessed any of the other secondary outcomes of this review.

Implementation‐related data

Implementation fidelity

In the study by Chen 2016, 85% of the nursing staff attended the educational sessions in the intervention group. The main reasons for non‐attendance of nurses were lack of motivation, loss of leisure time, and conflict with the work schedule. In the study by Liu 2017, three of the eight nursing homes from the intervention group achieved the required 90% of correct items on the checklist in the first week of the implementation period, three facilities achieved this goal in the second week and two facilities in the third week. No relapse was observed during the entire duration of the study. The main questions that arose during the first two weeks of the implementation stage concerned the interpretation of the assessment scores and the timing of the re‐evaluation after the pain treatments.

Two studies assessed the use of non‐pharmacological and pharmacological treatments (Chen 2016; Liu 2017). In the study by Chen 2016, the number of weekly non‐pharmacological treatments for pain relief increased in both study groups after the intervention period and returned to the baseline level at follow‐up (intervention group: baseline 0.21 ± 0.61, postintervention 0.63 ± 1.19, follow‐up 0.27 ± 0.96; control group: baseline 0.13 ± 0.56, postintervention 0.24 ± 0.76, follow‐up 0.16 ± 0.53). In the study by Liu 2017, the frequency of non‐pharmacological treatments increased in the intervention group (baseline 0.29 ± 0.96, follow‐up 9.09 ± 5.34) and increased slightly in the control group (baseline 0.25 ± 0.59, follow‐up 0.43 ± 1.07). The number of different types of non‐pharmacological treatments used increased in both study groups (intervention group: baseline 0.32 ± 0.74, follow‐up 1.11 ± 0.77; control group: baseline 0.38 ± 0.82, follow‐up 0.61 ± 0.91).

The number of weekly pharmacological treatments for pain relief in the study by Chen 2016 increased after the intervention period and returned to the baseline level at follow‐up (baseline 0.63 ± 0.94, post‐intervention 0.78 ± 2.31, follow‐up 0.64 ± 0.93) and increased in the control group during the study period (baseline 0.35 ± 0.67, post‐intervention 0.47 ± 0.89, follow‐up 0.71 ± 1.19). In the study by Liu 2017 the MQS III score for pain medication was almost unchanged in both groups (intervention group: baseline 5.16 ± 13.28, follow‐up 5.13 ± 13.31; control group: baseline 8.27 ± 16.64, follow‐up 8.76 ± 16.66). For psychotropic medication the score increased in the intervention group (baseline 20.75 ± 16.58, follow‐up 27.25 ± 28.21) and slightly increased in the control group (baseline 16.72 ± 28.33, follow‐up 18.94 ± 30.03).

Chen 2016 assessed the number of weekly referrals and found nearly no change in both study groups (intervention group: baseline 0.12 ± 0.52, follow‐up 0.20 ± 0.56; control group: baseline 0, follow‐up 0.11 ± 0.37).

In the study by Ersek 2016a, the assessment score (representing the number of performed assessments that were defined in the algorithm) increased in the intervention group (baseline 32.2 ± 2.3, postintervention 35.7 ± 2.2, follow‐up 39.9 ± 3.2) and slightly increased in the control group (baseline 27.5 ± 3.2, postintervention 29.8 ± 3.1, follow‐up 28.3 ± 3.3). The treatment score (representing the number of delivered non‐pharmacological and pharmacological treatments) was nearly unchanged in both study groups (intervention group: baseline 66.4 ± 2.1, postintervention 67.9 ± 2.4, follow‐up 66.6 ± 2.0; control group: baseline 65.0 ± 1.3, postintervention 66.2 ± 1.3, follow‐up 66.3 ± 1.8).

Barriers and facilitators to implementation

Ersek 2016a identified several structural and process barriers to the implementation of the intervention: staff and leadership turnover, high resident‐to‐staff ratio, government regulations (e.g. limited access to controlled substances), lack of time, physicians' negative attitudes about nurses' pain management skills, and fears of addiction and over‐sedation from staff, family and residents. The identified structural and process facilitators of implementation were: a strong and supportive leadership, involvement of different target groups in the pain assessment (i.e. nursing staff and other professionals), development and implementation of new policies and procedures to guide clinical practice, availability of necessary resources (e.g. education, forms, experts), consistency of staff, and government regulations (e.g. publicly‐reported quality indicators) (Ersek 2012; Ersek 2014; Ersek 2016a).

Discussion

Summary of main results

We included three trials that evaluated an algorithm‐based pain management intervention for people with dementia living in nursing homes. The algorithms varied in content and delivery, but the core components were comparable. The studies used different instruments to measure pain intensity: all studies used proxy‐rating instruments (two studies used the PAINAD‐C (Chen 2016; Liu 2017) and one study the Iowa Pain Thermometer (Ersek 2016a)), and two studies additionally used self‐rating instruments (Chen 2016; Ersek 2016a). Two of the studies offered pain education to the control group (Chen 2016; Ersek 2016a) and one study compared the intervention with usual care (Liu 2017).

In one study, fewer than half of the participants experienced pain, and the mean values of self‐rated and proxy‐rated pain in both study groups were below the threshold of pain that needs any treatment at baseline and follow‐up. We are uncertain whether an algorithm‐based pain management intervention has an effect on self‐rated and proxy‐rated pain intensity in people with no pain or almost no pain compared with pain education.

Only two studies investigated the effects of the intervention in participants with mild to moderate pain at baseline. One study compared the intervention with pain education and there was low‐certainty evidence that an algorithm‐based pain management intervention may have little to no effect on pain assessed both by self‐ and by proxy‐rating (Ersek 2016a). In the second study the control group received no intervention (usual care) and we found low‐certainty evidence that an algorithm‐based pain management intervention may reduce pain assessed by proxy‐rating. No detailed information about the pain management in the control group was reported, e.g. number residents with pain assessment or number of participates receiving any pain treatment.

The results on proxy‐rated pain are difficult to interpret. The validity of proxy‐rated pain is questioned (Zwakhalen 2018), and the study by Leong 2006 found a weak correlation between two different proxy‐rating instruments and self‐rated pain in people with dementia. We found some differences between the self‐ratings and proxy‐ratings of pain at baseline: in the study by Chen 2016, the pain scores assessed by proxy differed between the intervention and control group to some extent and were quite similar in the self‐rated pain scores; however, these differences could also be explained by chance. In the study by Ersek 2016a, proxy‐rated pain scores were higher than self‐rated pain scores in both study groups (assessed by the same instrument), but there was no clear difference in the proxy‐ and self‐rated pain scores between the study groups and only two thirds of the participants completed the self‐assessment of pain. In addition, the direction of the effect differed between self‐ and proxy‐rated pain (Ersek 2016a). Also, the responsiveness of the proxy‐rating pain assessment instruments, i.e. the ability to detect changes over time after a pain treatment is applied, is unclear (Husebo 2016).

Another reason for the heterogeneous effects might be the degree of implementation of the interventions. Chen 2016 found a slight increase in use of non‐pharmacological and pharmacological treatments after the intervention, but the values returned to the baseline level at follow‐up. In the study by Ersek 2016a, more assessments had been performed in both study groups, but the use of treatments did not increase considerably. Only Liu 2017 observed a greater use of non‐pharmacological pain treatments in the intervention group, but not for pain medications. This indicates that the algorithms might not be implemented as intended in the included studies. However, based on the different instruments used to assess the use of non‐pharmacological and pharmacological treatments, the results should be interpreted with caution.

Only one study assessed challenging behaviour. Based on the large number of participants without pain and the low mean pain intensity in this study, we are uncertain whether an algorithm‐based pain management intervention reduces challenging behaviour (very low‐certainty evidence).

None of the studies assessed serious adverse effects and no information about the occurrence of any serious adverse effects was reported in the included studies. The other outcomes of interest were not addressed in the included studies.

In summary, there is no clear evidence for a benefit of an algorithm‐based pain management intervention in comparison with pain education for reducing pain intensity in people with dementia in nursing homes, but the intervention may reduce proxy‐rated pain compared with usual care. The number of studies and the number of participants per study was small, the certainty of evidence was low, and the implementation fidelity seems to be limited. Pain treatment is a multidisciplinary challenge, but the interventions in this review mainly targeted the nursing staff. Only the intervention tested by Ersek 2016a described a multidisciplinary perspective and comprised the formation of a multidisciplinary pain management team. However, partly negative attitudes of physicians regarding nurses' pain management skills were described as barrier in this study.

Overall completeness and applicability of evidence

The number of studies contributing to the outcomes of interest is small. There was a pronounced variation in the participants' pain levels and cognitive function, and we found some differences in self‐ and proxy‐ratings of pain in one study (as described above). Only one study assessed challenging behaviour, although pain is considered as one reason for agitation and aggression. No information about adverse events was assessed or reported in the studies. Based on the low‐certainty evidence and the limited implementation fidelity of the algorithms, the results of this review must be interpreted with caution.

Quality of the evidence

The certainty of evidence was very low or low due to the heterogeneity of the participants' pain intensity and the effects of the different studies, the small number of studies for each comparison, and the small number of participants in the included studies. The methodological quality of the included studies varied. Selection bias was mainly rated as unclear (the method of sequence generation and allocation concealment were each unclear in two studies). Blinding of personnel was not possible due to the nature of the interventions and in one study the outcome assessors for proxy‐rated pain were not blinded to group allocation (Ersek 2016a). There was a high risk of a unit‐of‐analysis error since none of the studies considered the cluster effect in the analyses. We judged the implementation fidelity of the interventions to be limited, and we found some evidence that the use of the algorithms did not lead to a greater use of non‐pharmacological or pharmacological treatments compared to pain education.

Potential biases in the review process

We made several efforts to reduce the risk of bias in the review process. Guided by the Cochrane Dementia and Cognitive Improvement Group, we conducted an intensive literature search, including relevant databases and trial registers, and we performed additional searches, e.g. backward and forward citation tracking for all included studies. Two review authors independently selected the studies, assessed the methodological quality, and extracted data. For any missing information, we contacted the study authors.

Based on the small number of studies, we were not able to investigate the risk of publication bias.

None of the included cluster‐randomised trials accounted for the cluster‐design in the analysis, and an ICC was not available for any of the studies. We did not perform a cluster‐appropriate analysis since we did not identify an appropriate ICCC. There is a high risk of a unit‐of‐analysis error.

Agreements and disagreements with other studies or reviews

We have not found any previous systematic review investigating the effects of algorithm‐based pain management.

Several studies have investigated different types of pain management intervention for people with dementia in nursing homes. A cluster‐RCT by Rostad 2018 investigated whether the regular use of a pain assessment tool twice a week for 12 weeks improved the delivery of pain treatment and reduced pain. The study did not find an increase in analgesic use or a reduction of pain.

One study investigated the effects of a treatment protocol comprising different types of pain medications on challenging behaviour (Husebo 2011). Participants were allocated to one of the treatment regimens, and no changes were possible during the follow‐up period. The study found that agitation and pain were reduced after eight weeks. However, the treatment protocol with a fixed treatment regimen did not fulfil the inclusion criteria of this review.

Another algorithm‐based intervention that aims to improve challenging behaviour in people with dementia in nursing homes is the serial trial intervention (Kovach 2006c). This intervention comprises several steps with assessments and treatments for common causes of challenging behaviour, such as different basic care needs and psychosocial and environmental needs. The intervention did not fulfil the inclusion criteria of this review because pain is only one aspect among others. A systematic review on this intervention found that this intervention is not more beneficial than education about addressing challenging behaviour for reducing challenging behaviour or pain (Manietta 2021).

Authors' conclusions

Implications for practice.

The results of this review suggest that there is no clear evidence for a benefit of an algorithm‐based pain management intervention in comparison with pain education for reducing pain intensity in people with dementia in nursing homes, but the intervention may reduce proxy‐rated pain compared with usual care. From an ethical perspective, an adequate pain treatment for people with dementia should be standard of care. The experience of pain is also associated with negative health outcomes (Hurt 2008; Rajkumar 2017). An adequate pain assessment for people with dementia and, in case of pain, an individual pain treatment with non‐pharmacological interventions or pain medication should be offered to all residents in nursing homes as usual care, as recommended by several guidelines (NICE 2016; Sirsch 2020). We also found no clear evidence that algorithm‐based pain management reduces challenging behaviour, but only one study investigated this outcome in a population with mild or no pain.

Implications for research.

Although the rationale for using an algorithm‐based approach to pain in dementia seems to be reasonable, the available evidence does not support the hypothesis that using an algorithm either improves the assessment or treatment of pain in people with dementia living in nursing homes. There is a need for further studies investigating the effectiveness of different approaches to implementing comprehensive pain management (including the use of an appropriate method for assessing pain in people with dementia, and adequate pain treatment with non‐pharmacological interventions or pain medication, if necessary) and to overcoming common implementation barriers (Brunkert 2020; De Witt Jansen 2017; Knopp‐Sihota 2019).

Further research should focus on the feasibility of interventions and strategies to implement the algorithm based on established frameworks, for example the Medical Research Council framework for the development and evaluation of complex interventions (Craig 2008; Skivington 2021), and implementation frameworks such as Reach and Efficacy ‐ Adoption, Implementation, and Maintenance (RE‐AIM) (Glasgow 1999), or the Consolidated Framework For Implementation Research (CFIR) (Damschroder 2009). More emphasis should be laid on techniques and strategies to increase the impact of the algorithms in the decision‐making process about pain management in clinical practice and to promote interdisciplinary co‐operation and communication.

There is also a need for well‐planned, prospectively registered and sufficiently powered randomised controlled trials that follow established methodological standards, such as using a concealed allocation, adequate blinding of participants and outcome assessors, and the use of an active control group (i.e. education about state of the art in pain assessment and management for people with dementia). Pain assessment should generally be based on a self‐rating, if possible. If the participants are not able to assess pain themselves, proxy‐rating instruments can be used, but raters have to be trained in use of the instruments to increase the validity of the proxy‐rating. If challenging behaviour will be investigated in addition to pain‐related outcomes, validated instruments should be used. Adverse events of non‐pharmacological treatments and pain medication should also be included as important outcomes. The studies also have to include a comprehensive process evaluation (Moore 2015; Skivington 2021).

To improve reporting quality, available guidelines for reporting of interventions' development and characteristics should be used (e.g. Hoffmann 2014b; Möhler 2015), as well as design‐related guidelines (EQUATOR network 2021).

History

Protocol first published: Issue 5, 2019

Appendices

Appendix 1. Sources searched and search strategies

Source	Search strategy	Hits retrieved
1. CENTRAL (The Cochrane Library) crso.cochrane.org/SearchSimple.php (Date of most recent search: 30 June 2021)	#1 MESH DESCRIPTOR Dementia EXPLODE ALL TREES 5075 #2 MESH DESCRIPTOR Delirium 509 #3 MESH DESCRIPTOR Wernicke Encephalopathy 4 #4 MESH DESCRIPTOR Neurocognitive Disorders 151 #5 dement*. 20176 #6 alzheimer*:TI,AB,KY 9913 #7 (lewy* adj2 bod*):TI,AB,KY 379 #8 (chronic adj2 cerebrovascular):TI,AB,KY 107 #9 ("organic brain disease" or "organic brain syndrome"):TI,AB,KY 133 #10 (benign senescent forgetfulness):TI,AB,KY 2 #11 (cerebr* adj2 deteriorat*):TI,AB,KY 10 #12 (cerebral* adj2 insufficient*):TI,AB,KY 1 #13 (major neurocognitive disorder*):TI,AB,KY 22 #14 #1 OR #2 OR #3 OR #4 OR #5 OR #6 OR #7 OR #8 OR #9 OR #10 OR #11 OR #12 OR #13 23788 #15 MESH DESCRIPTOR Residential Facilities EXPLODE ALL TREES 1610 #16 MESH DESCRIPTOR Halfway Houses EXPLODE ALL TREES 16 #17 MESH DESCRIPTOR Long‐Term Care EXPLODE ALL TREES 1089 #18 MESH DESCRIPTOR Homes for the Aged EXPLODE ALL TREES 575 #19 MESH DESCRIPTOR Nursing Homes EXPLODE ALL TREES 1241 #20 (homes for the aged):TI,AB,KY 596 #21 (residential facilit*):TI,AB,KY 210 #22 (nursing home*):TI,AB,KY 3517 #23 (care home*):TI,AB,KY 563 #24 ((care or nursing or residential or rest or old* people* or old folk* or group or geriatric or aged or elderly) adj2 (home or homes or facility or facilities)):TI,AB,KY 8685 #25 #15 OR #16 OR #17 OR #18 OR #19 OR #20 OR #21 OR #22 OR #23 OR #24 9670 #26 #14 AND #25 1722 #27 MESH DESCRIPTOR ANALGESICS EXPLODE ALL TREES 49185 #28 MESH DESCRIPTOR PAIN EXPLODE ALL TREES 43856 #29 MESH DESCRIPTOR Pain Management EXPLODE ALL TREES 2978 #30 MESH DESCRIPTOR NARCOTICS EXPLODE ALL TREES 16484 #31 pain*:TI,AB,KY 152628 #32 analges*:TI,AB,KY 53508 #33 #27 OR #28 OR #29 OR #30 OR #31 OR #32 195860 #34 #26 AND #33 167	June 2019: 167 March 2020: 27 June 2021: 31
2. MEDLINE In‐process and other non‐indexed citations and MEDLINE 1950‐present (Ovid SP) (Date of most recent search: 30 June 2021)	1 exp Dementia/ 2 Delirium/ 3 Wernicke Encephalopathy/ 4 Delirium, Dementia, Amnestic, Cognitive Disorders/ 5 dement*.mp. 6 alzheimer*.mp. 7 (lewy* adj2 bod*).mp. 8 (chronic adj2 cerebrovascular).mp. 9 ("organic brain disease" or "organic brain syndrome").mp. 10 "benign senescent forgetfulness".mp. 11 (cerebr* adj2 deteriorat*).mp. 12 (cerebral* adj2 insufficient*).mp. 13 "major neurocognitive disorder*".ti,ab. 14 or/1‐13 15 exp Residential Facilities/ 16 exp Halfway Houses/ 17 exp Long‐Term Care/ 18 exp Homes for the Aged/ 19 exp Nursing Homes/ 20 "homes for the aged".ti,ab. 21 "residential facilit*".ti,ab. 22 "nursing home*".ti,ab. 23 "care home*".ti,ab. 24 ((care or nursing or residential or rest or old* people* or old folk* or group or geriatric or aged or elderly) adj2 (home or homes or facility or facilities)).ti,ab. 25 or/15‐24 26 14 and 25 27 exp ANALGESICS/ 28 exp PAIN/ 29 exp Pain Management/ 30 exp NARCOTICS/ 31 pain*.ti,ab. 32 analges*.ti,ab. 33 or/27‐32 34 26 and 33 35 randomized controlled trial.pt. 36 controlled clinical trial.pt. 37 randomized.ab. 38 placebo.ab. 39 drug therapy.fs. 40 randomly.ab. 41 trial.ab. 42 groups.ab. 43 or/35‐42 44 exp animals/ not humans.sh. 45 43 not 44 46 34 and 45	June 2019: 275 March 2020: 24 June 2021: 44
3. Embase (Ovid SP) 1974 to 26 June 2019 (Date of most recent search: 30 June 2021)	1 Dementia/ 2 Delirium/ 3 Wernicke Encephalopathy/ 4 Delirium, Dementia, Amnestic, Cognitive Disorders/ 5 ("benign senescent forgetfulness" or ("normal pressure hydrocephalus" and "shunt*") or ("organic brain disease" or "organic brain syndrome") or ((cerebral* or cerebrovascular or cerebro‐vascular) adj2 insufficien*) or (cerebr* adj2 deteriorat*) or (chronic adj2 (cerebrovascular or cerebro‐vascular)) or (creutzfeldt or jcd or cjd) or (lewy* adj2 bod*) or (pick* adj2 disease) or alzheimer* or binswanger* or deliri* or dement* or huntington* or korsako*).tw. 6 "major neurocognitive disorder".ti,ab. 7 or/1‐6 8 exp Residential Facilities/ 9 exp Halfway Houses/ 10 exp Long‐Term Care/ 11 exp Homes for the Aged/ 12 "homes for the aged".ti,ab. 13 "residential facilit*".ti,ab. 14 "nursing home*".ti,ab. 15 "care home*".ti,ab. 16 ((care or nursing or residential or rest or old* people* or old folk* or group or geriatric or aged or elderly) adj2 (home or homes or facility or facilities)).ti,ab. 17 or/8‐16 18 7 and 17 19 exp analgesic agent/ 20 exp pain/ 21 exp narcotic agent/ 22 pain*.ti,ab. 23 analges*.ti,ab. 24 or/19‐23 25 18 and 24 26 randomized controlled trial/ 27 controlled clinical trial/ 28 random$.ti,ab. 29 randomization/ 30 intermethod comparison/ 31 placebo.ti,ab. 32 (compare or compared or comparison).ti. 33 ((evaluated or evaluate or evaluating or assessed or assess) and (compare or compared or comparing or comparison)).ab. 34 (open adj label).ti,ab. 35 ((double or single or doubly or singly) adj (blind or blinded or blindly)).ti,ab. 36 double blind procedure/ 37 parallel group$1.ti,ab. 38 (crossover or cross over).ti,ab. 39 ((assign$ or match or matched or allocation) adj5 (alternate or group$1 or intervention$1 or patient$1 or subject$1 or participant$1)).ti,ab. 40 (assigned or allocated).ti,ab. 41 (controlled adj7 (study or design or trial)).ti,ab. 42 (volunteer or volunteers).ti,ab. 43 trial.ti. 44 or/26‐43 45 25 and 44 46 from 45 keep 1001‐1233	June 2019: 1233 March 2020: 158 June 2021: 258
4. PsycINFO (Ovid SP) (Date of most recent search: 30 June 2021)	1 exp Dementia/ 2 exp Delirium/ 3 exp Huntingtons Disease/ 4 exp Kluver Bucy Syndrome/ 5 exp Wernickes Syndrome/ 6 exp Cognitive Impairment/ 7 dement*.mp. 8 alzheimer*.mp. 9 (lewy* adj2 bod*).mp. 10 deliri*.mp. 11 (chronic adj2 cerebrovascular).mp. 12 ("organic brain disease" or "organic brain syndrome").mp. 13 "supranuclear palsy".mp. 14 ("normal pressure hydrocephalus" and "shunt*").mp. 15 "benign senescent forgetfulness".mp. 16 (cerebr* adj2 deteriorat*).mp. 17 (cerebral* adj2 insufficient*).mp. 18 (pick* adj2 disease).mp. 19 (creutzfeldt or jcd or cjd).mp. 20 huntington*.mp. 21 binswanger*.mp. 22 korsako*.mp. 23 ("parkinson* disease dementia" or PDD or "parkinson* dementia").mp. 24 "major neurocognitive disorder".ti,ab. 25 or/1‐24 26 exp Residential Care Institutions/ 27 exp Halfway Houses/ 28 exp Long Term Care/ 29 exp Residential Care Institutions/ 30 exp Nursing Homes/ 31 "homes for the aged".ti,ab. 32 "residential facilit*".ti,ab. 33 "nursing home*".ti,ab. 34 "care home*".ti,ab. 35 ((care or nursing or residential or rest or old* people* or old folk* or group or geriatric or aged or elderly) adj2 (home or homes or facility or facilities)).ti,ab. 36 or/26‐35 37 25 and 36 38 exp Analgesia/ 39 exp Pain/ 40 exp Analgesic Drugs/ 41 exp Pain Management/ 42 exp Narcotic Drugs/ 43 pain*.ti,ab. 44 analges*.ti,ab. 45 or/26‐44 46 25 and 45 47 exp Clinical Trials/ 48 randomly.ab. 49 randomi?ed.ti,ab. 50 placebo.ti,ab. 51 groups.ab. 52 "double‐blind*".ti,ab. 53 "single‐blind*".ti,ab. 54 RCT.ti,ab. 55 or/47‐54 56 46 and 55 57 from 56 keep 1001‐1853	June 2019: 1853 March 2020: 118 June 2021: 185
5. CINAHL (EBSCOhost) (Date of most recent search: 30 June 2021)	S53 S39 AND S52 S52 S40 OR S41 OR S42 OR S43 OR S44 OR S45 OR S46 OR S47 OR S48 OR S49 OR S50 OR S51 S51 MH "Random Assignment" S50 MH "Single‐Blind Studies" or MH "Double‐Blind Studies" or MH "Triple‐Blind Studies" S49 MH "Crossover Design" S48 MH "Factorial Design" S47 MH "Placebos" S46 MH "Clinical Trials" S45 TX "multi‐centre study" OR "multi‐center study" OR "multicentre study" OR "multicenter study" OR "multi‐site study" S44 TX crossover OR "cross‐over" S43 AB placebo* S42 TX random* S41 TX trial* S40 TX "latin square" S39 S31 AND S38 S38 S32 OR S33 OR S34 OR S35 OR S36 OR S37 S37 TX analges* S36 TX pain* S35 (MH "Pain Management") S34 (MH "Analgesics, Opioid+") S33 (MH "Pain+") S32 (MH "Analgesia+") S31 S20 AND S30 S30 S21 OR S22 OR S23 OR S24 OR S25 OR S26 OR S27 OR S28 OR S29 S29 TX (care or nursing or residential or rest or old* people* or old folk* or group or geriatric or aged or elderly) N2 (home or homes or facility or facilities) S28 TX care home* S27 TX nursing home* S26 TX residential facilit* S25 TX homes for the aged S24 (MH "Nursing Homes+") S23 (MH "Long Term Care") S22 (MH "Halfway Houses") S21 (MH "Residential Care+") S20 S1 OR S2 OR S3 OR S4 OR S5 OR S6 OR S7 OR S8 OR S9 OR S10 OR S11 OR S12 OR S13 OR S14 OR S15 OR S16 OR S17 OR S18 OR S19 S19 TX "major neurocognitive disorder" S18 TX korsako* S17 TX binswanger* S16 TX huntington* S15 TX creutzfeldt or jcd or cjd S14 TX pick* N2 disease S13 TX cerebral* N2 insufficient* S12 TX "benign senescent forgetfulness" S11 TX "benign senescent forgetfulness" S10 TX "normal pressure hydrocephalus" and "shunt*" S9 TX "organic brain disease" or "organic brain syndrome" S8 TX chronic N2 cerebrovascular S7 TX deliri* S6 TX lewy* N2 bod* S5 TX alzheimer* S4 TX dement* S3 MH "Wernicke's Encephalopathy" S2 (MH "Delirium") or (MH "Delirium, Dementia, Amnestic, Cognitive Disorders") S1 MH "Dementia+"	June 2019: 229 March 2020: 32 June 2021: 12
6. Web of Science – core collection (ISI Wed of Science) (Date of most recent search: 30 June 2021)	TOPIC: (dement* OR alzheimer* OR "vascular cognitive impairment" OR "lew* bod*" OR CADASIL OR "cognit* impair*" OR FTD OF FTLD OR "cerebrovascular insufficienc*" OR AD OR VCI) ANDTOPIC: (Residential Facilities OR Halfway Houses OR Long‐Term Care OR Homes for the Aged OR Nursing Home* OR "care home*") AND TOPIC:(analges* OR pain* OR NARCOTICS)AND TOPIC: (randomly OR randomised OR randomized OR "random allocat*" OR RCT OR CCT OR "double blind*" OR "single blind*" OR "double blind*" OR "single blind*" OR trial)	June 2019: 305 March 2020: 373 June 2021: 41
7. LILACS (BIREME) (Date of most recent search: 30 June 2021)	Residential Facilities OR Halfway Houses OR Long‐Term Care OR Homes for the Aged OR nursing homes AND dementia OR alzheimers OR cognition OR cognitive AND ANALGESICS or PAIN or NARCOTICS	June 2019: 2 March 2020: 1 June 2021: 0
8. ClinicalTrials.gov (www.clinicaltrials.gov) (Date of most recent search: 30 June 2021)	Residential Facilities OR Halfway Houses OR Long‐Term Care OR Homes for the Aged OR nursing homes | Interventional Studies | dementia OR alzheimers OR cognition OR cognitive | ANALGESICS or PAIN or NARCOTICS	June 2019: 9 March 2020: 9 June 2021: 0
9. ICTRP (Date of most recent search: 30 June 2021)	Residential Facilities OR Halfway Houses OR Long‐Term Care OR Homes for the Aged OR nursing homes AND dementia OR alzheimers OR cognition OR cognitive AND ANALGESICS or PAIN or NARCOTICS	June 2019: 4 March 2020: 0 June 2021: 0
9. ALOIS (Cochrane Dementia and Cognitive Improvement Group Register) (CRS web) (Date of most recent search: 30 June 2021)	#1 pain INREGISTER #2 (long‐term care OR care home* OR residential home* OR nursing home*) INREGISTER #3 #1 AND #2	June 2019: 74 March 2020: 63 June 2021: 15
TOTAL before deduplication		June 2019: 4151 March 2020: 805 June 2021: 586
TOTAL after deduplication		June 2019:3428 March 2020: 735 June 2021: 491

Data and analyses

Comparison 1. Algorithm‐based pain management vs control.

Outcome or subgroup title	No. of studies	No. of participants	Statistical method	Effect size
1.1 Pain (self‐rating)	2		Mean Difference (IV, Random, 95% CI)	Totals not selected
1.1.1 VDS scale	1		Mean Difference (IV, Random, 95% CI)	Totals not selected
1.1.2 Iowa Pain Thermometer	1		Mean Difference (IV, Random, 95% CI)	Totals not selected
1.1.3 Iowa Pain Thermometer ‐ worst pain	1		Mean Difference (IV, Random, 95% CI)	Totals not selected
1.2 Pain (proxy‐rating)	2		Mean Difference (IV, Random, 95% CI)	Totals not selected
1.2.1 Iowa Pain Thermometer	1		Mean Difference (IV, Random, 95% CI)	Totals not selected
1.2.2 PAINAD‐C	1		Mean Difference (IV, Random, 95% CI)	Totals not selected
1.2.3 Iowa Pain Thermometer ‐ worst pain	1		Mean Difference (IV, Random, 95% CI)	Totals not selected
1.3 Challenging behaviour	1		Mean Difference (IV, Random, 95% CI)	Totals not selected

Characteristics of studies

Characteristics of included studies [ordered by study ID]

Chen 2016.

Study characteristics
Methods	Study design: cluster‐randomised controlled trial; not registered Duration of follow‐up: 3 months Study grouping: parallel group Total study duration (start/end date): September 2012 to November 2013
Participants	Setting: 6 dementia special care units located in Northern and Central Taiwan (3 clusters per group) Inclusion criteria: all residents of participating clusters with dementia were eligible to participate. Exclusion criteria: no information Participants/clusters Number of participants, clusters randomised: overall participants n = 195, clusters n = 6; IG participants n = 98, clusters n = 3; CG participants n = 97, clusters n = 3 Number of participants, clusters at follow‐up: overall participants n = 170; IG participants n = 88; CG participants n = 82 Loss to follow‐up/dropouts (reasons): overall participants n = 25; IG n = 10 (admission to hospital n = 5, died n = 5); CG n = 15 (discharge home n = 4, admission to hospital n = 10, relocation to other facility n = 1) Baseline Characteristics Age (mean ± SD): IG 82.7 ± 8.2, CG 83.6 ± 7.2 Gender (female): IG n = 49, 50.0%, CG n = 59 (60.8%) Cognitive status/dementia diagnosis: MMSE (mean ± SD) IG 8.3 ± 8.2, CG 6.7 ± 6.9; MMSE ≤ 10: IG n = 68 (69.4%), CG = 75 (78.9%) Care dependency, ADL performance: Barthel Index (mean ± SD) IG 50.9 ± 37.2, CG 46.0 ± 33.9 Pain (self‐rating): VDS (mean ± SD) IG 0.51 ± 0.78 (n = 98), CG 0.42 ± 0.61 (n = 97) Pain (proxy‐rating): PAINAD‐C (mean ± SD) IG 1.31 ± 1.31 (n = 98), CG 0.60 ± 0.90 (n = 97) Pain history (information about pain in the past): IG n = 48 (49.5%), CG n = 37 (38.1%) Challenging behaviour: CMAI‐C (mean ± SD) IG 35.5 ± 6.36, CG 35.61 ± 6.57 Pretreatment: no statistically significant differences in any demographic data between the experimental and control groups, except for educational level and average number of medications.
Interventions	Intervention: Pain Recognition and Treatment Protocol (PRT) (Step 1. primary pain assessment, Step 2. secondary pain assessment (comprehensive assessment), Step 3. pain treatment, Step 4. reassessment at regular intervals) Control: Pain education (an introduction to pain assessment and management strategies for older adults, and barriers to pain management).
Outcomes	Primary: Pain (self‐rating) (pain intensity via Verbal Descriptor Scale) Pain (proxy‐rating) (pain related behaviours via Chinese‐Pain Assessment in Advanced Dementia) Pharmacological strategies (number of weekly pharmacologic strategies) Non‐pharmacological strategies (number of weekly non‐pharmacologic strategies) Number of referrals per week Secondary: Challenging behaviour (Cohen‐Mansfield Agitation Inventory Chinese version)
Identification	Sponsorship source: National Science Council (NSC 99‐2314‐B‐039‐019) of Taiwan Country: Taiwan, ROC Institution: Institute of Clinical Nursing, National Yang‐Ming University
Notes	Cluster effect was not incorporated in the analysis (high risk of a unit‐of‐analysis error).
Risk of bias
Bias	Authors' judgement	Support for judgement
Random sequence generation (selection bias)	Low risk	"(...) simple randomization procedures." Authors responded on request, that coin‐tossing was used (unpublished information).
Allocation concealment (selection bias)	Unclear risk	No information reported. Authors responded on out request: "each facility was blindly randomized", but did not provide information about the method of allocation concealment. It is also not clearly reported whether the participants were identified before randomisation of clusters.
Blinding of participants and personnel (performance bias) All outcomes	Unclear risk	Study was described as a "double‐blind cluster randomized trial", but no further information was reported. No further information was provided from the study authors on request.
Blinding of outcome assessment (detection bias) Self‐reported outcomes	Low risk	No information about blinding of participants reported, but since the intervention was delivered at the cluster level and the control group also received a basic education addressing pain, we judged the risk of detection bias to be low.
Blinding of outcome assessment (detection bias) Proxy‐rated outcomes	Low risk	"Research assistants (RAs) with Bachelor of Science or higher degrees with majors in psychology or nursing blinded to the study hypothesis and group allocation undertook data collection"; "The C‐PAINAD was completed before the VDS, so as not to bias the RAs by the verbal reports of pain."
Incomplete outcome data (attrition bias) All outcomes	Low risk	"At baseline, 195 residents participated in this study. Of these, 25 residents were lost to follow up at 3 months after the intervention because of death, admission to the hospital, discharge home, or relocation to another facility."
Selective reporting (reporting bias)	Unclear risk	No information about study registration or study protocol
Other bias	Low risk	‐

Ersek 2016a.

Study characteristics
Methods	Study design: cluster‐randomised controlled trial; trial was retrospectively registered (NCT01399567) Duration of follow‐up: 6 months Study grouping: parallel group Total study duration (start/end date): September 2006 to January 2010 (described in study register)
Participants	Setting: 27 nursing homes in the greater Puget Sound area, USA. Inclusion criteria: all residents of participating clusters aged ≥ 65 years, with moderate to severe pain, and expected to remain at the facility for at least 6 months were eligible to participate regardless of cognitive function. Exclusion criteria: no information Participants/clusters Number of participants, clusters randomised: overall participants n = 485, clusters n = 27; IG participants n = 259, clusters n = 13; CG participants n = 226, clusters n = 14 Number of participants, clusters at follow‐up: overall participants n = 385, clusters n = 27; IG participants n = 203, clusters n = 13; CG participants n = 182, clusters n = 14 Loss to follow‐up/dropouts (reasons): overall participants n = 101, clusters n = 0; IG participants n = 56 (death n = 46, moved/hospitalised n = 6, dropped out due to illness n = 4); CG participants n = 45 (death n = 36, moved/hospitalised n = 7, dropped out due to illness n = 2) Baseline characteristics Age (mean ± SD): IG 83.9 ± 8.3, CG 84.3 ± 7.6 Gender (female): IG n = 181 (70%), CG n = 158 (77%) Cognitive status/dementia diagnosis: Cognitive Performance Scale (range 0‐6) (mean ± SD) IG 2.29 ± 1.46, CG 2.71 ± 1.38 Care dependency, ADL performance: no information Pain (self‐rating): Iowa Pain Thermometer Scale (mean ± SD) IG 5.2 ± 2.4 (n = 198), CG 5.3 ± 2.1 (n = 150) Pain (proxy‐rating): Iowa Pain Thermometer Scale (mean ± SD) IG 3.4 ± 2.6 (n = 253), CG 3.1 ± 2.6 (n = 224) Pain history (information about pain in the past): no information Number of painful conditions (range, mean ± SD): overall range 0 to 7, IG 1.91 ± 1.46, CG 1.63 ± 1.30 Pain intensity: MDS 2.0 pain intensity score (possible range 0 to 3) (mean ± SD) IG 1.32 ± 1.02, CG 1.06 ± 1.06 MDS Pain Summary (MDS pain intensity x MDS pain frequency) (possible range 0 to 5): (mean ± SD) IG 2.29 ± 1.67, CG 1.78 ± 1.67 Agitation: Pittsburgh Agitation Scale (possible range 0 to 16) (mean ± SD) IG 1.78 ± 2.38, CG 1.99 ± 2.54 Pretreatment: the only statistically significant difference at baseline was the mean MDS (minimum data set) pain summary score. Nursing homes were diverse regarding the size, quality and ownership. There were no statistically significant differences between the facilities in the intervention group and the control group (number of beds, Centers for Medicare and Medicaid star rating, number of deficiencies or type of ownership).
Interventions	Intervention: pain management algorithm containing 11 linked evidence‐based decision trees Control: pain education (basic principles of pain assessment and management for older adults)
Outcomes	Primary: Pain (proxy‐rating) (pain intensity via Iowa Pain Thermometer Scale) Secondary: Pain (self‐rating) (pain intensity via Iowa Pain Thermometer Scale) Process outcome: Adherence to the Intervention (Pain Management Chart Audit Tool)
Identification	Sponsorship source: National Institute of Nursing Research (R01NR009100). Country: USA Institution: University of Pennsylvania School of Nursing
Notes	Cluster effect was not incorporated in the analysis (high risk of a unit‐of‐analysis error).
Risk of bias
Bias	Authors' judgement	Support for judgement
Random sequence generation (selection bias)	Unclear risk	"Facilities were randomized singly or in matched pairs, although the final 3 unmatched facilities were randomized simultaneously. For the first 18 facilities, pairs of facilities that were similar in size (110 beds or >110 beds), ownership (for profit or not for profit), and quality (based on number of deficiencies or stars on the 5‐star quality rating system) were matched and randomized (1 to treatment and 1 to control with equal chance of assignment). Six of the first 18 facilities were not paired and were randomized singly with an equal chance of assignment to either condition. The last 9 facilities were randomized with an adaptive randomization that set the probability of each possible assignment according to the resulting balance in the allocation of ALG versus EDU on key facility characteristics." No information about the method of sequence generation, but it was generated by a statistician.
Allocation concealment (selection bias)	Low risk	"Following collection of all baseline measures at a facility, the principal investigator (ME) contacted the statistician with the name or names of the facilities that were to be randomized along with the limited information that was necessary to monitor balance between ALG and EDU facilities."
Blinding of participants and personnel (performance bias) All outcomes	Unclear risk	"Residents (...) were blinded to intervention or control." Personnel were not blinded to group allocation, but we have insufficient information to permit judgement of ‘low risk’ or ‘high risk'.
Blinding of outcome assessment (detection bias) Self‐reported outcomes	Low risk	“Residents and data collectors were blinded to intervention or control, with the exception of 1 team member who completed the PM‐CAT at baseline in several facilities. However, blinded data collectors completed the PM‐CAT at follow‐up.”
Blinding of outcome assessment (detection bias) Proxy‐rated outcomes	High risk	No information about blinding of personnel was reported; staff were probably aware of group allocation.
Incomplete outcome data (attrition bias) All outcomes	Low risk	Participants lost to follow‐up and reasons were comparable in both groups.
Selective reporting (reporting bias)	Unclear risk	Trial was retrospectively registered. The preplanned primary outcome (self‐reported pain) was changed to proxy‐rated pain because of the large number of residents with moderate and severe dementia who were unable to give self‐report.
Other bias	Low risk	‐

Liu 2017.

Study characteristics
Methods	Study design: cluster‐randomised controlled trial; trial was prospectively registered (CUHK_CCT00367) Duration of follow‐up: 12 weeks Study grouping: parallel group Total study duration (start/end date): June 2014 to August 2015
Participants	Setting: 17 nursing homes in Hong Kong, in two major geographic districts (Kowloon and the New Territories) Inclusion criteria: all residents of participating clusters aged > 65 years, lived in the nursing home for > 6 months, with officially diagnosed dementia of some form, at least one pain‐related diagnosis or pain condition (confirmed by a physiotherapist working in the participant’s home), with substantial communication deficit judged by the nurses based on two items in the interRai‐Home Care (HC) Assessment were eligible to participate. Exclusion criteria: all residents of participating clusters were excluded who suffered from any acute physiological/psychiatric illness during the data collection period, experienced distressing social circumstances such as bereavement, which could have affected their daily behavioural patterns. Participants/clusters Number of participants, clusters randomised: overall participants n = 128, clusters n = 17; IG participants n = 64, clusters n = 8; CG participants n = 64, clusters n = 9 Number of participants, clusters at follow‐up: overall participants n = 119; IG participants n = 59, CG participants n = 60 Loss to follow‐up/dropouts (reasons): overall participants n = 9; IG participants n = 5 (died n = 5); CG participants n = 4 (died n = 4) Baseline Characteristics Age (mean ± SD): IG 89.30 ± 7.11, CG 87.87 ± 7.6 Gender (female): IG n = 61 (88.4%), CG n = 46 (70.8%) Cognitive status: Cantonese MMSE (range 0 ‐ 30) (mean ± SD) IG 2.33 ± 4.28, CG 4.50 ± 5.82 Care dependency, ADL performance: Chinese‐Modified Barthel Index (C‐MBI) (range 0 ‐ 100) (mean ± SD) IG 25 ± 26.76; CG 17.63 ± 26.59 Pain (proxy‐rating): PAINAD‐C (mean ± SD) IG 5.61 ± 1.68 (n = 64), CG 5.41 ± 2.01 (n = 64) Pain history (information about pain in the past): no information Number of pain‐related diagnoses (mean ± SD): IG 1.10 ± 0.79, CG 1.03 ± 0.81 Number of pain sites (mean ± SD): IG 1.41 ± 0.56, CG 1.22 ± 0.45 Number of prescribed pain medications (mean ± SD): IG 1.23 ± 1.11, CG 1.22 ± 1.04 Pretreatment: no significant differences in the majority of demographic characteristics and baseline measurements between the experimental and control groups, except for gender and cognitive status.
Interventions	Intervention: observational pain management protocol (Step 1. assessing pain, Step 2a verifying scores, 2b verifying ‐ no evidence indicates pain, Step 3. interpreting scores, Step 4. using treatments, Step 5. evaluating and continued monitoring) Control: usual pain management strategies (based on the clinical assessment of the staff according to their experience and knowledge about the residents)
Outcomes	Primary and secondary outcomes were specified in the study register only (not in the study protocol or the other publications) as follows: Primary: Non‐pharmacological strategies (types and frequencies) Pharmacological strategies (pain medication and psychotropic agents via Medication Quantification Scale version III) Secondary: Pain (proxy‐rating) (pain related behaviour via Chinese‐Pain Assessment in Advanced Dementia) The sample size for the study was calculated based on the pain score.
Identification	Sponsorship source: ResearchGrants Council of the Hong Kong University Grants Committee under the 2013/14 Early Career Scheme (Ref no.559213) Country: Hong Kong Institution: School of Nursing, The Hong Kong Polytechnic University
Notes	Cluster effect was not incorporated in the analysis (high risk of a unit‐of‐analysis error).
Risk of bias
Bias	Authors' judgement	Support for judgement
Random sequence generation (selection bias)	Unclear risk	"16 homes accepted the invitation and were randomized to the control or experimental conditions according to the pre‐set randomization list generalized by an independent biostatistician. As several control homes had a relatively small number of participants who met the criteria for inclusion, one more control home was recruited to meet the required sample size." Unpublished information from the study authors: sequence was computer‐generated by an independent biostatistician, but one nursing home was recruited after randomisation and assigned to the control group.
Allocation concealment (selection bias)	Unclear risk	"The group allocations were sealed in opaque envelopes, which were not opened to release the group assignment until the baseline assessments for all of the participants in each home were completed." One nursing home was recruited after randomisation and assigned to the control group.
Blinding of participants and personnel (performance bias) All outcomes	Unclear risk	Participants and personnel were not blinded to group allocation (not possible), insufficient information to permit judgement of ‘low risk’ or ‘high risk'.
Blinding of outcome assessment (detection bias) Proxy‐rated outcomes	Low risk	"The participants’ exhibition of pain‐related behaviors was observed based on the C‐PAINAD in weeks 8 and 16 by an independent pain observer, who was a senior nursing student blinded to the study’s hypotheses, the group allocation, and the total consumption of pain treatments of all of the participants."
Incomplete outcome data (attrition bias) All outcomes	Low risk	"All analyses were performed according to intention‐to‐treat principles. To manage the missing data, the pooled results of the MEM analyses were presented with 10 multiple imputations."
Selective reporting (reporting bias)	Low risk	The trial was prospectively registered and a protocol was published; all outcomes were reported as planned.
Other bias	Low risk	‐

ADL = activity of daily living; CG = control group; CMAI‐C = Cohen‐Mansfield Agitation Inventory‐Chinese version; IG = intervention group; MDS = minimum data set; MMSE = mini mental state examination; PAINAD‐C = Pain Assessment in Advanced Dementia‐Chinese version; VDS = verbal descriptor scale

Characteristics of excluded studies [ordered by study ID]

Study	Reason for exclusion
Aigner 2018	Ineligible intervention
Breivik 2014	Ineligible study design
Cohen Mansfield 2007	Ineligible study design
Cohen Mansfield 2012	Ineligible intervention
Cohen Mansfield 2014	Ineligible study design
Cook 1998	Ineligible intervention
DRKS00011062	Ineligible intervention
Erdal 2018	Ineligible intervention
Fischer 2008	Ineligible intervention
Habiger 2016	Ineligible intervention
Hirsch 2011	Ineligible study design
Husebo 2011	Ineligible Intervention (participants were allocated to one fixed treatment regime)
Husebo 2014a	Ineligible intervention
Husebo 2014b	Ineligible study design
ISRCTN61397797	Ineligible intervention
Klapwijk 2018	Ineligible intervention
Kovach 2006a	Ineligible study design
Kovach 2006b	Ineligible study design
Kovach 2006c	Ineligible intervention
Kovach 2008	Ineligible study design
Kovach 2012a	Ineligible intervention
Kovach 2012b	Ineligible study design
Kovach 2012c	Ineligible study design
Kutschar 2020	Ineligible intervention
Lichtwarck 2018	Ineligible intervention
Lin 2009	Ineligible intervention
Maltais 2019	Ineligible intervention
McCabe 2015	Ineligible intervention
Nocerini 2012	Ineligible study design
Opie 2002	Ineligible intervention
Pieper 2011	Ineligible intervention
Pieper 2016	Ineligible intervention
Pieper 2018	Ineligible intervention
Rodriguez‐Mansilla 2015	Ineligible intervention
Sandvik 2014	Ineligible intervention
Wong 2009	Ineligible intervention
Yuen 2019	Ineligible intervention

Differences between protocol and review

We amended one predefined outcome in the protocol: 'implementation fidelity' was changed to 'process‐related outcomes, e.g. implementation fidelity'.

In the protocol, we planned to favour self‐rated over proxy‐rated pain measures. We did not follow this approach in the review since in one study with both a self‐ and a proxy‐rating only about two thirds of the participants completed the self‐rating of pain. However, we clearly described the use of self‐ and proxy‐ratings in the studies and addressed this topic in the discussion.

Mortality was planned as a primary outcome in the protocol. Based on a peer review comment, we moved mortality to the secondary outcomes.

Contributions of authors

RM planned the study. VL and RM wrote the study protocol with support of ES and RT. CM and VL selected studies, extracted data and performed the quality appraisal, supervised by RM. CM contacted the study authors. CM, VL and RM conducted the analyses, interpreted the data and wrote the draft of the review. All authors contributed to all drafts of the review.

Sources of support

Internal sources

• No sources of support provided

External sources

• Ministry of Education and Research, Germany

  Grant number 01GL1733; provided the salaries, travelling costs and consumables for CM and VL

• NIHR, UK

  This review was supported by the National Institute for Health Research (NIHR), via Cochrane Infrastructure funding to the Cochrane Dementia and Cognitive Improvement group. The views and opinions expressed herein are those of the authors and do not necessarily reflect those of the Systematic Reviews Programme, NIHR, National Health Service or the Department of Health

Declarations of interest

Christina Manietta: none known

Valérie Labonté: none known

Rüdiger Thiesemann: none known

Erika G Sirsch: none known

Ralph Möhler: none known

New