Dementia as a public health concern

With an aging population, dementia is becoming an increasingly important health issue in the United Kingdom (UK) and across the world. In 2015 it was estimated that 46.8 million people worldwide were living with dementia, and this number is expected to increase to 74.7 million in 2030 and 131.5 million in 2050.[1] The impact of this disorder on patients, their carers, families, and society is profound.[2]

Benefits of timely diagnosis

The World Alzheimer’s Report 2011 highlights the benefits of early diagnosis in allowing people with dementia and their families to make plans for the future, before their disease becomes too advanced.[3] Additional benefits include timely access to information, advice and support, and the person with dementia being able to express their wishes in a way which helps them to maximise their quality of life. Likewise current symptomatic treatments and future disease-modifying medications are likely to have most effect if prescribed early in the illness.[3] Earlier diagnosis is also likely to delay entry to care homes, thus reducing the costs to society of institutional care [4] and possibly also contributing to quality of life, given that most elderly people express a preference to stay in their own home as long as is practicable. However, at the current time, diagnosis often happens at a late stage in the illness, or following a crisis (e.g. after hospitalisation due to a fall) when opportunities for maximising quality of life have passed.[3, 5] Higher diagnosis rates, and diagnosis earlier in the course of the disease, are strategic aims for the UK government and National Health Service (NHS), as described in the National Dementia Strategy,[6] Prime Minister’s Dementia Challenge,[7] NHS England Dementia Identification Scheme,[8] and the GP Dementia Toolkit.[9]

The role of UK general practice in diagnosis

Diagnosis and medical care for people with dementia in the UK starts with recognition of symptoms by patients themselves, their families, or general practitioners (GPs/family physicians). Following referral, dementia symptoms are then investigated by memory assessment services or other specialist teams, where a diagnosis of dementia may be made. However, there is a “diagnosis gap” in general practice with only about a half to two-thirds of patients having a recorded dementia diagnosis, compared to numbers expected from epidemiological studies.[10, 11]

Dementia risk tools and prediction models

Risk scores and clinical prediction models may help clinicians to identify patients at risk of conditions such as dementia earlier than would be the case in routine clinical practice. This is a valuable strategy where early intervention may slow the progression of a disease. A number of risk prediction tools have been developed to stratify patients by risk of dementia [12, 13], drawing on demographic, health, lifestyle, functioning and cognitive factors, as well as blood based biomarkers, genetics and brain imaging. These tools have shown variable performance, with Areas Under the Receiver Operative Characteristic Curve (AUROC) ranging from 0.48 to 0.91 [14]. The majority of models have been built based on white volunteer samples that tend to have biases such as high education levels and better health than the general population, and very few models have been validated in a setting different from the one they were developed in. Models have also not utilised within-subject trajectories on predictor variables, instead mainly using cross-sectional variables [12]. The cost implications of these models also vary widely, as for example, brain MRI scans are much more expensive than administering a short cognitive paper-based test. The cost of any risk tool will determine recommendations about the use of the risk tool in clinical practice [12]. Risk scores which use self-report measures, or particularly, routinely collected health data, will have a cost advantage over those which require expensive tests which are outside the realm of usual clinical practice.

General practice patient records as a source of data for a prediction model

In the UK, GPs see registered patients throughout their lives and record longitudinal data in electronic records on medical and family history and ongoing medical events. GP patient records could therefore be a key data source for the development of a dementia risk prediction tool. GPs record data using a hierarchical system of codes covering symptoms, tests, referrals, medical history and diagnoses, as well as a complete record of prescribing and free text notes.[15] These facets of recording map onto the predictive factors explored in the development of previous tools. GP patient records are also available to researchers, as anonymised samples curated in databases such as The Clinical Practice Research Datalink (CPRD)[16] and The Health Improvement Network (THIN);[17] covering more than 6 million current UK patients.

The use of risk prediction tools in routinely collected health data may enable the early identification of those with dementia and those at higher risk of developing dementia. It could also improve the quality of patient records, by identifying patients who miss codes for their condition, for whatever reason. Once identified, patients can be added to disease registers and provided with ongoing proactive care. In research studies using electronic health record data, clinical prediction models are analogous to case detection algorithms. Algorithms can help to improve either the sensitivity or specificity (or both) of case detection, compared to a single code.

A few studies have attempted to create clinical risk prediction models for dementia from general practice data.[18–20] Walters et al. created a clinical prediction model for dementia from primary care data which had good discrimination for 60–79 year olds but various thresholds for high risk resulted in either a low sensitivity or a low positive predictive value.[20] This model performed poorly in patients over 80 years old. This team used 14 clinical variables, but many more conditions and medications have been associated with dementia. A model developed by another research team achieved a sensitivity of 72% and a specificity of 83%, but the predictors used were not published.[19]

GP patient record databases contain high dimensional data which is sparsely populated. Choosing predictors for a risk tool must generally be done a priori. With a view to improving upon previous risk prediction tools utilising general practice data, the aim of this study was to inform feature selection for a future dementia prediction tool. We aimed to study the literature to find clinical entities in general practice patient records which have previously been found to be associated with dementia. We conducted a systematic literature review to identify papers examining risk factors for dementia using primary care data, and then conducted a meta-analysis of all explored risk factors, in order to identify which were significantly associated with dementia.

Search strategy

Three searches were carried out. In the first search, PubMed and Web of Science, were searched between 7^th Janurary 2016 and 13^th February 2016 by author NG using the following search string: ((((dementia[Title/Abstract]) OR Alzheimer*[Title/Abstract])) AND (((("primary care"[Title/Abstract]) OR "Primary Health Care"[Title/Abstract]) OR family pract*[Title/Abstract]) OR general pract*[Title/Abstract])) AND (((diagnos*[Title/Abstract]) OR onset[Title/Abstract]) OR predict*[Title/Abstract]).

Second, Online bibliographies of publications from CPRD and THIN during June 2016 by EF using keywords “Dementia” and “Alzheimer*”. No date ranges for publications were specified.

Third, an updated search was conducted by EF on 26^th September 2017. PubMed and Web of Science, were searched using the search string defined above.

After duplicates were discarded, the following inclusion criteria were then applied:

1. Studies using routinely collected primary care data extracted from existing databases.
2. Studies that looked at all cause dementia, Alzheimer’s or vascular dementia as their outcome.
3. Studies that had a cohort or case-control design in which exposures to risk factors in dementia cases could be compared to exposures to risk factors in a control group reflective of the general practice population.
4. Risk factors were measured before dementia diagnosis.
5. Studies published in English.

Studies were excluded if they focused on: management, screening tools (except those developed using routine primary care data), imaging or novel biomarkers for dementia; the ability of GPs to diagnose dementia; barriers to diagnosis of dementia; concordance with guidelines; patient/caregiver experience of diagnosis; prediction of prognosis, survival or institutionalization; cognitive decline; presenting clinical guidance or consensus; were review articles; or were published in languages other than English.

Using these criteria, the articles extracted from the literature search were screened based on their title, then abstract, then finally full text analysis, by two authors (NG and EF).

Quality and Bias Assessment

Bias was assessed by author EF using the Risk of Bias Assessment Tool for Nonrandomized Studies (RoBANS).[22] Each study is assigned a category of low risk, high risk or unclear risk in 6 domains: selection of participants, confounding variables, measurement of exposures, blinding of outcome assessments, incomplete outcome data and selective outcome reporting.

Study quality was assessed by author EF using the Newcastle-Ottawa scale for the assessment of the quality of nonrandomized studies.[23] Each study is assessed under 3 domains: Selection of cases/controls (4 items), comparability of cases/controls (1 item) and measurement of exposure (3 items). For each item a study can receive 1 or 2 stars, giving a total out of 10 stars. There are different items for assessing case-control and cohort studies, but each scale totals to 10 stars.

To examine the risk of publication bias within analyses, we performed funnel plots for comparisons which had 10 or more studies. The Cochrane handbook states that “as a rule of thumb, tests for funnel plot asymmetry should be used only when there are at least 10 studies included in the meta-analysis, because when there are fewer studies the power of the tests is too low to distinguish chance from real asymmetry”.[24]

Data extraction

The following meta-data from studies were extracted into a table in Excel: author names, date of publication, country of origin, name of primary care database, data extraction period, type of dementia, case definition of dementia, number of cases and controls, control matching criteria, conditions assessed, medications assessed, follow up time, mean age of patients, and percentage of female patients.

The following study data were extracted by risk factor: total number of dementia cases, number of cases exposed to risk factor, total number of controls, and number of controls exposed to risk factor. Firstly, all results in each study were combined for a category of “all dementia”. Data on Alzheimer’s disease (AD) and Vascular Dementia (VaD) were extracted separately where available. For medication data, the value for “ever prescribed” was extracted from all studies, as dosage format varied. For eight studies where study data were not published in an extractable format, we emailed authors, resulting in data on three studies.

Meta-analysis

ReviewManager 5.3 was used to perform meta-analyses and estimate odds ratios (ORs) for all risk factors. Initially a fixed effects analysis was conducted for each meta-analysis, but where results were heterogeneous (I² > 50%) a random effects analysis was used and is presented. The I² statistic describes the percentage of variation across studies that is due to heterogeneity rather than chance.[24] No correction for multiple comparisons was performed as the outcome were analysed as stated in the study protocol and no post hoc analysis was performed.

Sensitivity analysis

Twelve papers included in the analysis drew their data from the same database (CPRD). Resampling may artificially increase the precision of estimates, reduce heterogeneity and narrow confidence intervals. For every risk factor which had 3 or more CPRD studies in the “all dementia” category, we removed all but one CPRD paper per analysis. The CPRD paper spanning the longest time period was retained; where two studies had equal time spans, the study with the larger sample size was retained.

Results of the searches

The initial database search identified 3,836 unique papers. Following title sifting, 421 abstracts from the database search, and 20 from searches of the CPRD bibliography and THIN publications list were examined, and of these, 94 papers were examined in full text. 19 papers met inclusion criteria and were included in the analysis. A second database search was performed which elicited 555 unique papers from 2016 onwards. Of these 64 abstracts were examined, and 8 papers in full text. One further paper was found to meet inclusion criteria. Thus 20 papers are included in the final analysis (Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. PRISMA flow diagram for study selection.

https://doi.org/10.1371/journal.pone.0194735.g001

Quality and Bias Assessment

Scoring is presented in S1 Table. There were 15 case-control studies all scoring 8 or 9 out of 10 stars on the Newcastle-Ottawa Scale. All 15 studies were categorized as low risk on all domains of the RoBANS, except for one study which had an unclear risk of bias in its handling of confounding variables. There were 5 cohort studies, of which one scored 7, three studies scored 8 and one study scored 9 on the Newcastle-Ottawa Scale. All 5 studies were categorized as low risk on all domains of the RoBANS, except for one study which had an unclear risk of bias in its handling of confounding variables. No studies were excluded on the basis of quality or bias.

We produced funnel plots for two comparisons: diabetes (12 studies) and hypertension (10 studies). Neither funnel plot showed asymmetry, suggesting the risk of publication bias was low (S1 Fig).

Characteristics of studies

The characteristics of studies are shown in Table 1. Fourteen studies were based on UK data (12 from CPRD)[20, 25–37], three from the Netherlands,[18, 38, 39] two from Germany[40, 41] and one from Denmark.[42] Patients had a median follow up time of 5.25 years, the median age of participants was 80.6 years and the median proportion of females in the studies was 63%. Data for studies were recorded in primary care databases between 1990 and 2014. Study sizes ranged from 199 to 930,395 participants. More recent studies tended to have larger sample sizes and data extracts spanning longer periods.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Study characteristics.

https://doi.org/10.1371/journal.pone.0194735.t001

Associations between risk factors and dementia

All associations between dementia and other conditions are shown in Table 2. Forest plots for comparisons with 4 or more studies can be seen in S2 Fig. Results are described by type of predictor, categorised as lifestyle, neuropsychiatric, cardiovascular, other conditions, and cognitive symptoms.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Meta-analysis for lifestyle factors and conditions.

https://doi.org/10.1371/journal.pone.0194735.t002

Associations between medication and dementia

Fewer studies reported on associations with medication, so many results in this section are from single studies. Results are reported in full in Table 3. Results are described by type of medication, categorised by neuropsychiatric, cardiovascular and diabetes, anti-inflammatory, and other types of medication.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Meta-analysis for medications.

https://doi.org/10.1371/journal.pone.0194735.t003

Sensitivity analysis

Twelve UK papers drew their samples from the same study population (CPRD). Fig 2 shows overlap between the data extraction periods for CPRD studies.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Data extraction periods for CPRD studies.

https://doi.org/10.1371/journal.pone.0194735.g002

The meta-analysis was repeated with all but one CPRD study removed for every risk factor which had 3 or more CPRD studies in the all dementia category: BMI, smoking, depression, coronary heart disease, hypertension, dyslipidemia, atrial fibrillation, heart failure, diabetes, COPD, inflammatory conditions, and statins.

Results are presented in Table 4 and forest plots in S3 Fig. A reduction in the strength of association was found for high BMI (OR 1.09; 95% CI 0.93–1.28), all other risk factors retained their associations (or lack thereof).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Sensitivity analysis results.

https://doi.org/10.1371/journal.pone.0194735.t004

Implications and future directions

This review and meta-analysis is designed to guide feature selection for creating a clinical risk tool for dementia in primary care. The types of risk factors consistently associated with dementia suggest that a patient record quality improvement model, or case detection algorithm, could feasibly be developed using primary care data, as the most predictive features include signs and symptoms of dementia that will have already been picked up by GPs. However, a risk stratification model, aiming to identify patients at increased risk of dementia when no clinical signs have yet been identified, may be more problematic, given the few clinical conditions consistently found to be associated with dementia.

Notably absent from the literature, despite a thorough examination of primary care database studies, were other elements in a primary care record which could be utilised in a predictive model for dementia. Patterns of attendance, missed appointments, missed prescriptions, and consultations by family members could all be used as indicators of behaviour change, and might increase the predictive validity of a model. The demonstration of their absence from the literature is a useful outcome of this review, and future research should undertake investigation as to whether these would be valuable predictors.

Strengths and limitations

This is the first meta-analysis using routinely collected primary care data to establish a range of risk factors for dementia. We only included articles published in English, and only represented countries which have an established primary care database system. A single author extracted data from the included studies, which may have introduced error. We also only searched two medical literature databases, although we supplemented this with a targeted search of UK primary care database literature. This gives our work greater relevance to the UK but potentially limits the generalisability of our results beyond the UK primary care system.

We conducted a review of bias in individual studies, and we generated funnel plots to examine systematic heterogeneity for the best populated comparisons. We found little evidence of bias within studies, due to the data all being gathered in real time in routinely collected health records, and then extracted retrospectively, thus diminishing the risk of selection bias, detection bias, performance bias and attrition bias. However, there may have been a publication bias both in studies that were published (by only including significant associations in the results) and negative studies going unpublished. If this were the case, were we to recover these additional unpublished findings we would expect to find even fewer of the risk factors to show associations with the outcome.

Other characteristics of the data also urge cautious interpretation. First, heterogeneity was high for almost all analyses, which was unexpected due to the consistent nature of data generation between studies. This heterogeneity may have been caused by variation in the way participants, exposures and outcomes were defined in studies, such as the variation in average age of participant. The varying practices and coding systems in different countries may also have played a part as well as variation in coding practices of clinicians. For example, clinical entities were defined differently between studies, as some case definitions required a single Read code to define dementia and others a more complicated combination of codes. In many cases, code lists for the risk factors were not available. This may have limited the value of pooling results from studies. We attempted to mitigate the effects of heterogeneity by conducting random effects modelling for many analyses, but further thought may be needed to understand the source of this heterogeneity. Such heterogeneity may be a wider issue to be considered in all meta-analyses performed on routinely collected health record studies.

Another source of heterogeneity could have come from the timing of the measurement of exposures, as this varied between studies. In most studies, very little temporal information was given about when risk factors occurred or were measured, and the best description of our understanding is that they were present at some point in the record. More work is needed to unpick exactly when different risk factors have the greatest association with subsequent dementia, as longitudinal information and time-at-risk could usefully be incorporated into a clinical prediction model.

Additionally, the problem of resampling from the CPRD population meant that cases could not be assumed to be independent. However, our sensitivity analysis approach showed that for analyses with three or more CPRD studies, there were minimal differences between including one versus all CPRD studies, with the exception of BMI >30.

A further potential criticism of the overall endeavour of creating a risk prediction tool from primary care records is that GPs tend to diagnose dementia late in the course of the disease. Thus the data is not optimal for training an early risk identifier, as all studies assessing associations between risk factors and dementia will be identifying risk factors for late-diagnosed dementia on the whole. Additionally, those individuals who do get a diagnosis of dementia may differ systematically from the undiagnosed, and the low level of diagnosis in primary care records (possibly around 50% [11]) means that these differences could have substantial effects on our understanding of associations with various risk factors. As primary care records are human artefacts, created during the course of human interactions with only clinical purposes in mind, researchers must also be mindful of bias due to the underlying reasons or timings for recording, as well as choice of codes or free text, which may have multiple individual or systemic influences. Counter to this, the fact that the data is gathered prospectively in real time means that if the risk tool allows a long run up period, a wide collection of indicators will be found, including those linked to the early stages of the disease. Data collected in real time is also less biased by knowledge of the outcome. One future approach to mitigate the lack of coded diagnosis in the records would be to search for evidence of dementia or cognitive impairment in the textual, narrative, parts of the records, such as letters from specialists.

Conclusions

This study provides a summary of the current literature on risk factors for dementia as ascertained from routinely collected primary care records. Few factors were consistently associated with dementia, and many gaps in the literature for examining potential risk factors were found. These findings are of potential value for feature selection for building a clinical risk prediction tool to aid the identification of cases of dementia in primary care.