Antidepressant use at the threshold: people prescribed antidepressants around the time of a dementia diagnosis

Abstract

Background

Antidepressant use is common in people with dementia, and many start around the time of diagnosis. We hypothesised that those who commenced antidepressants at dementia diagnosis started for dementia-related symptoms, distinct from those with longstanding prescriptions, resulting in different characteristics in electronic records.

Methods

We carried out a retrospective cohort study using linked primary care (Lambeth DataNet) and mental health (Clinical Record Interactive Search) data for patients with dementia in south London, UK in 2009–22. Antidepressant prescription starting within one year before or after dementia diagnosis identified as ‘new’. Dementia and anxiety were described using coded fields, a rating scale for neuropsychiatric symptoms and natural language processing of full-text.

Results

Of 3713 patients with dementia, 28% were prescribed an antidepressant in the year of dementia diagnosis. 42% of these were ‘new’ prescriptions. Compared to the no antidepressant group, ‘new’ starters were more likely to be female (odds ratio (OR) 1.33, 95% 1.08–1.64) and have vascular dementia (OR 1.44, 1.13–1.83). Compared to longstanding antidepressant group, ‘new’ had similar depression and anxiety symptoms, but fewer coded mentions (OR 0.50, 0.39–0.65), especially those ‘new’ from non-White ethnicities. Deprescribing was equally unlikely in ‘new’ and longstanding (6.3% vs 5.5% per year of follow-up).

Conclusion

There is a high prevalence and lack of deprescribing, despite guidelines cautioning antidepressants. This points to unmet needs at the threshold of dementia. Clinicians should ensure good documentation of indication for antidepressants and proactive medication reviews for people with dementia.

Graphical Abstract

Graphical Abstract.

Key Points

• Antidepressants have poor evidence of efficacy, and the possibility of adverse effects, in dementia.

• In the year of a dementia diagnosis 28% had an antidepressant prescribed, nearly half (42%) started less than a year before.

• Natural language processing showed depression symptoms in those taking antidepressants high, regardless of when commenced.

• Deprescription of antidepressant was low, at 6% per year of follow-up.

Background

The threshold period where dementia is suspected, investigated and newly confirmed can be a confusing and distressing time. The differential diagnosis for symptoms of depression and anxiety during this time includes depressive and anxiety disorders, neuropsychiatric symptoms of dementia and reactions to the uncertainty during the assessment process and changing social roles [1, 2]. Depression is a risk factor for dementia, and symptoms of depression often feature as a prodrome of dementia [3, 4]. Depression may also have distinct neurobiological features when appearing in someone with dementia [5, 6].

The prevalence of antidepressants in adults of all ages is ~23%, rising with female sex and older age [7, 8]. Some people experiencing depression and anxiety around the time of dementia diagnosis may benefit from antidepressants [9]; however, prescribing for older people, particularly those with dementia, has elevated risks [10]. Non-antidepressant strategies are considered first-line in people with dementia, including psychological treatment for depression and anxiety, personalised activities for agitation and personalised sleep management for insomnia [1].

We aimed to characterise antidepressant prescribing around the time of dementia diagnosis, the extent to which depression and anxiety were documented, and the profiles of individuals who had recently initiated an antidepressant, especially those without a recorded diagnosis of depression or anxiety. Our findings aim to support guidance and resource planning for the clinical challenge of managing mental health symptoms in people when dementia is suspected or recently diagnosed.

Methods

We assembled a retrospective cohort of people with dementia and identified people prescribed antidepressants in the years around diagnosis. Documentation of depression or anxiety was studied, alongside demographics, dementia features and cessation of antidepressant.

Data source

We used electronic health record data from the Clinical Record Interactive Search (CRIS) at South London and Maudsley NHS Foundation Trust (SLaM) [11], and a pre-established linkage to Lambeth DataNet (LDN). CRIS provides pseudonymised structured and free text data from SLaM, a provider of dementia and mental health care for four London boroughs [12]. LDN provides structured data about people registered with a GP in the socially and ethnically diverse London Borough of Lambeth (about 405 000 residents), one of SLaM’s catchment boroughs [13]. Natural Language Processing (NLP) has been deployed extensively in CRIS to extract a range of entities from text. Linkage includes all eligible Lambeth patients unless they have opted out of CRIS or LDN. Linkage and extraction was by the CRIS Linkage Service; investigators also had access to CRIS.

Sample

We selected patients with a documented dementia diagnosis in LDN or CRIS as described previously [11] using the ICD-10 fields in CRIS and Read codes in LDN. The date of the first documented diagnosis of dementia was the index date, and eligible patients had an index date between 2009 and 2022, aged 65 or over, registered with a Lambeth GP and with a SLaM contact at some point from 2 years prior to 4 years post-index.

Exposure

Antidepressant prescription was ascertained from the LDN prescription fields. The antidepressant positive group are those who were issued a prescription for antidepressants between the index date and 12 months later (the index year). We allocated each patient to one of three groups:

• (i) ‘New start’: prescribed antidepressant in the index year, but not in the year 12–24 months before the index date, targeting those starting in the two years between 12 months before and 12 months after the first documented dementia diagnosis.

• (ii) ‘Longstanding’: prescribed antidepressant in the index year and in the year 12–24 months before index (started >1 year before diagnosis).

• (iii) ‘None’: not prescribed antidepressants in the index year.

A visual explanation of the group allocations is shown in Supplementary Appendix 1.

We identified antidepressants through the British National Formulary [14], grouped into SSRI, tricyclic (TCA) and ‘other’ antidepressants, and matched these to the codes used by LDN (drugs + medical devices, DMD, shown in Supplementary Appendix 2) to search. Using a window of observation from 2 years before diagnosis until 4 years post-diagnosis, we also noted the class of antidepressants, including the use of more than one of these in any year. Deprescription was defined as no antidepressant prescription for a complete year during follow-up. Follow-up ended after 4 years, at death, or if a patient left LDN according to registration data.

Other variables

The derivation of all variables is shown in Supplementary Appendix 3A and timing in Supplementary Appendix 3B. Demographic factors were age at diagnosis, sex, marital status, ethnicity and neighbourhood deprivation. The Charlson comorbidity index [15] (excluding dementia) from primary care codes was a measure of physical health detailed previously [11]. Dementia factors were subtype, severity at diagnosis and difficulty of activities of daily living (from closest HoNOS 65+) [16].

Depression, anxiety and neuropsychiatric documentation

To study diagnoses and symptoms, we used:

• (i) Codes related to depression and anxiety: Coded mentions were primary care Read code in LDN from a list [17] or ICD-10 codes from the categories F32, F33 and F4 in CRIS at any date.

• (ii) Full text mentions of symptoms: Using pre-existing NLP tools that identify symptoms of depression and anxiety synonyms [18], we assessed mentions in the CRIS text recorded in the 6 months either side of the index date. These are reported as being out of 21 included concepts for depression and 20 trigger words for anxiety, shown in Supplementary Appendix 4.

• (iii) A rating scale: Neuropsychiatric symptoms were derived from the clinician-completed Health of the Nation Outcome Scale (HoNOS) 65+ [16] in CRIS, the closest to the index date. Sum of scores (0–4) for each of: agitation, self-injury, substance use, depression, hallucinations/delusions and other problem (which may be anxiety) up to 24, the actual distribution being a range of 0–18 with mean of 3.

Analyses

We carried out the following analyses:

• Antidepressant factors: We identified prescription prevalence in the index year, approximate onset, grouping the cohort into ‘new start’, ‘longstanding’ and no antidepressant. We described the classes prescribed.

• Demographic and clinical factors: Characteristics were summarised using proportions or means and standard deviations for each group. Univariate regression models were used to quantify associations between groups using linear or logistic models, as appropriate for each dependent variable, giving mean difference or odds ratio with 95% confidence intervals.

• Depression/anxiety measures: These were summarised and compared between groups, as for demographic and clinical factors. Those with ‘new start’ of antidepressant were divided into two groups depending on whether they had a code for depression/anxiety, and clinical and depression/anxiety measures compared as previously.

• De-prescription: Deprescription was identified as above and in Appendix 1 in supplementary data. The number of deprescriptions per year of follow-up was calculated and compared between ‘new start’ and ‘longstanding’ groups using chi-squared tests.

Data were organised using R/R-studio and analysed in STATA and excel.

The number of demographic, clinical and anxiety/depression factors was 23 for each of the two contrasts (new vs none and new vs longstanding). To avoid false discovery due to multiple comparisons, we defined significance using the Benjamini–Hochberg corrected P-value [19] less than .05. Supplementary Appendix 5 shows the results with raw and corrected P-values. All other tables show symbols indicating the significance of the corrected P-values.

Post-hoc power calculations were performed, and provided in Supplementary Appendix 6.

Missing data and bias

Items with missing data were demographic factors of marital group and ethnicity, where an unknown group was added. In clinical features, dementia subtype contained an ‘unknown’ category for ‘unspecified’ subtype and missing. Dementia severity could only be calculated if there was an MMSE or HoNOS in the 6 months either side of the index date (79%), and if not, severity was unknown. For the other HoNOS-related outcomes, we used the closest to diagnosis, which was within six months of diagnosis for 76%. When calculating neuropsychiatric scores, we excluded those with no HoNOS ratings (9%).

Those people with no records within 6 months either side of the index date would score zero on depression and anxiety NLP scores if included. We were not directly able to ascertain who these were, so to reduce the risk of bias, if a person had no MMSE nor HoNOS in that time-period (21%), they were be excluded from the NLP results. These exclusions are evident in the cohort size of Tables 2 and 3.

Table 2.

Depression, anxiety and neuropsychiatric factors divided by antidepressant status.

		Antidepressant status			New vs none		New vs long
Characteristics	Full cohort	New start	Long-standing	None	Odds ratio or mean difference (95% CI)	siga	Odds ratio or mean difference (95% CI)	siga
No. of patients	3713	441	599	2673
Depression/anxiety codesb  ever (n = 3713)
Depression	32.5%	52.2%	68.5%	21.1%	4.07 (3.30–5.01)	**	0.50 (0.39–0.65)	**
Anxiety	21.7%	29.7%	44.9%	15.2%	2.36 (1.87–2.97)	**	0.52 (0.40–0.67)	**
Either	41.9%	64.2%	78.8%	30.0%	4.18 (3.39–5.17)	**	0.48 (0.37–0.64)	**
Neuropsychiatric ratingc  closest to diagnosis (n = 3387)
Combined scale mean (SD)	3.0 (2.5)	3.8 (2.5)	3.6 (2.7)	2.7 (2.4)	MD: 1.13 (0.87–1.40)	**	MD: 0.29 (0.04–0.61)
Natural Language Processing tool scored  six months either side of diagnosis (n = 2934)
Depressive symptoms, mean (SD)	1.9 (2.0)	2.8 (2.3)	2.4 (2.5)	1.7 (1.7)	MD: 1.10 (0.86–1.34)	**	MD: 0.37 (0.09–0.71)	**
Anxiety symptoms, mean (SD)	2.0 (2.0)	2.7 (2.3)	2.5 (2.4)	1.8 (1.8)	MD: 0.87 (0.66–1.14)	**	MD: 0.17 (−0.10 to 0.50)

Comparison between groups with odds ratio (OR) or mean difference (MD) with 95% confidence interval and significance based on P-values corrected for multiple comparisons. Bold text for OR or MD represents significant associations.

^aSignificance according to P-values corrected for multiple testing where ^#P < .1, ^*P < .05 and ^**P < .01.

^bRead codes from LDN and ICD-10 codes from CRIS.

^cSum of six symptom subscales from HoNOS.

^dUsing existing natural language processing tools: number of distinct depression symptoms (maximum 20); number of distinct anxiety synonyms (maximum 21).

Table 3.

Characteristics of antidepressant new starters with separation by the documentation status of depression and anxiety.

	New starters	Depression or anxiety code documented in LDN or CRISa		Any vs no code
Characteristics		Any codes	No codes	Odds ratio or mean difference (95% CI)	sigb
N	441	283	158
Socio-demographic status
Age at diagnosis, mean (SD)	81.3 (7.3)	81.1 (7.2)	81.8 (7.6)	MD: −0.69 (−2.11–0.74)
Female gender	62.6%	63.3%	61.0%	1.08 (0.72–1.62)
Married	30.7%	29.6%	32.6%	0.87 (0.55–1.36)
Non-White Ethnicity	37.2%	31.6%	47.3%	0.51 (0.34–0.78)	**
More deprivedc	45.4%	46.3%	43.7%	1.11 (0.75–1.64)
Health status
Comorbidity scored mean (SD)	2.0 (1.9)	2.0 (1.9)	2.0 (1.8)	MD: −0.04 (−0.41–0.33)
Dementia factors—scores from assessment closest to diagnosis
Subtype
Alzheimer’s disease	54.7%	55.8%	52.5%	1.14 (0.77–1.69)
Vascular or mixed	23.4%	22.3%	25.3%	0.84 (0.54–1.33)
Other	5.7%	5.7%	5.7%	0.99 (0.43–2.30)
Unspecified	16.3%	16.2%	16.5%	0.99 (0.58–1.67)
Dementia severity at diagnosis
Mild dementia	28.8%	33.2%	20.9%	1.88 (1.19–2.97)	**
Functioning (n = 3387)
Problems with activities of daily livinge	66.7%	65.1%	69.4%	0.82 (0.53–1.27)
Depression and anxiety factors
Neuropsychiatric symptomse  closest to diagnosis(n = 3387)
Combined scale, mean (SD)	3.8 (2.5)	3.9 (2.4)	3.7 (2.8)	MD: 0.18 (−0.35–0.71)
Natural Language Processing tool scored  six months either side of diagnosis (n = 2934)
Depressive symptoms, mean (SD)	2.4 (2.3)	2.6 (2.4)	1.9 (2.2)	MD: 0.66 (0.22–1.11)	**
Anxiety symptoms, mean (SD)	2.2 (2.3)	2.5 (2.4)	1.7 (2.0)	MD: 0.79 (0.35–1.23)	**

Comparison by odds ratio (OR) or mean difference (MD) with significance from pairwise regression between the documentation groups. Bold text for OR or MD signifies significant association.

^aRead code from LDN or ICD-10 code from CRIS.

^bSignificance where ^*P < .05 and ^**P < .01 (no correction for multiple testing).

^cIndex of Multiple Deprivations in 30% most deprived in England, based on approximate address.

^dBased on Charlson comorbidity index (with dementia not scored).

^eUsing scale on Health of the Nation Outcome Scale (HoNOS) 65+ scale for ADLs.

^fSum of symptom subscales from HoNOS.

^gUsing existing natural language processing tools: number of distinct depression symptoms (maximum 21); number of distinct anxiety synonyms (maximum 22).

A completed RECORD reporting checklist [20] has been included as Supplementary Appendix 7.

Results

Cohort

The cohort assembly is shown in Figure 1, with 3713 patients included in the analyses. Of the cohort, 2539 (68%) had codes for dementia in both CRIS and LDN, while 719 (19%) had codes in CRIS only, so that 87% were definitely known to have dementia in the specialist service. Documentation of at least one MMSE or HoNOS rating within 6 months either side of the index date was present in 2934 (79%). The mean follow-up time was 2.8 years (SD 2.6), with 21% alive and in LDN after four years.

Figure 1.

Cohort flow chart showing the formation of a dementia cohort through combining patients with a code for dementia from two sources (Clinical Records Interaction Search and Lambeth DataNet).

Characteristics of the full cohort are in Table 1. The mean age at dementia diagnosis was 82 years, 59% were female, and the most commonly recorded ethnicity was White (61%) with 6% unknown. The most common dementia subtype was Alzheimer’s disease (62%), while 31% had mild dementia at diagnosis.

Table 1.

Cohort characteristics, with comparison between groups defined by antidepressant status.

		Antidepressant status			New vs none		New vs Long
Characteristics	Full cohort	New start	Long-standing	None	OR or MD (95% CI)	siga	OR or MD (95% CI)	siga
N	3713	441	599	2673
Socio-demographic status
Age at diagnosis, mean (SD)	81.6 (7.3)	81.3 (7.3)	80.7 (7.6)	81.9 (7.1)	MD: −0.59 (−1.32 to 0.14)		MD: 0.67 (0.22–1.56)
Female sex	58.6%	62.6%	68.8%	55.6%	1.33 (1.08–1.64)	*	0.76 (0.59–1.54)	.
Married	30.6%	31.3%	27.7%	30.7%	0.97 (0.77–1.23)		1.16 (0.64–1.54)
Non-White Ethnicity	38.7%	37.2%	30.4%	40.8%	1.13 (0.92–1.38)		0.68 (0.53–0.88)
More deprivedb	49.0%	45.4%	48.8%	49.7%	1.12 (0.88–1.43)		1.17 (0.88–1.59)
Health status
Comorbidityc (mean, SD)	2.1 (1.9)	2.0 (1.9)	2.5 (2.1)	2.1 (1.9)	MD: −0.09 (−0.29 to 0.10)		MD −0.51 (−0.74 to −0.27)	**
Dementia factors—scores from assessment closest to diagnosis
Subtype
Alzheimer’s disease	61.8%	54.7%	53.9%	64.8%	0.66 (0.53–0.80)	**	1.03 (0.76–1.24)
Vascular or mixed	18.5%	23.4%	19.5%	17.5%	1.44 (1.13–1.83)	*	1.26 (0.93–1.69)
Other	4.9%	5.7%	7.9%	4.0%	1.43 (0.91–2.23)		0.71 (0.43–1.17)
Unspecified	14.8%	16.3%	18.7%	13.7%	1.23 (0.93–1.62)		0.85 (0.61–1.17)
Dementia severity at diagnosis
Mild	31.4%	28.8%	30.6%	31.4%	0.92 (0.74–1.15)		0.88 (0.68–1.16)
Moderate	23.4%	24.7%	23.0%	23.4%	1.10 (0.87–1.39)		1.08 (0.81–1.43)
Severe	24.7%	29.3%	27.0%	24.7%	1.15 (0.89–1.50)		1.13 (0.92–1.38)
Unknown	20.5%	17.2%	19.4%	20.5%	0.87 (0.66–1.13)		0.81 (0.59–1.11)
Functioning (n = 3387)
Problems with ADLsd	62.0%	66.7%	66.0%	60.3%	1.32 (1.05–1.64)	#	1.03 (0.79–1.35)

^aSignificance according to corrected P-value: ^#P < .1, ^*P < .05 and ^**P < .01.

^bIndex of multiple deprivations in 30% most deprived in England, based on approximate address.

^cBased on Charlson comorbidity index (with dementia not scored).

^dActivities of Daily Living, using ADL subscale on Health of the Nation Outcome Scale (HoNOS) 65+. Bold OR or MD signifies significant association.

Antidepressant-related features

In the index year, there were 31 746 prescriptions for 1040 patients (28%). For those prescribed an antidepressant in the index year, 441 (441/1040, 42%) were ‘new start’, and 599 were ‘longstanding’. Both antidepressant groups had a mean follow-up of 2.9 years (SD 2.6) post index diagnosis.

Appendix 7 in supplementary data shows prescribing in the index year by antidepressant class. From the 1040, 53% received at least one prescription for an SSRI, 37% for ‘other’ antidepressants and 21% for a TCA. TCAs were prescribed for 28% (168/599) of the longstanding group and 13% (55/440) of the new start group, giving an odds ratio of 0.37 (95% confidence interval 0.26–0.51). Twelve percent received a prescription for more than one of those classes in the index year, rising to 36% when including 2 years before to 4 years after diagnosis.

Demographic and clinical factors

Table 1 describes the three antidepressant groups. When compared with no antidepressant, the ‘new start’ group had more female patients (OR 1.33, 1.08–1.64), vascular subtype (OR 1.44, 1.13–1.83) and ADL impairment (OR 1.32, 1.05–1.64) but no relationship with age, marital status, ethnicity, deprivation or comorbidity. When compared with longstanding antidepressant, the ‘new start’ group showed a mean comorbidity score about half a point lower (−0.51, −0.74 to −0.27) than the longstanding group, and a suggestion of difference in ethnicity that did not reach significance.

Depression/anxiety factors

Of the full cohort, 42% had at least one depression or anxiety code in GP records, and 53% had at least one recorded neuropsychiatric problem on the HoNOS, mean 3.0 on the combined scale. A mean of 1.9 symptoms of depression, and 2.0 anxiety-related words were identified by NLP. In Table 2, the ‘new start’ group had significantly higher scores on all of these measures compared to the no antidepressant group. Comparisons between ‘new start’ and ‘longstanding’ groups showed ‘new start’ had a lower prevalence of depression/anxiety codes (OR 0.48, 0.37–0.64) but slightly higher depression NLP score (2.8 vs 2.4), and no significant difference on neuropsychiatric rating or anxiety NLP scores.

In the ‘new starters’ group, 36% did not have depression and anxiety codes in primary or specialist databases, and may have had an alternative indication. Table 3 separates those with and without formal depression or anxiety documentation. Those from a white ethnicity and with mild dementia are more likely to have codes. Those without codes had a lower mean score for NLP depression and anxiety, but a similar neuropsychiatric rating.

De-prescription

Of the 1040 prescribed antidepressants in the index year, 175 (17%) stopped all antidepressants for a year during follow-up. In the ‘new start’ group, 81 people (of 441) stopped antidepressants during a mean 2.9 years of follow-up: a mean of 6.3% per year of follow-up. In the ‘longstanding’ group 95 people (of 599) stopped antidepressants during a mean follow-up of 2.9 years: a mean of 5.5% per year. There was no significant difference between the groups (yearly data, P = .557).

Discussion

We defined a cohort of 3713 people with dementia using routinely collected information, and defined a ‘threshold’ of dementia as one year prior to the first documented diagnosis of dementia to one year after. This is a time of suspicion and assessment, confirmation of dementia and adjustment. The prevalence of antidepressant prescription was 28% in the year of dementia diagnosis. The 441 patients who had a ‘new’ antidepressant prescription in this period accounted for 42% of those prescribed antidepressants. Compared to those without antidepressants, the ‘new’ group were more likely to be women (OR 1.33, 1.08–1.64) and to have vascular dementia (OR 1.44, 1.13–1.83). To address confounding by the risk factors of depression, we also compared with longstanding antidepressant group, showing similar in demographics except lower comorbidity in the ‘new’ group (mean 2.0 vs 2.5). Symptom measures were higher in the ‘new’ group, reaching significance for depression (2.8 vs 2.4), but not anxiety (2.7 vs 2.5). The 36% of ‘new’ group without coded entries for depression/anxiety differed in ethnicity, with those from non-White ethnicities substantially less likely to have a code (OR 0.51, 0.34–0.78), with lower NLP symptom scores (depression mean 1.9 vs 2.6, anxiety mean 1.7 vs 2.5) but similar neuropsychiatric rating (mean 3.9 vs 3.7). During follow-up of those with antidepressant prescriptions in the index year, 36% were prescribed more than one class of antidepressants and 17% (a mean of 6% per year) stopped antidepressants.

Around 30% of were prescribed TCAs, despite being cautioned in older people due to risk of cognitive decline, falls and cardiovascular issues [21, 22]. The lower use of TCAs in the ‘new’ group may mean prescribers were less keen to start TCAs in older people but may continue existing prescriptions. SSRIs were used more commonly, and though safer than TCAs, may cause QT prolongation, hyponatraemia and gastrointestinal complications in older people [23].

Vascular dementia was more common in the new start group (23%) compared to non-antidepressant group (18%), which supports the hypothesis that late-life depression may be linked to vascular pathology [24]. If depression is a marker of cerebrovascular disease, antidepressants, for which most of the evidence comes from younger adults with no vascular pathology, may be ineffective or inappropriate [5].

Depression/anxiety codes were found for 30% with no antidepressant, 79% with longstanding and 64% with ‘new’ antidepressant prescription. Therefore 36% of the ‘new’ group were without codes, similar to the 33% with no codes for depression in a UK cohort prescribed antidepressants [7]. Under-recording, a phenomenon common to dementia and mental disorders [11], will account for some of this gap. There are also alternative indications, such as pain, migraine and irritable bowel; and this may explain the higher comorbidity score in the longstanding antidepressant group. In the ‘new’ antidepressants group, some without depression/anxiety may be prescribed for distressing symptoms of dementia and pre-dementia, such as agitation, sleep disturbances or apathy [9, 25, 26], which may account for the slightly higher neuropsychiatric score despite lower NLP symptom scores in these patients.

The 36% of the ‘new’ starters with no depression or anxiety codes were significantly more likely to have a non-White ethnicity. Although care needs to be taken with this result, as it is not corrected for confounding, it may represent bias in healthcare. Previous work in the UK found that patients from Black ethnicities with depression were less likely to have this recorded and more likely to receive pharmacological treatment instead of psychotherapy [27]. In dementia care, people from minoritised communities have been prescribed more potentially harmful anticholinergic medication and antipsychotics on average, and fewer cognitive enhancers [28].

Where antidepressants were prescribed in the index year, they were continued in every year of follow-up in 83%. Give the prior evidence of poor efficacy in people with dementia [5, 29, 30], seen in the 36% who combined or switched antidepressant classes (a proxy for hard-to-treat depression [31]), and with no proven action of antidepressants against deterioration in dementia [10, 21], the lack of deprescribing suggests either a lack of medication reviews, a concern about stopping antidepressants or both [32–34]. This leads to prolonged and potentially unnecessary antidepressant use.

Strengths and limitations

The linked data sources used in this study provide a near-unique resource for these analyses in the level of detail available from multiple care providers, including comprehensive prescription data, primary care presentation coding, specialist care scales and metadata on mental health and dementia derived from NLP. This allows comparisons that would not be possible using only coded data or only one health provider. However, the data are drawn from a single area of London, so wider generalisability cannot be assumed. Although the setting is restricted, we suggest that the themes that come up in the discussion will be similar throughout the UK and beyond.

Routinely collected data inevitably reflects recording biases, which can introduce misclassification in cohort selection and in the ascertainment of clinical characteristics. For example, we previously showed that dementia was more frequently documented when cognitive enhancers were prescribed [11], and a similar pattern may occur for depression and antidepressants, potentially inflating recorded rates of depression among those receiving such prescriptions. Our inclusion criteria may also introduce bias: requiring contact with specialist services could disproportionately select individuals with comorbid mental health difficulties, and NLP outputs depend partly on the volume of documentation, which may be longer for reasons unrelated to symptom severity. The depression NLP tool has been extensively validated in CRIS, whereas the anxiety tool is less developed and may be less reliable. In addition, using diagnosis codes as the index date with a two-year look-back may miss earlier dementia-related treatment, given the average 3.5-year interval between symptom onset and diagnosis [35], may miss the real threshold of dementia-related treatment. Limited documentation of non-pharmacological and social interventions further restricts our ability to account for these factors.

As our aim was to describe prescribing behaviour rather than infer causality, we did not adjust for confounding, and our methods did not allow for a consistent ordering of antidepressant start and symptom measures, so efficacy cannot be inferred. Due to a lack of access in LDN to source text fields in the primary care record, we cannot search for more information on decisions about prescribing, and such question may be answered best by different approaches, such as qualitative research, which could also be helpful to capture the voices of patients and carers.

Implications for clinical practice

Our study confirmed that long-term antidepressant treatment in people with dementia has become common and that some of those starting an antidepressant around the time of diagnosis may have different reasons for starting to those with longstanding prescriptions. Clinicians should be aware, when considering an antidepressant for potentially time-limited symptoms, that it is likely to be continued, and assess whether other support available could delay commencement. Minoritised communities, such as people with non-White ethnicities in this study, appear to have less optimal treatment, so access to alternative treatment and support should be tailored to better meet these needs. Documentation of the reasons for starting an antidepressant, and sharing of this, may be necessary for ensuring all and future prescribers are informed. Integrating validated scales into routine clinical practice [36, 37] can objectively document symptom progression to support decisions about antidepressant initiation and discontinuation. Medicine reviews should be scheduled regularly in people with dementia to consider the risks and benefits of all long-term medication.

Conclusions

In a cohort from London, UK, over a quarter of people recently diagnosed with dementia were prescribed an antidepressant, and around half of these started around the time of the dementia diagnosis. There is a consensus that keeping medication burden as low as possible is preferable in people with dementia, and therefore social support and psychological treatments would be preferred over medication, although this requires the availability of appropriate services. Once antidepressants are started, they are usually continued, so there is an important role for primary and secondary care collaboration at the threshold of dementia to prevent initial prescribing, and a need for universal annual dementia medication reviews to stop any medications doing more harm than good.

Declaration of Conflicts of Interest

The authors wish to acknowledge funding not directly related to this publication: DATAMIND HDR UK Mental Health Data Hub MR/W014386 (R.S.), Department of Psychiatry at Lausanne University Hospital (CHUV) (B.P.M.), NIHR Applied Research Collaboration South London (R.S.), NIHR HealthTech Research Centre in Brain Health (R.S. and C.M.), SPIN-Dementia Network+ (K.A.S.D.), UK Prevention Research Partnership (Violence, Health and Society) MR-VO49879/1 (R.S.), University of Lausanne Tremplin Grant (B.P.M.).

Declaration of Sources of Funding

This work was partially funded by the National by the National Institute for Health and Care Research (NIHR) Biomedical Research Centre (BRC): Maudsley. The views expressed are those of the author(s) and not necessarily those of the NIHR or the Department of Health and Social Care.

Data Availability

The data used in this study are available at CRIS at South London and Maudsley BRC hub, but restrictions apply for data confidentiality, see https://www.maudsleybrc.nihr.ac.uk/facilities/clinical-record-interactive-search-cris/information-for-researchers/