This study harnessed routinely-collected memory clinic data to define two SCD samples according to different operationalizations of the Jessen et al. research criteria [4], and compared the incidence rate of dementia between them. The main finding was that the age-corrected incidence rate of dementia was over threefold greater in SCD patients when the Jak/Bondi (versus Winblad) criteria were used to exclude MCI.

Interestingly, the SCDJak/Bondi sample (n = 185) was larger than the SCDWinblad (n = 86) sample. The retrospective application of the Jak/Bondi criteria resulted in 111 patients originally diagnosed with MCI (according to Winblad criteria) being reclassified as SCD. Whilst some patients were reclassified from SCDWinblad to MCIJak/Bondi, these were fewer in number (n = 12). Interestingly, none of these 12 individuals progressed to dementia during follow-up. Whilst the increased dementia incidence rate in SCDJak/Bondi thus appears to be driven by the inclusion of patients formerly classified as MCIWinblad, it may also reflect that 12 patients who did not progress to dementia were included in the SCDWinblad but not the SCDJak/Bondi sample. Importantly, the SCDJak/Bondi sample was older than the SCDWinblad sample. Dementia risk increases with age [38] – it is thus likely that, by broadening the definition of SCD to include individuals with worse cognition, a greater number of older adults with incipient cognitive decline were captured.

The SCDJak/Bondi (versus SCDWinblad) sample had lower baseline scores on the CAMCOG-R, category fluency, and both immediate and delayed LM recall. A previous study of cognitively unimpaired participants found that scores on the 5-item preclinical Alzheimer’s cognitive composite (PACC5) – which measures global cognition, category fluency, verbal memory and executive function – are inversely associated with cerebral amyloid-β load [39]. Moreover, an earlier study from the Essex Memory Clinic found that poorer global cognition and delayed verbal recall predicted incident AD in a combined SCD/MCI sample [40]. The present finding that, versus SCDWinblad, SCDJak/Bondi is characterized by poorer global cognition, category fluency and verbal recall, as well as an increased risk of dementia (predominantly due to AD) is thus broadly in-keeping with prior work linking specific cognitive profiles to early AD.

We observed dementia incidence rates of 15 and 76 per 1,000 person-years in SCDWinblad and SCDJak/Bondi, respectively. Slot et al. [7] leveraged six memory clinic cohorts to estimate the incidence rate of dementia in 1,530 patients with SCD (mean ± SD age: 67 ± 9 years); five of the cohorts excluded MCI on the basis of a single impaired score, while the remaining cohort excluded MCI via Jak/Bondi criteria. Pooling across cohorts, Slot et al. reported a dementia incidence rate of 20 per 1,000 person-years. In a study using National Alzheimer’s Coordinating Center (NACC) data, Ward et al. [41] calculated that the incidence rate of dementia was 134 per 1,000 person-years in 3,428 patients with MCI (age: 76 ± 7 years) diagnosed on the basis of a single impaired score. The incidence rate in the current SCDWinblad sample thus appears broadly comparable to that observed in the multicenter study by Slot et al. [7] (which largely employed a similar approach to excluding MCI). The incidence rate in the SCDJak/Bondi sample is intermediate between Slot et al. and the MCI sample from NACC [41]. Given the SCDJak/Bondi sample comprised 74 patients originally classified as SCDWinblad, plus 111 reclassified from MCIWinblad, this intermediate incidence rate appears plausible. Moreover, a previous study by Rhodius-Meester et al. [42] reported prognostic data for three SCD samples, one of which was ascertained by the German Dementia Competence Network (DCN; n = 269). Importantly, the DCN use Jak/Bondi criteria for MCI [43]; their SCD operationalization thus appears analogous to SCDJak/Bondi. Rhodius-Meester et al. [42] did not report dementia incidence rates, but they did include sufficient data to calculate these – we calculated a rate of 54 per 1,000 years for the DCN sample. This rate appears lower than that observed for SCDJak/Bondi in the current study (76 per 1,000 years); however, the current patients were older than the DCN sample (age: 73 ± 9 versus 66 ± 8 years). Whilst the incidence rate of dementia in SCDJak/Bondi may thus be higher than expected for an SCD sample, it remains broadly congruent with findings from other cohorts.

Whilst fewer patients developed dementia in the SCDWinblad (versus SCDJak/Bondi) sample, a greater proportion had MCI at their last available assessment (30% versus 11%). Whilst this may caution against conceptualizing SCDWinblad as a reliably ‘benign’ or ‘stable’ phenotype, the between-sample difference in MCI incidence does not affect the overall conclusions of this study; the majority of SCDJak/Bondi patients who progressed to dementia would likely have transitioned through the MCI stage (not reflected in the ‘final’ proportions with MCI reported above). Moreover, MCI does not always progress to dementia [44], and can revert (i.e., improve) in around 8% of cases in clinical settings [45]. Thus, whilst a greater proportion of patients with SCDWinblad progressed to MCI, overall this phenotype had a markedly better prognosis (i.e., a lower dementia incidence rate) versus SCDJak/Bondi. One question not answerable using the current design is whether there is a difference in the rate of cognitive decline between SCDWinblad and SCDJak/Bondi; this could be a fruitful line of investigation for future research.

Recall that Jak/Bondi criteria employ a conservative cut-off to classify cognitive scores as impaired (-1 SD below the normative mean), but that more than one impaired score is required for MCI (see Methods/Table 1). Bondi et al. [14] retrospectively applied the Jak/Bondi MCI criteria to a sample of Alzheimer’s Disease Neuroimaging Initiative participants, including 846 patients with MCIWinblad, and 304 cognitively unimpaired individuals (without SCD). Following reclassification, 401 participants had MCIJak/Bondi, while 749 were cognitively unimpaired. Compared to the original MCIWinblad sample, MCIJak/Bondi was characterized by more consistent cognitive impairment; a greater proportion of APOE ε4 carriers; more AD-like CSF profiles; and a greater dementia incidence rate. In the present study, the application of Jak/Bondi criteria similarly classified a smaller number of individuals as MCI (here resulting in a greater number with SCD). In summary, Jak/Bondi criteria take a more conservative (versus Winblad) neuropsychological approach to defining MCI, resulting in fewer, more impaired ‘cases’ of MCI, with a greater risk of dementia (i.e., more specific for predicting progression). Conversely, Jak/Bondi criteria may be more prone to miss subtle cognitive deficits (i.e., less sensitive for predicting progression [46]). The use of Jak/Bondi criteria to rule out MCI thus appears to capture an SCD phenotype with worse objective cognition/prognosis.

Whilst Jak/Bondi and Winblad criteria take different approaches to operationalizing MCI, they are not mutually exclusive, resulting in a degree of ‘overlap’ between the current SCD samples. An alternative would be to categorize patients into three non-overlapping groups: those with SCD irrespective of MCI criteria (SCDWinblad/SCDJak/Bondi), and those with SCD under one criteria but MCI under the other (i.e., SCDWinblad/MCIJak/Bondi and MCIWinblad/SCDJak/Bondi). This approach may improve prognostic predictions. Unfortunately, this study lacked statistical power to investigate this empirically. In any case, this may be a primarily research-oriented question, as most clinical settings only utilize one type of MCI criteria.

The prognostic implications of different approaches to excluding MCI during the ascertainment of SCD have previously been discussed [11, 43], but empirical data have been lacking. Nevertheless, there is increasing recognition that SCD may not be synonymous with entirely ‘normal’ objective cognition, and that minor neuropsychological deficits have prognostic value in SCD. Using DELCODE data, Wolfsgruber et al. [9] demonstrated that, at the group level, patients with SCD have minor neuropsychological deficits (approximately 0.25–0.5 SD in magnitude) versus controls. The same group recently showed that SCD patients with (versus without) minor neuropsychological deficits had a faster cognitive decline and increased risk of MCI [10]. The DELCODE investigators operationalized MCI as a deficit of at least 1.5 SD on any test – an approach comparable to that used for the SCDWinblad sample in the current study. In spite that the SCD sample in DELCODE most closely aligns with the sample with better cognition/prognosis in the current study, those DELCODE SCD participants with subtle cognitive deficits continued to have a worse prognosis. In summary, despite that objective cognition is unimpaired in SCD, variation in scores (comfortably within the normal range) is linked to prognosis, both at the between-individual and between-sample levels. Whilst this suggests that the neuropsychological cut-offs employed by MCI criteria miss subtle – yet prognostically meaningful – cognitive deficits, more sensitive/thorough tests may be required to capture them, which are not available in all clinical settings.

The finding that excluding MCI via more stringent criteria yielded an SCD sample with worse cognition/prognosis is, arguably, unsurprising. Nevertheless, this finding remains important; there is significant heterogeneity in dementia incidence rates across the SCD literature [5, 47], and recent critiques have questioned the prognostic value of an SCD ‘diagnosis’ [48, 49]. Attempts to explain the heterogeneous dementia incidence rates in SCD have generally been unsuccessful. A recent meta-analysis evaluated numerous potential moderators of dementia incidence in SCD studies (including how SCD is defined, demographic/genetic factors, recruitment source, and follow-up duration) [6], but no significant moderators were identified (note: statistical power may have been lacking). Interestingly, neither the type of criteria used to exclude MCI, nor objective cognition more generally, were explored as candidate moderators. Whilst, in clinical practice, prognostic evaluation is primarily informed by individual patient characteristics, the present work suggests that the particular approach used to define MCI in a given clinical setting has important prognostic implications for individuals with SCD [4].

Reviewing the operationalization of SCD in research settings for the SCD-I, Molinuevo et al. [11] encouraged greater harmonization of SCD characterization across studies, to facilitate comparisons and evidence synthesis. Nevertheless, the authors acknowledged that variation in how SCD is defined has advanced scientific understanding in the field. Indeed, the authors did not recommend specific MCI criteria for SCD studies, because: the SCD/MCI boundary may not be clearcut, neuropsychological batteries vary across settings, and the choice of ‘liberal’ or ‘conservative’ MCI criteria depends on research questions. We echo the recommendations of the SCD-I: the optimum approach taken to exclude MCI in future SCD studies will likely depend on the research question. For example, in clinical trials targeting AD, the Jak/Bondi criteria may be optimal for excluding MCI (in order to ‘enrich’ the sample for preclinical AD and maximize statistical power [50]).

The current work has a number of strengths. The present use of routinely-collected clinical data is no doubt a strength, given dementia risk is elevated in clinical, but not community-dwelling SCD populations [7]. This also increases the generalizability of the work to settings which assess and provide prognosis to individuals with cognitive symptoms. Moreover, all patients in the study underwent a thorough medical, neurological and psychiatric evaluation, and were assessed using a range of neuropsychological tests. This thorough characterization enabled the derivation of two samples fulfilling SCD criteria, differing only in the neuropsychological criteria used to operationalize (i.e., exclude) MCI. Furthermore, patients were followed-up for an average of 3 years, enabling dementia incidence rates to be estimated. The diagnosis of dementia (including subtype) was made by consensus, which improves diagnostic accuracy compared to individual clinicians [51]. Lastly, the calculation of the incidence rate ratio made use of Mantel-Haenszel methods which enabled adjustment for the confounding effect of age.

This study also has limitations. To derive adequate samples of SCD patients, we used all available data (collected over three decades). Jutten et al. [52] discuss how ‘average performance’ on cognitive tests can drift over time, rendering norms outdated. Our reclassification of patients according to ‘fixed’ cognitive norms may thus have delineated subtly different SCD phenotypes at different times (although the temporal distribution of assessments appeared comparable between samples – see Table 2). This limitation partly reflects that the data were collected during routine healthcare at a single center [15]. Efforts to replicate the current findings in data collected within tighter timeframes (perhaps using prospective multicenter designs) are thus welcomed. The administration of TMT in this study differed from the typical procedure – here, patients were discontinued at two errors, whereas in other settings, assessors correct all errors without discontinuing patients (though the potential impact of this on results appears limited; see supplementary Discussion note 1). The MCI criteria utilized in this study vary in the precision with which they define cognitive impairment. The Winblad criteria require ‘impairment on objective cognitive tasks’, while the Jak/Bondi criteria feature specific cutoffs according to population norms. Both samples had a very high proportion of white individuals; whilst this is in-keeping with the characteristics of the population served by the memory clinic (see supplementary Discussion note 2), further research in samples with better representation of other ethnic groups will be required to assess the extent to which the current findings generalize in other populations. Further, the normative data used to calculate z-scores for each domain did not adjust for identical demographic factors; for verbal fluency and psychomotor speed/executive function, age- and education-adjusted norms were used, whereas the available norms for episodic memory only adjusted for age. Nor did we have access to ethnically and culturally appropriate test norms; a previous study reported that using ‘combined’ ethnicity norms can result in misclassifying scores for some groups [53]. In the interests of standardization and replicability, future SCD investigators are encouraged to characterize samples using ethnically and educationally appropriate norms, and to employ MCI criteria with precise cutoffs. Furthermore, the available tests permitted only three cognitive domains to be operationalized for Jak/Bondi criteria; given one of the ‘routes’ to diagnosis of MCIJak/Bondi is one impaired score within each assessed domain, this could have led to overdiagnosis of MCI, though Bondi et al. [14] also utilized three cognitive domains in their validation study. Considering the specific measures used to characterize cognitive domains, the tests used by Bondi et al. [14] for the language and episodic memory domains appeared less correlated than in the present study (see supplementary Discussion note 3). This could have resulted in a degree of diagnostic misclassification in the current study; future investigators are encouraged to utilize a greater number/range of cognitive domains, as well as sufficiently independent constituent tests, in order to mitigate this possibility. The defining characteristic of SCD is a self-experienced decline in cognitive capacity [4]. A retrospective case note review was required to ascertain the presence/absence of this feature for some patients, and this could not be determined in a small number, resulting in their exclusion. More generally, we did not have access to a standardized measure of subjective cognition (e.g., the SCD interview [54]), which precluded a comparison of subjective cognitive profiles between samples. This could be a fruitful topic for future investigations. Lastly, there is increasing interest in the role of biomarkers for prognostication in SCD [55]. Unfortunately, a lack of comprehensive biomarker data prevented the characterization of SCD according to (e.g.) the ATN framework [56]; we recommend that future projects include biomarker data where possible.