This study involved community-dwelling older persons recruited between March 2022 and September 2024, as part of a nationally funded project in Singapore aimed at developing artificial intelligence tools to detect early cognitive impairment in the community (Project PENSIEVE) [44]. Community-dwelling individuals were invited to participate if they met the following criteria: (1) higher risk of cognitive impairment (ie, aged ≥65 y [26]) and had at least one of the 3 chronic diseases (ie, diabetes mellitus, hypertension, or hyperlipidemia); this criterion was included a priori to focus on individuals with at least some risk of cognitive impairment, in line with current literature suggesting the limited benefit of screening among low-risk individuals) [26,45]; (2) ability to follow simple instructions in English or Mandarin (due to limitations in the available assessment language); and (3) presence of an informant (eg, family member or friend) who knew the participants well. Individuals were excluded if they had severe visual impairment that would affect their ability to complete neuropsychological assessments (to ensure generalizability, participants were included as long as they could see pictures on a piece of paper held in front of them). No participants were excluded for reasons related to missing data, as we implemented strict data collection procedures (eg, mandatory data field) and routine data audits throughout the study.

Sources of recruitment included 14 community roadshows conducted by the study team, clients of our community partners, home visits by community volunteers who partnered with us, media publicity (radio, online articles, and posters), and word-of-mouth referrals from participants who had completed research assessments. To ensure the recruited samples were representative of the community, the study’s publicity materials emphasized the key message of “detect dementia early” (along with direct referrals to memory clinics in the event of significant findings), rather than the conventional invitation to participate in research (which may inadvertently attract a distinct group of individuals). Examples of these publicity materials (eg, study banner, poster, and brochure) are provided in Method S1 in Multimedia Appendix 1.

All participants received comprehensive assessments, which included semistructured interviews with both participants and their informants, detailed neuropsychological testing, and observational notes of participants’ behavior during assessments. Full descriptions of the comprehensive assessments are available in Method S2 in Multimedia Appendix 1, with further details on each assessment tool provided in Method S3 in Multimedia Appendix 1. Diagnoses of MCI and dementia were determined via consensus conference by 3 dementia specialists. Dementia was diagnosed using the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, criteria [46], which require the presence of cognitive concerns (reported by the individual or a knowledgeable informant), impairment in instrumental activities of daily living (eg, managing money or medications), and objective cognitive deficits. MCI was diagnosed using the modified Petersen criteria [47], which require the presence of cognitive concerns (reported by the individual or a knowledgeable informant), absence of impairment in instrumental activities of daily living, and the presence of objective cognitive deficits. Normal cognition was diagnosed when participants did not meet criteria for dementia or MCI.

AD8 [30] is an 8-item, informant-based questionnaire that assesses changes in a participant’s cognition and function over the past few years. For each item, the informant rates whether there has been a change in the participant: 1=yes (a change) and 0=no (no change) or don’t know. Responses to the 8 items are summed to provide a total score, with higher scores reflecting greater cognitive problems. AD8 has been shown to be useful in detecting varying severities of cognitive impairment [37,38]. Informant AD8 has also been previously validated in Singapore [48].

To derive an ultra-brief questionnaire, the study team focused on candidate questions that assess the two key criteria for diagnosing MCI/dementia [46,47]: (1) the presence of cognitive concerns as reported by the individual or a knowledgeable informant; and (2) impairment in instrumental activities of daily living (iADL). The presence of cognitive concerns was evaluated using validated questions related to subjective cognitive decline (SCD), by asking participants or informants: “Do you feel like your (or your family member’s) memory is becoming worse?” and “Are you worried that your (or your family member’s) memory is becoming worse?” These 2 sets of questions have been validated in previous studies [49,50] and have been shown to be useful for capturing early symptoms of cognitive decline [51-55].

Impairment in iADL was assessed using the locally validated modified Lawton scale [56], with informants asked about difficulties in various domains (public commuting, grocery shopping, managing money, using the telephone, taking medications, preparing meals, doing housework, and doing laundry). The original responses on the modified Lawton scale included 4 options (ie, unable to do at all, needs some help, needs no help, and never needed to do this), which were collapsed in this study into two options: (1) yes and (2) no/never needed to do this.

In total, 15 candidate questions were considered for the ultra-brief questionnaire: 12 cognition-related items and 3 basic demographic variables (eg, age, sex, and years of education), given their potential correlation with cognition. The exact wording and response options for all 15 items are provided in Table 1.

The study sample was randomly split into a 70% training sample and a 30% test sample. The training sample was used to develop an ultra-brief questionnaire that best distinguished MCI/dementia from normal cognition, whereas the test sample was used to evaluate the actual performance of this questionnaire.

In the training sample, a best-subset approach [57] with 5-fold cross-validation was used to select the optimal combination of items from the 15 candidate questions. The best-subset approach is an efficient, computationally intensive method for variable selection [4,20,58,59] in which logistic regression is used to exhaustively evaluate all possible combinations of the candidate questions, identifying models with the lowest prediction errors. Five-fold cross-validation was then used to select the most parsimonious model within 1 SE of the best-performing model. After identifying the optimal model from the best-subset approach, we further refined the model by considering potential inclusions of quadratic terms for continuous variables (eg, age and years of education); quadratic terms with P<.05 were incorporated into the final model to improve fit. The final selected model was then applied to the test sample to generate predicted probabilities of cognitive impairment (ie, MCI/dementia), with the model variables constituting the new ultra-brief questionnaire.

Predicted probabilities were computed from logistic regression using the following equation:

Predicted probability=e Logit1 + e Logit

where

with each β representing the regression coefficient for its respective variable in the model.

In the test sample, the predicted probabilities were used to compute the area under the receiver operating characteristic curve (AUC), thereby assessing the actual performance of the ultra-brief questionnaire in discriminating MCI/dementia from normal cognition. In general, an AUC of 0.7 to 0.8 is considered acceptable discrimination, 0.8 to 0.9 is considered excellent, and more than 0.9 is considered outstanding [60]. Comparisons of AUC were conducted using the nonparametric method proposed by DeLong et al [33,59,61,62], with analyses stratified by education subgroups (ie, ≤10 and >10 y of education based on median split). A 2-cutoff approach [63-67] was adopted for the ultra-brief questionnaire. The first cutoff was chosen to yield high sensitivity and negative predictive value (>80% each) and was used to rule out MCI/dementia (ie, when probability scores fell below the first threshold). The second cutoff was selected for high specificity and positive predictive value (>80% each), identifying those very likely to have MCI/dementia. This 2-cutoff approach is recommended in recent literature [67], as it enhances the performance of cognitive assessment tools [63-66], reduces the effects of prevalence on tool performance [64], and allows prioritization of scarce health care resources for individuals who truly require further cognitive assessments [63].

As a secondary analysis, the performance of the ultra-brief questionnaire was also evaluated for distinguishing dementia from nondementia. Additionally, 2 sensitivity analyses were conducted in the test sample to evaluate the robustness of results when the prevalence of MCI/dementia was readjusted to reflect the average prevalence in most communities:

The best-subset approach was performed with the “bestglm” [57] package in R (version 4.4.0; R Foundation for Statistical Computing). All remaining analyses were conducted in Stata (version 18; StataCorp LLC). No a priori sample size calculation was performed; the final sample size was determined pragmatically based on participants recruited between March 2022 and September 2024. Post hoc power analyses confirmed that the test sample provided robust power (90%) to distinguish MCI/dementia from normal cognition (α=.05, two-sided test). Power was also reasonably sufficient for participants with ≤10 years of education (72%) but was limited in those with >10 years of education (19%) due to fewer positive cases. The power calculations were conducted using PASS software (version 15.0.5; NCSS, LLC) and the Hanley and McNeil formula [74], with further details provided in MethodS4 in Multimedia Appendix 1.

The study received ethical approval from the SingHealth Centralized IRB of Singapore (reference number: 2021/2590). Informed consent was obtained from all participants. Before obtaining informed consent, the mental capacity of participants was briefly assessed in accordance with the Mental Capacity Act of Singapore [75]. In the event of concerns about mental capacity, informed consent by proxy was then obtained from the legally authorized next-of-kin. Participants who completed the research assessments received Singapore Dollar $80 as compensation for their time, inconvenience and transportation costs.

A total of 1856 participants were included, of whom 255 (13.7%) had MCI/dementia. Participant characteristics are presented in Table 2, with a median age of 72 years and a median education of 10 years (range 0-23 y). Corresponding informant characteristics are provided in Table S1 in Multimedia Appendix 1, with informants primarily comprising spouses (897/1856, 48.3%) and children (506/1856, 27.3%). The sample was randomly split into a training sample (1299/1856, 70%) and a test sample (557/1856, 30%). Table S2 in Multimedia Appendix 1 shows that the training and test samples had comparable demographic characteristics (P>.05 across covariates).

Table 3 presents the top models identified through the exhaustive search method in the training sample (n=1299). iADL–medications and worry about cognition–informant were among the most useful items in detecting MCI/dementia, whereas iADL–meals and iADL–grocery were among the least useful. Following 5-fold cross-validation, the model with 4 items was identified as the most parsimonious among the top models (Figure S1 in Multimedia Appendix 1). This model included 2 informant questions (iADL–medications and worry about cognition–informant), plus 2 demographic variables (age and years of education). These items were then selected to constitute the new ultra-brief questionnaire, henceforth denoted as IQ2+ (Informant Questionnaire for cognitive impairment–2 items plus demographics). Further model refinement considered the inclusion of quadratic terms for age and years of education, resulting in the addition of a quadratic term for years of education (P=.004) in the final model. Thus, the final IQ2+ model comprised five predictors: iADL–medications, worry about cognition–informant, age, years of education, and the quadratic term for years of education. Responses from the IQ2+ questionnaire can be converted to predicted probabilities using an interactive web-based calculator we have created [76]. A screenshot of the calculator is shown in Figure 1.

Table 4 presents the AUC results for IQ2+ and AD8 in the test sample (n=557). Predicted probabilities from the new IQ2+ demonstrated excellent performance in distinguishing MCI/dementia from normal cognition (AUC 85.3%, 95% CI 80.4%‐90.2%), which was significantly better (P=.003) than a nested baseline model (comprising age, years of education, and the quadratic term of years of education; AUC 78.0%, 95% CI 72.6%‐83.4%). In contrast, AD8 had an AUC of 76.1% (95% CI 69.6%‐82.6%), which was not significantly different (P>.99) from that of the baseline model. For the detection of dementia, both IQ2+ and AD8 had AUCs >90%, which were significantly higher (P<.05) than the baseline model (83.2%). AUC results remained largely similar across education subgroups and in the 2 sensitivity analyses where the prevalence of MCI/dementia was increased to reflect average prevalence in most communities (ie, 20% [68-73] and 35% [69,70,72,73], respectively).

Test statistics for IQ2+ are plotted in Figure 2A. Adopting a 2-cutoff approach, the lower cutoff (probability ≥12%) had 80.3% sensitivity and 96.0% negative predictive value and was used to rule out MCI/dementia (for individuals with probability scores below the cutoff), whereas the upper cutoff (probability ≥52%) had 99.0% specificity and 80.8% positive predictive value and identified those likely to have MCI/dementia (ie, to rule in MCI/dementia). These 2 cutoffs provide an intermediate range between them (grayed area in Figure 2A), identifying those who warrant further assessment. The optimal cutoffs varied slightly with changing prevalence of MCI/dementia, as shown in Figure 2B and 2C. Detailed results on test statistics for IQ2+ are also available in Tables S3–S5 in Multimedia Appendix 1.

The 2 cutoffs of IQ2+ effectively identified three risk categories for cognitive impairment: (1) less likely to have cognitive impairment, (2) higher risk of cognitive impairment, and (3) likely to have cognitive impairment. These 3 categories, along with their cross-tabulation with the final diagnoses, are presented in Table 5. In the first category (ie, less likely to have cognitive impairment), 88% to 96% of individuals had normal cognition. In the second category (ie, higher risk of cognitive impairment), 25% to 45% of individuals were diagnosed with MCI. In the third category (ie, likely to have cognitive impairment), 81% to 84% of the individuals had MCI/dementia, with a large proportion having dementia (47%-54%). Distinctions between these 3 risk categories of IQ2+ are also visible in the box plots in Figure 3A–3C. The first category (white region with probability scores below the lower cutoff) identified those with normal cognition, the third category (dark gray region with probability scores above the upper cutoff) identified almost all individuals with dementia, and the second category (light gray region between the lower and upper cutoffs) mostly captured those with MCI. In contrast, AD8 showed poorer discrimination between normal cognition and MCI, with discernible floor effects in the normal cognition and MCI groups as seen in the box plots in Figure 3D–3F. This is also reflected in the test statistics of AD8 (as presented in Tables S6–S8 in Multimedia Appendix 1), which demonstrated low sensitivity across its cutoff scores (with maximum sensitivity of 69.6%‐73.4%).

In this study, we developed and validated an ultra-brief informant questionnaire (IQ2+) for the detection of cognitive impairment in literacy-diverse communities. Using a rigorous best-subset approach with cross-validation, we identified a parsimonious 4-item model that was most useful in detecting MCI/dementia, comprising 2 informant-based questions (assistance with medications and worry about cognition), along with age and years of education. In the independent test sample, IQ2+ achieved an AUC of 85.3%, outperforming the widely used AD8 (AUC=76.1%) and a baseline demographic model (AUC=78.0%). In contrast, AD8 was not significantly better than the baseline demographic model. The robust performance of IQ2+ was consistent across education subgroups and under varying prevalence scenarios, highlighting its potential utility for case finding in diverse community settings. Adopting 2 optimal cutoffs [63-67], IQ2+ demonstrated high sensitivity, specificity, negative predictive value, and positive predictive value.

Despite being widely recommended [26,39-42], the AD8 performed poorly in detecting MCI/dementia and was not substantially better than a basic demographic model comprising age and education. This finding aligns with recent literature, which demonstrates AD8’s lower performance for detecting early cognitive impairment (AUC 61%‐69%) compared to its performance in clear-cut dementia cases (AUC 89%‐93%) [48,77]. AD8 was originally developed for the purpose of detecting dementia [30], and its items tend to focus on identifying conspicuous changes in iADL (ie, hallmarks of the onset of dementia) [46]. As a result, AD8 may possibly not be as sensitive or consistent in detecting subtle changes associated with early cognitive impairment [30,78]. In contrast, the 2 informant questions in IQ2+ may plausibly be more sensitive to early cognitive changes. Worry about cognition—especially when reported by a knowledgeable informant—has been shown to be an early symptom of cognitive decline [51]. This item requires informants to compare current abilities with previous premorbid abilities, making it more likely to detect subtle cognitive decline over time. It also prompts informants to apply a “threshold” to decide whether they are worried about the cognitive decline, which serves as a filter to identify individuals with meaningful changes. Similarly, the ability to manage one’s medications is a complex task that possibly involves multiple cognitive domains [16], such as language (for understanding medication instructions), executive function (for scheduling medication intake and problem-solving when medications are not taken on schedule), and memory (for tracking medication intake). A subtle decline in any of these domains may potentially manifest as increasing difficulty in managing medications independently, serving as an early symptom of cognitive decline [79].

At the population level, IQ2+ can serve as a quick and efficient risk stratification tool, facilitating appropriate triage for further assessment. As summarized in Table 6, low-risk individuals (<12% probability of MCI/dementia) may be reassured and advised to repeat the test in 3 to 5 years or if circumstances change (eg, appearance of new symptoms). Intermediate-risk individuals (~25%‐45% probability of MCI/dementia) can be directed to further assessments using brief cognitive tests [4,32-34] to provide more conclusive evidence of cognitive impairment [27]. High-risk individuals (>80% probability of MCI/dementia) may benefit from direct referral to memory clinics for further clinical management. This approach, as summarized in Table 6, aligns with the 2-stage strategy for active case finding of cognitive impairment (ie, subjective reports, followed by brief cognitive tests) [27,28]. It offers a scalable approach to case finding in large populations by conserving resources for cognitive testing and reserving them for individuals who truly need further verification [27]. To ensure that IQ2+ remains useful in diverse populations, its optimal cutoffs can also be adjusted depending on the prevalence of MCI/dementia in different populations (as shown in Figure 1), thus providing a more tailored solution for identifying individuals at varying risk levels.

Essentially, IQ2+ may have 2 plausible use cases in the community. First, it allows family members who have concerns about a loved one’s cognition to complete the questionnaire online (accessible at [76]), providing an immediate result with risk score and a brief interpretation of its implications. The results can also be downloaded as a PDF file for further discussion with health care providers. This approach mirrors a well-established practice in the field of diabetes mellitus, where the public is encouraged to complete an online risk test [80] as an initial screening to determine the need for further diagnostic tests [81]. By enabling self-appraisal, this use case of IQ2+ leverages family members for efficient case finding and empowers them to proactively support the brain health of older persons.

Second, IQ2+ can also be used by health care workers who routinely conduct comprehensive geriatric assessments in the community. Since 2019, the WHO has advocated for routine assessment of intrinsic capacity among community-dwelling older persons, which requires comprehensive evaluations of cognition, mobility, nutrition, hearing, vision, and mood [29]. Given the extensive range of components to cover, it can be challenging to include routine cognitive testing in all geriatric assessments. IQ2+, which can be completed in <1 minute, is well-suited to be embedded within initial geriatric assessments to prioritize individuals who require further cognitive testing. Given its excellent performance across education subgroups (Table 4), IQ2+ offers a practical tool that may potentially be broadly implemented in diverse populations, including in underserved populations in LMICs, which currently have the largest number of undiagnosed cognitive impairment [82]. This is also in line with the 2024 Lancet Commission’s call to address the unmet need for cognitive screening tools that are also suited for individuals with lower literacy in LMICs [35].

Several limitations are notable. First, IQ2+ was developed for older individuals aged ≥65 years. It is unclear whether IQ2+ would be useful for individuals aged <65 years. Second, IQ2+ would benefit from further validations in other cultures and languages, as cultural differences may affect how informants respond to its items and could potentially modify its performance in different settings. Third, IQ2+ requires the use of a web-based calculator to determine the probability of cognitive impairment. This approach leverages technology to automate test scoring, result visualization, and interpretation. However, it may limit accessibility compared to pen-and-paper tests that use raw scores as a cutoff (eg, AD8), particularly in settings with limited access to technology. Fourth, one of the core questions of IQ2+ involves difficulty managing medications. This item was identified as the most useful item in our exhaustive search (Table 3) and has some face validity (as it detects subtle changes in a complex task of managing medications [16,79], and the need to take regular medications may also be a proxy for higher risk due to the presence of chronic diseases). However, this question may be less useful among healthier older individuals who do not need to take medications regularly. Fifth, although IQ2+ can be a useful case finding tool, it is not intended to replace comprehensive clinical and neuropsychological assessments, which provide more definitive diagnoses as well as granular information on specific cognitive deficits [20,83,84].

Using rigorous methodology, this study developed an ultra-brief questionnaire that enables untrained laypersons to screen for cognitive impairment in <1 minute. Despite its brevity, the questionnaire demonstrated excellent performance in detecting MCI/dementia, outperforming the well-established AD8. The questionnaire can be completed by members of the public who have concerns about a family member’s cognition (accessible at [76]) or embedded within community geriatric assessments to prioritize cognitive testing. Its excellent performance was consistent across education subgroups and varying prevalence scenarios, supporting its utility for case finding in diverse populations, including underserved communities and LMICs.