A multistage random sample of 7501 households with at least 1 older adult resident (Singapore citizen or permanent resident) was purchased from the Department of Statistics of Singapore representing approximately 2.2% of all households with an older adult resident in Singapore in 2022. Household addresses were sampled across all 5 geographic regions of Singapore (Central, East, North, Northeast, and West). To increase the participant count and precision of statistical estimates within minority subgroups, we oversampled addresses with predominant (1) Malay and (2) Indian residents, and with (3) at least 1 resident aged 75 years or older by a factor of 2.

Individuals residing in the random sample of households were eligible for inclusion in the TARGET cohort if they were (1) a Singapore citizen or permanent resident, (2) aged 60‐105 years, (3) spoke at least 1 of the 4 official languages in Singapore (English, Mandarin [or a Chinese dialect], Malay, and Tamil), (4) ambulatory with or without mobility aids, (5) able to see with or without glasses, and (6) able to hear with or without hearing aids. Individuals with at least one of the following criteria at the time of recruitment were excluded: (1) possible delirium or dementia defined as an Abbreviated Mental Test score <6 [48]; (2) blood pressure outside the normal range (defined as ≥200/110 mm Hg or ≥180/100 mm Hg and heart rate >120 per minute or ≤90/60 mm Hg); (3) any symptoms of chest discomfort, breathlessness, dizziness, or profuse sweating; or (4) any acute cardiac event in the past 30 days.

An invitation letter was mailed to each household in the random sample stating that participation in the study was completely voluntary, and the study team could be contacted to opt out of the study. Trained field interviewers then conducted home visits 2‐4 weeks later to households that did not opt out. All willing potential participants within each household were first screened to determine their eligibility against the inclusion and exclusion criteria. Multiple residents from the same household were able to participate if they independently met the eligibility criteria. Once participants consented to the study, they were able to complete the baseline assessment on the same day, if convenient, or reschedule the assessment for a later date. Interviewers made up to 5 visits at differing times of day to each residence to recruit participants. Households, which remained uncontactable after 5 visits, were excluded due to nonresponse.

At least 1 participant was recruited from 1674 households, giving a household recruitment rate of 22.3% (1674/7501). The household participant yield was 1.3 participants per household (range: 1-4). Anecdotally, the field interviewers reported that the households approached were generally reluctant to interact with strangers, coinciding with active national scam prevention campaigns targeting older adults in Singapore during the recruitment period. Demographic characteristics were compared between household responders and nonresponders. There was no significant difference in the age group of the household reference person (primary resident) between household responders and nonresponders. However, differences in housing type and region were identified. The proportion of private households (ie, residential properties that are not subsidized, built, and managed by the Singapore government) was significantly higher among nonresponders. Response rates also significantly differed by geographic region (Figure 1). The highest response was within the Northeast Region (577/1516, 38.1%) while the West Region (215/1549, 13.9%) yielded the lowest response rate.

Table 1 summarizes key measures, variables, and outcomes. Both primary and secondary data sources were used. Primary data collected included epidemiological survey data (eg, sociodemographic, anthropometric, cognitive, psychosocial, physical, and environmental data), biomechanical data, clinical imaging, and spatial performance data to provide a holistic understanding of fall and fracture risk factors. Additionally, secondary data in the form of electronic medical records (EMRs) containing data on health status, health care use, prescriptions, procedures, and mortality status are being linked from the Singapore Ministry of Health’s TRUST (Trusted Research and Real-World Data Utilisation and Sharing Tech) platform [49] for participants who have consented to this linkage.

A face-to-face computer-assisted personal interview was conducted with each participant at their residence or other preferred place of convenience using a structured questionnaire to capture sociodemographic, environmental, physical, sensory, medical, cognitive, and psychosocial characteristics related to fall and fracture risk at baseline. Anthropometric measurements (height, weight, and bilateral hand grip strength) were also taken. Responses were collected using the survey software Qualtrics (Qualtrics). The survey was designed by TARGET study investigators and administered by trained field interviewers in the participant’s preferred language (English, Mandarin or Chinese dialect, Malay, or Tamil). The baseline interview took approximately 2 hours to complete, including the screener, consent-taking, and gait assessment detailed in the following text.

As a component of the home-based interview, all participants performed a 5-minute walk inside or outside their home, in an open physical space, with the freedom to walk at their preferred pace and navigate turns, while wearing 6 ZurichMOVE inertial measurement units (IMUs) motion sensors [65] affixed to designated anatomical locations (feet, wrists, trunk, and head; Figure 2).

There has been growing research and clinical interest in the use of wearable sensors for identifying fall risk through biomechanical data, specifically, dynamic gait signatures [37,66,67]. Gait and balance assessments are a core component of fall risk assessments in clinical practice. However, existing gold standard assessments such as the Timed Up and Go test and Berg Balance Scale have shown poor predictive accuracy (pooled sensitivity ranging from 0.25 to 0.32) in predicting prospective falls [68,69]. As the majority of falls occur while walking, dynamic gait characteristics such as step length, step time, variability, and asymmetry need to be factored into fall risk assessments for accurate risk prediction [66,70]. Wearable sensor systems can objectively measure a wide range of dynamic gait characteristics, which can then be used to train predictive models for fall risk assessment [71]. IMU sensors, which acquire data using accelerometers and gyroscopes, have gained research and clinical interest as a possible alternative to the current gold standard optical motion capture systems for gait analysis [72]. Compared with optical motion capture, IMU sensors are compact, inexpensive, easier to operate, and can be combined with functional walking tests for gait assessment in the home environment [73].

The IMUs embedded within ZurichMOVE are time-synchronized and have a dynamic range of ±16 g for the accelerometer and 2000° per second for the gyroscope at a sampling frequency of 200 Hz and have been shown to reliably assess gait and identify gait differences in healthy and neurological populations across laboratory settings and real-world environments [74,75]. Triaxial acceleration and angular velocity signals recorded during the walk test are first filtered and analyzed with the validated gait event detection algorithm [70,71], identifying key phases in the gait cycles over the 5 minutes. From these phases, key spatiotemporal parameters such as stride time, length, cadence, and double-limb support time were computed for every participant. Furthermore, intercycle variations are evaluated during the entire walking trial to capture fluctuations and variability in movement patterns, indices associated with deficits in gait, balance, and risk of falling [37,76,77].

After the home-based interview, participants who reported that they did not have bilateral hip replacement were invited to a clinic visit at the Singapore Eye Research Institute for a dual-energy X-ray absorptiometry (DXA) scan. Each participant underwent 3 DXA scans (hip, lumbar spine, and whole-body) within a 20‐ to 30-minute session using the Hologic Horizon W scanner (Hologic, Inc). The hip and lumbar spine scans were used to assess areal bone mineral density (aBMD), which was then used to calculate the T-score to evaluate osteoporotic status. Participants with a T-score of ≤−2.5 were classified as osteoporotic, those with a T-score between −2.5 and −1 were classified as osteopenic, and those with a T-score of >−1 were classified as normal, while the whole-body DXA scan captured clinical anthropometric data and body composition, including measurements of fat mass, lean mass, and total mass across the trunk, limbs, and head. These data were used to categorize participants into the low muscle mass group, defined as appendicular lean mass/height² of <7 kg/m² for males and <5.4 kg/m² for females [78]. Additionally, a subset of participants completed 3D-optical body shape scans using the TC2-19M scanner (TC2 Labs LLC), wearing tight-fitting clothing (Spandex top and pants). Participants held an A-line stance for approximately 9 seconds during scanning, which generated a triangular mesh representing their body shape. Data from both the whole-body DXA and 3D optical scans were used to estimate trochanteric soft tissue thickness (TSTT) [79]. TSTT refers to the thickness of the soft tissues, including skin, fat, and muscle, over the greater trochanter of the femur. Clinical risk factors were also gathered, and the risk for major osteoporotic fractures (MOFs) and hip fractures was assessed using the Fracture Risk Assessment Tool (FRAX) without aBMD for the entire cohort. For those who had a DXA scan, FRAX was calculated with the aBMD from the femoral neck. A high risk for MOF was assigned if FRAX-MOF ≥9%, and a high risk for hip fractures was assigned if the FRAX–Hip Fracture (FRAX-HF) score is ≥2% [80]. Age- and sex-adjusted analyses were conducted for all variables, except FRAX-related variables, as age and sex are already incorporated into the FRAX calculation algorithm.

Despite widespread use, studies have shown that aBMD measurements from DXA scans and the 10-year fracture probabilities of hip or MOFs computed using FRAX [61] often fail to detect individuals at risk, as only between 28% and 61% of hip fractures occur in individuals with osteoporosis [62,81,82]. Although FRAX incorporates 12 risk factors and provides country- and ethnicity-specific scores, research suggests that FRAX may be less accurate in Asian populations, compared with Caucasians [63,64]. Finite element models (FEMs) have demonstrated strong potential for improving fracture risk prediction, especially given the limited sensitivity of current standard-of-care methods. Subject-specific FEMs based on computed tomographic scans provide direct estimates of femoral strength and have outperformed aBMD in predicting future fractures, with 1 recent study demonstrating this superiority over a follow-up period of up to 16 years [83-85]. Emerging research from Caucasian cohorts also suggests that FEM-based strength estimates derived from DXA outperform aBMD alone in identifying individuals with hip fractures [86,87]. This body of evidence paves the way for FEM-based fracture risk assessment, particularly when adapted for use with clinical DXA scans for primary screening and with tomographic scans for opportunistic screening, as a more accurate alternative to current methods.

Participants who completed the DXA scan were also invited to play a virtual reality (VR) spatial navigation game during the clinic visit to assess their cognitive ability. The SPACE (Spatial Performance Assessment for Cognitive Evaluation) is a novel iPad serious game designed to identify deficits in spatial navigation indicative of cognitive impairment and potential fall risk [88]. SPACE has been validated with participants across community and clinical cohorts in Singapore. Studies revealed that SPACE is well suited for older adults [88], and its performance metrics correlate with scores from the Montreal Cognitive Assessment (MoCA); predict hippocampal atrophy beyond age, education, and standard neuropsychological tests [89]; and distinguish between dementia severity levels with high accuracy [90].

Studies have highlighted the potential application of VR technology for identifying markers of cognitive impairment and fall risk [88]. While analysis of cerebrospinal fluid and positron emission tomography can reliably detect dementia-related pathophysiological changes, these methods are invasive and costly, which limits their use for population-based screening [91]. Blood biomarkers are a promising alternative, but these are still undergoing clinical validation and are not yet widely available for routine diagnostic use [92]. In contrast, neuropsychological tests, such as the Mini Mental State Examination and MoCA, are relatively easier to deploy but have shown limited sensitivity to detect early cognitive impairment [93,94]. Critically, most cognitive assessments do not include tests that assess deficits in spatial navigation ability. This lacuna is surprising since spatial disorientation is one of the earliest signs of declining cognitive ability. Indeed, the hippocampus and the entorhinal cortex, which play an instrumental role in navigation, are also among the earliest brain regions to be affected by Alzheimer disease pathophysiology [95-98]. VR technology now offers an accurate, cost-effective, and safe method to assess navigation deficits indicative of cognitive impairment and fall risk in a quick and portable manner [88,99].

In SPACE, participants assume the role of an astronaut tasked with exploring an unfamiliar planet by completing a series of spatial navigation tasks with the help of a companion robot. Each task is specifically designed to probe different aspects of their spatial navigation abilities. SPACE includes a detailed training phase in which players can familiarize themselves with the controls while being assessed on basic visuospatial and motor skills. After the initial training phase, participants perform 5 spatial tasks (ie, path integration, egocentric pointing, mapping, associative memory, and perspective taking) designed to recruit critical brain regions involved in spatial navigation. During gameplay, both performance data and metadata (eg, position and orientation, time, coordinates, and screen interaction) are stored and used to produce individual scores for the various tasks [88].

Prospective follow-up of the TARGET cohort is currently ongoing. The cohort is being followed up via telephone every 3 months from the baseline interview for 2 years to determine (1) whether they have experienced a fall or fall-related injury over the past quarter, and (2) their cognitive status every 6 months measured using an abbreviated version of the MoCA designed for telephone administration (T-MoCA [51]). Interviewers may also visit participants to conduct the follow-up assessments in person if the participant has difficulties communicating clearly over the phone. As of May 24, 2026, T₆ (18-month follow-up) is complete for all participants, with 80.5% (1844/2291) successfully followed up (48/2291, 2.1% deaths, and 399/2291, 17.4% lost to follow-up), and 55.3% (1267/2291) of participants have completed T₈ (24-month follow-up). The study protocol allows for participants with missing follow-up data to be recontacted at subsequent follow-ups to maximize data completeness and cohort retention. Prospective outcomes will be validated against national EMR data. These data will provide granular diagnostic, procedural, and laboratory data on falls, fractures, and broader age-related diseases for the cohort.

Finally, EMRs for each TARGET cohort member who provided consent for record linkage at baseline (1243/2291, 54.3%) are being obtained through Singapore Ministry of Health’s TRUST platform [49]. TRUST is a secure data exchange platform that enables the fusion and sharing of anonymized health-related research and real-world data for research purposes between public institutions and the public and private sectors. Linked datasets to be requested for the TARGET cohort will include health records (from acute hospitals, polyclinics, general practice clinics, and medical centers), health (emergency department presentations), and birth and death registry (mortality). EMR data will be used to analyze longitudinal pathology, morbidity, and mortality patterns associated with falls and broader age-related diseases in older adults. Analysis will be performed on the deidentified fused datasets in the secure TRUST Workspace.

Table S1 in Multimedia Appendix 1 compares key characteristics between participants who completed and did not complete the clinic visit components (DXA and 3D body scans and SPACE game), and participants who consented and did not consent to the EMR linkage to evaluate potential selection bias. Overall, participants who did not complete the clinic visit components were more likely to be older and have greater ADL/IADL difficulty.

All researchers were required to complete the Collaborative Institutional Training Initiative online program for research ethics and compliance education. Fieldwork interviewers, including university interviewers and interviewers from a survey research company, were required to undergo training and fieldwork shadowing and to demonstrate competency in the relevant procedures and questionnaire administration prior to being approved to perform household recruitment with participants. Study documents, including participant information sheets, informed consent forms, and questionnaires, were translated into Chinese, Malay, and Tamil languages by a certified translation company. The study team validated the translated documents to ensure that the meaning is consistent with the original version. A computer-assisted personal interview questionnaire was designed in Qualtrics using logic and rules to assist fieldwork interviewers to conduct proper inclusion and exclusion criteria eligibility checks. A pilot study (n=18) was undertaken to test recruitment processes and ensure that interviewers were familiar with the recruitment and interviewing procedures. Data collected by all fieldwork interviewers were checked periodically by key study team personnel. Blood pressure machines and weighing scales were checked and certified by the supplier before being issued to the interviewers. All faulty study instruments are either serviced or replaced depending on the availability of spare parts to ensure measurement accuracy.

The primary outcomes of interest are the incidence of falls and fractures over the 2-year prospective follow-up period. Falls are defined following the World Health Organization’s definition as “an event which results in a person coming to rest inadvertently on the ground, floor, or lower level” [100]. Fracture events will include both self-reported and electronically validated incident fractures identified during follow-up and through linkage with EMR data. MOF will include fractures of the hip, clinical spine, wrist, and humerus, consistent with the FRAX definition. Hip fractures will be analyzed separately due to their substantial morbidity, mortality, and health care burden among older adults.

Prospective outcomes collected through quarterly follow-up interviews will be validated against linked EMR using relevant diagnostic and procedural coding systems, including ICD (International Classification of Diseases) codes, available within the TRUST-linked datasets. Time-to-event analyses will be conducted to evaluate predictors of incident falls and fractures over the follow-up period.

Primary analyses will focus on the development and validation of novel personalized risk prediction models integrating variables from the epidemiological survey, gait assessment, clinical imaging, SPACE, and EMR data to segment the older adult population into different fall and fracture risk levels. Risk prediction models will be built using state-of-the-art machine learning, deep learning, and statistical techniques. The performance of the novel risk prediction models will be evaluated against existing gold standard assessments (eg, STEADI algorithm, Three Key Questions, and FRAX) and validated against the prospective data. Secondary analyses will focus on (1) developing predictive models of fracture risks among older adults based on holistic body composition parameters obtained from DXA and whole-body 3D shape scans, (2) assessing the performance of SPACE to detect cognitive impairment and its predictive usefulness in detecting falls, and (3) assessing the psychological and sociodemographic correlates of fall risk.

Cost-effectiveness analyses of screening strategies will be evaluated using decision tree models. The analyses will use Markov models to simulate the long-term costs and quality-adjusted life years over the lifespan. Prospective epidemiological data collected from TARGET will be used as input parameters for event rates and diagnostic accuracy. Performance of novel risk prediction tools (ie, sensitivity, specificity, and accuracy) will be evaluated against current methods of evaluating falls and fracture risk to determine the incremental cost-effectiveness ratios of different screening strategies. A budget impact analysis will determine the short- and medium-term fiscal implications of a national implementation of these strategies.