Falls are one of the leading causes of disability, hospitalization, and mortality among older adults [1,2]. Early detection of functional decline is critical for fall prevention; however, conventional assessments such as the Short Physical Performance Battery (SPPB) rely heavily on manual observation, which is time-consuming and subject to interrater variability [3,4]. With the rapid advancement of digital health technologies, artificial intelligence (AI) has emerged as a promising tool for objective, real-time mobility screening [5-7].

Recent studies have demonstrated that AI-based motion analysis can improve the accuracy and scalability of functional assessments in older adult populations [3,6,8-10]. Unlike static paper-based tests, AI can dynamically track skeletal movement, quantify gait and balance parameters, and provide continuous feedback. These innovations enable early detection of mobility decline and offer opportunities for personalized interventions [5,7,9-11]. Despite these advances, limited large-scale evidence exists regarding whether AI-assisted screening, when combined with adaptive training, leads to measurable improvements in functional mobility outcomes among older adults.

Recent JMIR studies have also highlighted the feasibility of integrating SPPB with broader fall risk assessment tools [12], the application of explainable AI for fall prediction [13], and the potential of digital and telemedicine-based prevention programs [14,15]. Moreover, wearable technology is increasingly considered a viable option for health care monitoring [16].

This study investigates the effectiveness of AI-assisted dynamic postural control screening combined with adaptive training in improving mobility outcomes among older adults. The research makes three key contributions:

This study addressed the following research question: can AI-assisted dynamic postural control screening combined with adaptive training significantly improve functional mobility outcomes (SPPB score, gait speed, and sit-to-stand performance) compared with standard care in older adult populations?

This study had the following research hypotheses:

While prior studies have demonstrated that the SPPB is a reliable predictor of functional decline [12,13], its traditional administration is limited by subjectivity and reliance on manual timing or scoring [14]. AI-assisted motion analysis offers a scalable solution by automating the assessment process, increasing measurement precision, and enabling real-time feedback [15,16].

This study builds upon existing evidence by integrating AI-based skeletal tracking with SPPB assessments and extending the application to a large older adult cohort. The primary aim is not only to demonstrate feasibility but also to evaluate whether AI-assisted screening, when combined with adaptive training prescriptions, leads to measurable improvements in mobility outcomes. Imran et al [10] and Santos et al [11] specifically support the dual rationale of integrating AI into both functional assessment and individualized prescription, underscoring the novelty of this approach.

By conducting a quasi-experimental study with more than 2000 participants, this research provides large-scale evidence on the effectiveness of AI-assisted interventions. Unlike prior work that focused mainly on prediction or assessment, our approach tested whether real-time AI-driven prescriptions translated into tangible improvements in balance, gait, and sit-to-stand performance.

This quasi-experimental study was conducted in Keelung, Taiwan, over a 6-month period. A total of 2005 older adults (≥65 years) were recruited through stratified sampling from 12 community centers and 9 health care institutions.

The experimental group (n=1451) was recruited primarily from community centers. Participants underwent AI-assisted SPPB assessments and received individualized exercise prescriptions dynamically adapted by the AI system. Prescriptions were adjusted at subsequent assessments according to participants’ performance.

The control group (n=554) was recruited from health care institutions. Participants underwent AI-assisted assessments and continued with standard care (ie, routine health advice and monitoring by their institutions) but did not receive AI-tailored prescriptions. This ensured ethical parity while isolating the effect of AI-enabled personalization.

Importantly, baseline differences reflected recruitment sources: the experimental group had a mean SPPB score of 11.2 (SD 1.8), whereas the control group had a mean score of 10.4 (SD 1.8). These differences were expected, given the healthier status of community-dwelling participants compared with clinical recruits. Independent-sample tests confirmed that baseline differences did not bias the interpretation of intervention outcomes. A summary of participant recruitment, group allocation, and follow-up assessments is presented in Figure 1.

Although not prespecified, a post hoc power analysis using G*Power (Heinrich Heine University Düsseldorf) indicated that a minimum of 788 participants per group would be required to detect a medium effect size (Cohen d=0.4) with 90% power at α=.05. Our actual sample sizes (1451 and 554) exceeded this requirement, ensuring sufficient statistical power and confirming that the large cohort was methodologically and ethically justified.

Functional mobility was assessed using the SPPB, which is widely validated as a predictor of functional decline and fall risk [4,11-13]. Wearable and sensor-based technologies further demonstrate feasibility for continuous monitoring and objective evaluation [16]. The SPPB included (1) balance tests, (2) 4-meter gait speed, and (3) the 5-times sit-to-stand test. Assessments were conducted with AI-assisted skeletal tracking to improve accuracy and objectivity.

The AI system replaced traditional manual scoring by extracting quantitative motion features in real time. This consolidated statement replaces prior repeated descriptions, improving readability while preserving scientific accuracy. Key features included the following:

In a prior pilot study of 320 older adults, the AI algorithm achieved greater than 90% accuracy in classifying functional mobility categories compared with physiotherapist ratings. Test-retest reliability was high (intraclass correlation coefficient >0.88) for gait speed and sit-to-stand measures, supporting the robustness of the system. The architecture and workflow of the AI system are summarized in Figure 2, showing the process from skeletal tracking and feature extraction to automatic scoring and feedback generation.

This study was approved by the Institutional Review Board of En Chu Kong Hospital (ECKIRB1071204). All participants were informed about the study objectives and procedures and provided written informed consent before enrollment. No financial incentives or compensation were provided for participation. All data were anonymized, and confidentiality and data security measures were strictly implemented to protect participant privacy in accordance with institutional and national ethical guidelines.

In the experimental group, the AI system generated individualized exercise prescriptions that were adapted at subsequent assessments. Algorithms adjusted the type, intensity, and progression of exercises based on gait speed thresholds, sit-to-stand performance, and balance stability metrics.

This ensured tailored interventions beyond generic training exposure. The control group used the AI system solely for assessment, without adaptive prescriptions. The adaptive prescription framework is presented in Figure 3, which illustrates how SPPB-derived performance metrics (ie, balance, gait, and sit-to-stand performance) informed AI-driven adjustments in exercise type, intensity, and progression.

Participants in both groups were assessed at baseline and 6-month follow-up. Monthly interim assessments were conducted to monitor progress. All assessments were video recorded through the AI skeletal tracking system and processed automatically.

Data collected included balance scores from SPPB balance tasks; gait speed (cm/s) from the 4-meter walk; sit-to-stand time (s) from the 5-repetition test; and demographics such as age, sex, and recruitment source.

AI-generated feedback reports were shared with trainers to guide intervention adjustments in the experimental group.

All statistical analyses were performed using SPSS (version 27.0; IBM Corp). Descriptive statistics (means and SDs) were calculated for all measures. Between-group comparisons were conducted using independent-samples t tests to compare improvements in SPPB, gait speed, and sit-to-stand performance between the experimental and control groups. Within-group changes were examined using repeated-measures ANOVA to examine changes from baseline to follow-up within each group. Participation-level analysis was performed using one-way ANOVA to compare outcomes among subgroups with high (>75%), moderate (50%‐75%), and low (<50%) participation rates, to explore the relationship between adherence and functional outcomes. Effect sizes were calculated using Cohen d for pairwise comparisons, and η² was reported for ANOVA models; 95% CIs were reported for all effect size estimates. Statistical significance was set at a 2-tailed α=.05.

A total of 2005 participants were analyzed, including 1451 in the experimental group and 554 in the control group. The experimental group had a mean age of 73.4 (SD 6.1) years, with 841 (57.9%) female participants. The control group had a mean age of 74.1 (SD 6.4) years, with 338 (61.0%) female participants.

Baseline differences in walking speed were noted. As shown in Table 1, the experimental group demonstrated a mean baseline gait speed of 117 cm/s, compared with 105 cm/s in the control group. These values align with cut-off thresholds for functional decline in mobility. Baseline functional measures showed expected differences due to recruitment sources (community vs health care institutions):

An SPPB score of ≤9 was considered indicative of high fall risk, while scores ≥10 represented normal to moderate function, consistent with prior validation studies. Independent-samples tests confirmed that although baseline functional status differed, these differences did not bias the interpretation of intervention outcomes.

Participation rates strongly influenced outcomes [17]. The differences across participation levels are summarized in Table 2. The experimental group was divided into high (>75%), moderate (50%‐75%), and low (<50%) attendance subgroups.

One-way ANOVA confirmed significant differences across participation levels for all three functional outcomes. As shown in Table 2, higher participation was associated with greater improvements in SPPB scores, gait speed, and sit-to-stand performance.

This study demonstrated that AI-assisted dynamic postural control screening combined with adaptive training significantly improved functional performance among older adults over a 6-month period. Improvements were observed in overall SPPB scores, gait speed, and sit-to-stand performance, with effect sizes ranging from moderate to large (Cohen d=0.39‐0.62). Furthermore, the participation rate was positively associated with functional gains, indicating a dose-response relationship.

Our results align with previous AI- and sensor-based fall risk assessment studies [3,5,9,13,18,19]. Importantly, our findings expand on recent JMIR research exploring digital health interventions for fall prevention [14,15]. Classic reviews and epidemiological studies have long established that falls among older adults are preventable through multifactorial interventions [20]. In addition, the screening component itself warrants emphasis. Early and objective detection of deficits in mobility, balance, and gait is a prerequisite for timely intervention, particularly as conventional screening tools such as the SPPB are subject to variability and may not scale efficiently in community or clinical settings [4,12]. By embedding AI-based screening in our design, we demonstrated not only intervention efficacy but also the feasibility of scalable, objective, and standardized risk detection, which justifies the inclusion of the screening part of the study. This represents a shift from using AI solely for risk identification toward using it as a tool for delivering individualized, dynamic interventions. While conventional screening tools such as the SPPB are validated, they are limited by interrater variability and scalability issues. Few existing AI interventions have incorporated both real-time skeletal tracking and adaptive exercise prescriptions, making our approach distinct from most prior AI studies. In contrast, our study demonstrates that AI-assisted screening combined with adaptive prescriptions can not only identify deficits but also drive functional improvements in a large older adult cohort. A structured comparison of prior AI applications and this study is summarized in Table 3, highlighting how our work advances beyond assessment toward intervention.

The findings suggest that AI-assisted systems have potential for broad application in community and clinical settings. By automating data capture and analysis, the AI platform provides objective, real-time, and scalable assessments that reduce reliance on subjective manual scoring. The adaptive prescription framework further allows for individualized exercise programs that can be implemented in day care centers, senior fitness programs, and long-term care facilities. Such systems may enhance efficiency for health care staff while promoting functional independence among older adults.

Several limitations should be noted. First, fall incidence was not directly measured; therefore, conclusions should be interpreted as functional performance improvements rather than confirmed reductions in fall events [21]. Second, the recruitment of the experimental group from community centers and the control group from health care institutions introduced baseline differences in functional status. Although statistical analyses confirmed that these differences did not bias the outcomes, they may limit comparability. Third, the study was conducted in Keelung, Taiwan, which may limit generalizability to populations with different cultural, socioeconomic, or rural backgrounds. Fourth, the follow-up period of 6 months reflects short-term to midterm outcomes; longer-term effects remain unknown.

Future studies should extend follow-up to 12 to 24 months to examine the sustainability of functional gains and to prospectively track fall incidence [22]. Multisite validation across diverse populations, including rural communities, lower-income groups, and culturally varied settings, is required to establish generalizability. Further work should also investigate adherence, engagement, and the accessibility of AI technologies, particularly among populations considered vulnerable who may face barriers to digital health adoption.

This study demonstrated that AI-assisted dynamic postural control screening, when combined with adaptive training, significantly improved functional mobility among older adults. Improvements were evident in SPPB scores, gait speed, and sit-to-stand performance, with effect sizes in the moderate-to-large range. These results, based on a large sample of more than 2000 participants, provide strong support for the feasibility of implementing AI-assisted programs at scale. Importantly, participation levels showed a dose-response relationship with outcomes, highlighting the role of consistent engagement in maximizing functional benefits.

While the findings are encouraging, several considerations remain. Fall incidence was not directly measured; therefore, conclusions should be interpreted as functional improvements rather than confirmed reductions in fall risk. The 6-month study duration limits insights into long-term sustainability, and the recruitment of participants primarily from community centers and health care institutions in Keelung, Taiwan, may constrain generalizability. Furthermore, the accessibility and usability of AI-based systems for older adults in rural or low-resource settings require further examination.

In summary, this study contributes large-scale evidence that AI-assisted dynamic postural control screening with adaptive training can enhance functional performance in older adults. Future research should extend longitudinal follow-up to 12 to 24 months, prospectively measure fall events, and conduct multisite trials across diverse cultural and socioeconomic contexts. Addressing adherence, technology accessibility, and digital literacy will be essential to ensure equitable adoption and long-term impact of AI-assisted older adult care solutions.