Are We Missing the Environmental Factors in AI-Based Fall Risk Models?: A Systematic Review

Are We Missing the Environmental Factors in AI-Based Fall Risk Models?: A Systematic Review | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Systematic Review Are We Missing the Environmental Factors in AI-Based Fall Risk Models?: A Systematic Review Jiyoun Song, Boeun Kim, Min-Jeoung Kang, Shuxuan Li, Lingjie Liu, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8723907/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 4 You are reading this latest preprint version Abstract Background Falls commonly occur in home environments where environmental conditions can contribute to fall risk. Identification and mitigation of environmental hazards are critical components of fall prevention. However, artificial intelligence (AI)-based fall prediction models have largely focused on individual-level predictors, with limited attention to home environmental hazards despite their modifiable role in fall risk. Objective To systematically review how environmental factors are incorporated into existing AI-based fall risk prediction models and summarize reported AI approaches and model performance among community-dwelling older adults. Methods This systematic review followed PRISMA guidelines. Six electronic databases (PubMed, Embase, CINAHL, Cochrane Library, Web of Science, and Scopus) were searched from inception through December 2025. Eligible studies applied AI-based models to predict falls among older adults in community settings and incorporated environmental factors as model inputs. Results Of more than 10,000 records identified, nine studies met final inclusion criteria. Six used supervised machine learning with structured data, while three employed computer vision or robotics-based approaches. Environmental factors were heterogeneously represented, ranging from checklist-based indicators to sensor- and vision-derived measures. When included, environmental features contributed meaningful information by improving discrimination or identifying actionable home hazards (AUC-ROC ranged from 0.67 to 0.76). Conclusion Environmental factors remain underemphasized in AI-based fall prediction models. Greater integration of standardized and context-aware environmental information may enhance the relevance and preventive utility of AI-based fall risk prediction in community settings. Fall Risk Assessment Predictive Modeling Environmental Factors Systematic Review Nursing Informatics Figures Figure 1 Figure 2 Introduction Falls pose a critical challenge to healthy aging and remain one of the leading causes of injury among community-dwelling older adults. 1 The incidence of falls continues to rise as the global population ages, with approximately one in four older adults experiencing at least one fall each year. 1 Among older adults, falls are associated with serious physical, psychological, and social consequences, including fractures, traumatic brain injuries, functional decline, and long-term disability. 2 , 3 Older adults aged 75 years and older experience the highest rates of fall-related hospitalization and are particularly vulnerable to prolonged recovery and subsequent loss of independence. 4 These outcomes directly undermine well-being and healthy aging and contribute to greater healthcare utilization, long-term care needs, and economic burden. 5 Recent advances in artificial intelligence (AI) and machine learning (ML) have generated growing interest in their application to fall risk prediction and prevention. A rapidly expanding body of literature has explored AI-driven approaches for fall prediction, including machine learning algorithms, deep learning models, and sensor- or wearable-based systems. 6 – 9 Traditionally, fall risk assessment has relied on standardized clinical tools and screening instruments developed to identify individuals at elevated risk. Commonly used tools include the Berg Balance Scale (BBS) to assess functional balance, 10 Functional Reach Test, and various fall risk questionnaires that assess fall history, mobility limitations, medication use, and comorbidities. 11 These instruments have been widely implemented in clinical and community settings given their simplicity and straightforward administration. Data derived from these traditional assessment tools have also been used as input features in AI–based fall prediction models, either independently or in combination with additional clinical, functional, and sensor-derived variables. 12 However, despite their widespread adoption, traditional fall risk assessment tools demonstrate significant limitations in capturing dynamic and environmental determinants of fall risk Falls result from the interaction of multiple factors, including individual characteristics, physical and psychological impairments, and environmental conditions. 4 , 13 Accurate fall prediction, therefore, requires approaches capable of integrating comprehensive, diverse and interrelated risk factors. While individual-level risk factors (e.g., age, comorbidities, medication use, and prior falls) and functional-level risk factors (e.g., balance, mobility, and functional limitations) have been extensively studied, 14,15 home environmental hazards remain relatively understudied despite being an integral and modifiable component of fall risk among community-dwelling older adults. Notably, a majority of falls occur at home, contributing to 79.2% of fall-related emergency department visits, 16,17 underscoring the central role of home settings in fall prevention efforts. Common hazards include inadequate lighting, slippery surfaces, loose rugs, unsafe handrails, and cluttered pathways, 4,18 which often interact with age-related functional limitations, increasing fall risk and highlighting the need for systematic approaches to environmental risk identification and mitigation. AI-based models offer unique advantages in handling complex, high-dimensional data and identifying non-linear interactions among multiple risk factors. Despite these advances, it remains unclear to what existing AI-based fall prediction models account for environmental risk factors which play a modifiable role in fall risk. To address this knowledge gap, this study aims to systematically synthesize the existing literature on the inclusion and representation of environmental factors in AI-based fall risk prediction models, and summarize the algorithms and performance metrics reported in studies of falls among community-dwelling older adults. Methods This systematic review followed the guidelines of Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 19 The protocol was registered in the PROSPERO (CRD420261279513). Eligibility Criteria (1) Inclusion Criteria Studies were eligible for inclusion if they met the following criteria: (1) focused on community-dwelling older adults aged 60 and above from any country; (2) utilized AI or technology-based models to assess, predict, or prevent fall risks, including machine learning models or mobile health technologies; (3) assessed or predicted fall risk as a primary outcome; (4) incorporated environmental factors as one of the input variables or predictors in the model; (5) were data-based primary studies, including model development, validation, and evaluation studies reporting empirical data on model outcomes, accuracy, or effectiveness; and (6) were written in English. (2) Exclusion Criteria Studies were excluded if they: (1) focused on institutionalized populations such as nursing home residents; (2) employed non-technology or non-AI-based approaches, such as manual assessments or traditional fall risk screening methods; (3) were qualitative studies, dissertations, conference abstracts, editorials, commentaries, or literature reviews. Search Strategy The literature search strategy was developed in collaboration with a librarian and conducted across six electronic databases (i.e., PubMed, Embase, CINAHL, Cochrane Library, Web of Science, and Scopus) on December 19, 2025. In addition, a manual review of the reference lists of included studies was conducted to identify additional eligible articles. No publication date limits were applied to ensure comprehensive retrieval of all relevant articles. A comprehensive search methodology was designed using both free-text keywords and Medical Subject Headings (MeSH) terms ( Appendix 1 ). The search strategy incorporated three major concept groups combined with Boolean operators: (1) Falls-related terms (e.g., "Accidental Falls"[MeSH], fall, falls, falling, fallen); (2) AI and technology-related terms (e.g., "Artificial Intelligence"[Mesh] machine learning, deep learning, telemedicine, digital health, mHealth, sensors, smart home technology); and (3) Older adult-related terms (e.g., aged[MeSH], older adult, elder*, senior*, geriatric*). Study Selection and Data Extraction Search results were imported into Covidence (Veritas Health Innovation, Melbourne, Australia), and duplicates were removed prior to screening. Four reviewers [JS, BK, MK, and WJ] conducted a two-stage screening process. In the title and abstract screening stage, two reviewers independently assessed each record, and potentially relevant articles advanced to full-text review. During the full-text assessment, two reviewers again independently evaluated each article against eligibility criteria. At both stages, disagreements were resolved through consensus with a third reviewer who had not participated in the initial assessment. In order to finalize the form, two authors created a standardized data extraction form and pilot-tested it on one article before finalizing the form. Each included study was independently extracted by two reviewers, with discrepancies resolved by a third reviewer. Extracted information included publication details (author, year, country, data source), study characteristics (aim, design, sample size, inclusion/exclusion criteria, recruitment), details of the AI model (type of approach, algorithms used, input variables), fall risk factors (including environmental factors and how they were measured) and model performance metrics (accuracy, sensitivity, specificity, AUC-ROC, F-score). Study limitations and potential conflicts of interest were also recorded. Findings were synthesized into evidence tables summarizing study characteristics, AI approaches and risk factors, and model performance outcomes. Quality Appraisal The methodological quality of included observational studies was assessed using the National Heart, Lung, and Blood Institute (NHLBI) Quality Assessment Tool for Observational Cohort and Cross-sectional Studies. 20 The tool consists of 14 criteria evaluating clarity of research objectives, definition of the study population, participation rates, consistency of recruitment methods, adequacy of sample size justification, temporal relationship between exposure and outcome measurement, validity and reliability of exposure and outcome measures, blinding of outcome assessors, completeness of follow-up, and control for confounding variables. Each study was independently appraised by two reviewers, with discrepancies resolved through discussion with a third reviewer. Ethical Considerations This systematic review utilized only publicly available published data and did not involve human subjects. Therefore, institutional review board (IRB) approval was not required. Results An overview of the review process is depicted in Fig. 1 . A total of 21,760 articles were identified through the initial literature search, including PubMed (n = 2,185), Embase (n = 3,804), CINAHL (n = 948), Cochrane Library (n = 854), Web of Science (n = 5,981), and Scopus (n = 7,988). No additional studies were identified through manual searching of reference lists and research reports. After the removal of duplicate articles (n = 8,918), the remaining articles (n = 12,842) were screened. Of these, 12,817 were excluded after title and abstract review. During the full-text screening, an additional 16 articles were excluded for not meeting the inclusion criteria. Throughout the exclusion process, the majority of excluded studies did not include environmental risk factors, which were a key inclusion criterion. Consequently, a total of nine studies were included. Overall, the methodological quality of included studies was variable (Fig. 2 ). Most studies clearly stated research objectives and adequately defined study populations, but common limitations included incomplete reporting of sample size justification, limited handling of confounding, and insufficient information on exposure and outcome measurement reliability. Studies using vision- and sensing-based approaches, which rely on unstructured visual or depth data captured in real-world environments, showed greater uncertainty across several quality domains compared with structured-data–based machine learning studies, which used predefined, tabular inputs. Overview of Included Studies Table 1 summarized the overview of included studies. The included studies were conducted across multiple countries, with the majority originating from the United States 21 , 22 and European countries (Belgium, 23 Switzerland, 24 Spain, 25 and the United Kingdom 26 ), alongside representation from Asian countries including China 27 and Thailand. 28 Six studies (66.7%) used structured datasets, such as electronic health record (EHR), home health assessments or community-based registries, to train machine-learning models to predict future falls. 21 , 23 , 25 , 27 – 29 In these studies, longitudinal follow-up periods and large samples were used to evaluate predictive performance across a wide range of demographic, clinical, and environmental variables. These included studies encompassed diverse populations with sample sizes ranging from 304 to 59,028 participants. The largest study by Lo et al. 21 analyzed data from 59,028 unique patients aged over 65 years using retrospective EHR with available OASIS-C assessment and MAHC-10 fall risk assessment data. Lathouwers et al. 23 examined 82,580 community-dwelling older adults aged 60 and over in Belgium, recruited through stratified random sampling by sex and age from census data; after excluding missing data, 24 input features and 33,346 entries remained for analysis. In contrast, three studies used sensor- or computer-vision–based experimental designs. 22 , 24 , 26 Du et al. employed robotics-enabled environmental screening incorporating luminosity and spatial measurements, 22 while Cloix et al. validated depth-based stair detection using controlled motion capture settings. 24 Moore et al. developed AI algorithms within simulated environments or structured testing conditions to detect environmental hazards or risky activities related to falls. 26 These three studies did not provide detailed participant recruitment or demographic information. Table 1 Overview of Included Studies Author (Year of publication) Study region Study design Aim Date Source Participants characteristics Chen et al. (2023) 27 China Cohort study To build prediction models for falls and fall-related injuries among community-dwelling older adults in China China Health and Retirement Longitudinal Study (CHARLS), a national cohort of middle-aged and older Chinese adults (mainly ≥ 45 years) across 28 provinces. Launched in 2011 with follow-up in 2013, 2015, and 2018. This study used data from the latest two waves (2015 and 2018). • Recruitment: Not applicable. • Inclusion criteria: Not applicable. • Exclusion criteria: (1) aged 10% missing data in individual variables. • Total participants: 5,818. Cloix et al. (2016) 24 Switzerland Other: Technology Evaluation Study To evaluate a low-complexity algorithm to detect descending stairs and curbs of any shape, specifically designed for low-power real-time embedded platforms The cameras were located at 78 cm height with a tilted angle of 35°. Images were captured at 512 × 384 and 640 × 480 pixel resolution with the stereo camera and the Red Green Blue-Depth (RGBD) sensor. • The assessment of our approach was carried out using thirteen scenes of descending stairs and curbs. • Recruitment: Not applicable. • Inclusion criteria: Not applicable. • Exclusion criteria: Not applicable. • Total participants: Not applicable. Du et al. (2014) 22 USA Other: Engineering Feasibility and Prototype Validation Study To develop a robot-based system for in-home environment screening that supports both manual and autonomous fall risk assessment. Real-time sensor data collected by a mobile robot (Turtlebot2) in a simulated in-home environment. • Recruitment: Not applicable. • Inclusion criteria: Not applicable. • Exclusion criteria: Not applicable. • Total participants: Not applicable. Lathouwer et al. (2022) 23 Belgium Cohort study To identify fall risk factors using a quality-of-life questionnaire covering biological, behavioral, environmental, and socio-economic domains. Questionnaire developed by the Belgian Ageing Studies research group of the Vrije Universiteit Brussel. • Recruitment: Stratified random sampling (by sex and age) from census data of community-dwelling adults aged ≥ 60 in Belgian municipalities. • Inclusion criteria: Community-dwelling older adults aged ≥ 60 living in Belgium. • Exclusion criteria: Not applicable. • Total participants: 33,346 community-dwelling older adults; 84 questionnaire items excluded due to missing data, resulting in 24 input features and 33,346 valid entries. Lo et al. (2019) 21 USA Cross sectional study To devise a machine learning pipeline using existing home health care data to predict fall risk among older adults, with the goal of building clinical decision support tools for fall prevention Outcome and Assessment Information Set (OASIS-C) – mandatory assessment for home health care patients; Electronic Health Records (EHR) – supplemental demographic data. • Recruitment: Not applicable (retrospective EHR-based study). • Inclusion criteria: Age ≥ 65 years; available OASIS-C assessment data and MAHC-10 fall risk assessment data. • Exclusion criteria: Not applicable. • Total participants: 59,028 unique patients. Millet et al. (2023) 29 Spain Cohort study To predict recurrent falls in the older population using machine learning techniques, with the aim of reducing the number of falls and their consequences. Getafe University Hospital’s Geriatric Falls Unit (data obtained from the Hospital’s Electronic Health Records) between January 2017 and December 2021. Older adults who experienced a fall requiring medical attention were referred to this unit for personalized treatment and monitoring to prevent repeated falls. • Recruitment: Not applicable. • Inclusion criteria: Adults who were treated at the Geriatric Falls Unit for at least one fall (mean age 80.3±7.7). • Exclusion criteria: Not applicable. • Total participants: 304 older adults. Moore et al. (2024) 26 UK Other: A mixed-method study To develop an AI algorithm using wearable video-based eye-tracking and IMU gait data, and to explore how people with Parkinson’s disease perceive the use of video and AI technologies in daily life for understanding their fall risk. For developing the AI model, pre-trained MS COCO weights were initialized and fine-tuned on a local dataset (80:20 split). Ten hours of local video data (> 1500 frames) were manually extracted and annotated using the LabelImg tool, with 4 categories and 18 classes selected for relevance to fall risk or privacy. A total of 1542 frames were selected and annotated, including: (1) scripted route data from 7 young adults covering 10 interior and 10 exterior environments; (2) 240 manually extracted frames from 3 PwPD, ensuring unseen data for testing; and (3) 4 additional first-person-view exterior videos (~ 240 min) downloaded under CC-BY licence and annotated. • Recruitment: – PwPD participants were recruited locally through networks within Northumbria University. – Seven young adults (6M:1F, 23–32 years) were recruited to wear video glasses solely for generating AI training data; young adults were recruited through word of mouth. • Inclusion criteria: – For PwPD: • Clinical diagnosis of Parkinson’s disease; • Prior experience in wearable-gait research; • Familiarity with technology (e.g., smartphones, tablets, applications); • Willingness to attend focus groups and consent to audio-recording; • English-speaking and literate. – For young adults: Not formally specified beyond age and functional independence. • Exclusion criteria: – For PwPD: Significant cognitive impairment. – For young adults: Any functional impairments. • Total participants: Video dataset included frames from 7 young adults and 3 PwPD; focus group size was n = 4. Panyakaew et al. (2021) 28 Thailand Cross sectional study To explore the prediction of falling in Parkinson's disease patients using a machine learning-based approach. Recruited from the Chulalongkorn Center of Excellence for Parkinson’s Disease and Related Disorders, Faculty of Medicine, Chulalongkorn University, Thailand, between January and December 2019. • Recruitment: Patients with a clinical diagnosis of PD with Hoehn & Yahr (H&Y) stage 1–4, with or without a history of falls, were recruited from the same center ( www.chulapd.org ) between January and December 2019. • Inclusion criteria: (1) Clinical diagnosis of PD based on standard diagnostic criteria; (2) Hoehn & Yahr (H&Y) stage 1–4; (3) able to ambulate within community residences; (4) able to follow simple commands with a score ≥ 21 on the Thai version of the Mini-Mental Status Examination (MMSE); and (5) able to complete the ABC-16 scale independently. • Exclusion criteria: (1) Hoehn & Yahr stage 5; (2) coexisting disorders contributing to postural instability and falls, including stroke, ataxia, neuropathy, and visual, vestibular, or proprioceptive problems; (3) patients taking sedative medications. • Total participants: 305. Pérez-Ros et al. (2019) 25 Spain Other: Cohort trial nested case-control design To determine the predictive factors for falls in functionally independent community dwelling older adults. Data was collected from a prospective, longitudinal study of 374 community-dwelling older adults aged ≥ 70 years in La Ribera, Valencia, Spain, recruited from primary care centers between December 2013 and May 2014, with a 12-month follow-up to assess falls. • Recruitment: A publicity strategy was used from December 2013 to May 2014, including posters and notifications in retirement centers, meetings with managing authorities of retirement associations and centers, and telephone calls to individuals enrolled in primary care centers. Family telephone calls were also used to reduce dropouts among individuals lacking the physical means to attend fall-prevention education sessions. • Inclusion criteria: (1) Aged ≥ 70 years; (2) Barthel Index ≥ 60 (functionally independent); (3) Independent walking (with or without technical aids, but not assisted by another person); (4) Resident in the La Ribera region. • Exclusion criteria: (1) Life expectancy < 6 months; (2) Blindness or deafness; (3) Serious psychiatric illness (schizophrenia, major depression, bipolar disorder, panic disorder); (4) Moderate to severe cognitive impairment (MEC-Lobo < 24); (5) Did not sign informed consent. • Total participants: 374. AI-based Fall Risk Prediction Models Table 2 summarizes the characteristics of AI-based fall risk prediction models included in this review, including the type of AI algorithms used, environmental and non-environmental risk factors incorporated, reported model performance metrics, and key study limitations. Table 2 AI-Based Fall Risk Models and Environmental Risk Factors Author (Year of publication) Type of AI algorithm Environmental Risk Factors Non-Environmental Risk Factors Model Performance Limitation Chen et al. (2023) 27 Machine learning • AI algorithm: logistic regression (LR); support vector machine (SVM); random forest (RF); adaptive boosting (AdaBoost); light gradient boosting machine (LightGBM) Structure of building; handicapped facilities; kitchen availability; flush toilets; cooking fuel type; Internet availability; house tidiness; house temperature • Measures of environmental risks: inquiry-based survey measures Biological, behavioral, and socioeconomic predictors including sex, vision, diabetes, liver disease, memory-related disease, disabilities (brain, vision, physical), ADL, IADL, experience of falling, hip fracture history, smoking, alcohol use, sleep duration, medications (antihypertensive, lung, heart, kidney, memory, arthritis), depressive symptoms, dental care, relationship with children, health satisfaction, systolic BP, walking speed, hand strength, white blood cell count, blood urea nitrogen, glucose; injury model includes marital status, stroke, dyslipidemia medication, digestive medication, income, life satisfaction, lung function, abdominal obesity The LR model showed good performance with an AUC-ROC of 0.739 (sensitivity 0.707) for falls and an AUC-ROC of 0.757 (sensitivity 0.654) for fall-related injuries. • Fall risk impact: Baseline falling experience was the most important predictor. A history of falls was the strongest predictor of future falls, and house tidiness was an important environmental predictor in both fall and fall-related injury models. Variables were treated as categorical; causal inference was not possible; predictors were only measured at baseline, without accounting for changes over time. Cloix et al. (2016) 24 Computer vision for stairs and curbs detection • AI algorithm: Binary prediction algorithm (Algorithm 1) using upper/lower depth map (Dp) from stereo images; Three-class prediction algorithm (Algorithm 2) using full depth map (D) divided into sub-depth maps; thresholds TG/TR and TGi/TRi applied to ground distance and ratio of valid pixels • Environmental risks identified: Stairs and curbs • Measures of environmental risks: depth difference from stereo disparity; ratio of pixels below ground level; spatial location of drop (upper vs. lower half of image); predicted “danger zone” proximity using trapezoid projection Not applicable The system distinguished more than 94% dangerous scenes from safe scenes with an overall recognition rate of 91% at very low resolution, operating in real time and robust to indoor/outdoor lighting conditions. • Fall risk impact: Not applicable. HDR cameras can better capture scenes under bright sunlight to avoid saturation that causes stereo image matching to fail, and DSPs may assist with motion blur removal before disparity computation to improve the SAD stereo image matching. Du et al. (2014) 22 Computer vision, robotics, and shared control systems • AI algorithm: Gmapping (Rao-Blackwellized Particle Filter SLAM); OctoMap (3D mapping using octrees); costmap-based obstacle avoidance planning; AMCL (Adaptive Monte Carlo Localization) Poor lighting; cluttered or narrow spaces; obstructed walkways; limited visibility (telepresence) • Measures of environmental risks: luminosity (lux) thresholds: ≥401 = bright, 201–400 = medium, < 200 = dark; lighting condition displayed on web interface (green = bright, yellow = medium, red = dark) Not applicable • Model performance: Not reported. • Fall risk impact: Not applicable. The system is a prototype and requires major improvements, including automation of assessment tasks, addition of object recognition functions for independent hazard detection, and better error handling and recovery methods to ensure reliability and stability for use in real world. Lathouwer et al. (2022) 23 Machine learning • AI algorithm: Random Forest model Housing issues, housing change, environmental vulnerability • Measures of environmental risks: not reported 24 variables including biological (loneliness, sex), behavioral (physical vulnerability, physical effort, physical activity, mental activity, help required, help available, mode of transportation, psychological vulnerability), and socioeconomic factors (age class, education, civil status, surrounding density, homeownership, home type, number of children and grandchildren, insecurity, neighborhood organization, social vulnerability) The model reached an average of 73% accuracy. • Fall risk impact: The classification model identified 24 predictors of falling (2 biological, 8 behavioral, 11 socioeconomic, and 3 environmental factors), each contributing 4.5–6.5% to the overall fall risk. Further work is needed to improve data quality, address multicollinearity, explore additional ML methods (e.g., SVM, neural networks), and enhance accuracy through hyperparameter tuning and larger datasets. Lo et al. (2019) 21 Machine learning • AI algorithm: Random Forest Clutter or obstacles; inadequate lighting; unsafe flooring; lack of assistive devices or improper use • Measures of environmental risks: assessed via OASIS-C structured items recorded by home health care clinicians during mandatory assessments age; sex; language group; borough of residence; number and severity of diagnoses; pain levels; cognitive impairment; visual impairment; frequency of assistance with ADL/IADL; therapy visit frequencies; living situation The OASIS model achieved a balanced accuracy of 0.62, an AUC-ROC of 0.67 (95% CI: 0.66–0.68), and an average precision of 0.10, outperforming the baseline MAHC-10 scoring system. • Fall risk impact: A random forest model using OASIS-C and EHR data provided improved fall risk prediction compared with the MAHC-10 tool, with age, severity of diagnoses, therapy visits, pain, and ADL/IADL assistance identified as top predictors supporting individualized fall-prevention strategies. The study relied on data from a single home care agency, and model precision remained low due to highly imbalanced outcomes; nearly 95% of patients in the cohort had no reported falls. Millet et al. (2023) 29 Machine learning + Natural language processing • AI algorithms: Random Forest; Decision Tree; Logistic Regression; LightGBM; Support Vector Machine (SVM); K-Nearest Neighbors (KNN) Stairs at home; teleassistance; home accessibility; elevator; stairs management (from selected 25 features) • Measures of environmental risks: Not clarified BMI; weight; age; gait speed; height; standing balance; walking cane; grip strength; sitting balance; foot support; tele-assistance; lift; antiplatelet drugs; cognitive impairment; anticoagulants; hip fracture; dizziness; depressive syndrome; fear; memory loss; stable turns; disorientation Best performance achieved by the Bagging ensemble using Random Forest, which attained 75.8% accuracy, 70.0% sensitivity, 80.5% specificity, and an AUC of 75.3%. Among base models, Random Forest alone performed strongly with 76.9% accuracy and 85.0% specificity, though with lower sensitivity (65.5%). Ensemble methods generally improved robustness and balance across metrics. • Fall risk impact: Accurate prediction of recurrent falls in older adults using routine clinical data. Reliance on retrospective data from a single hospital, the semi-structured nature of clinical notes requiring extensive NLP preprocessing, relatively low model sensitivity, and the absence of dynamic risk updates or validation in prospective or interventional settings. Moore et al. (2024) 26 Deep learning model • AI algorithm: YOLOv8 object detection algorithm Stairs; doorway; shower; sink; toilet; table; bed; signage; chair; animal; wet surface; mat/rug/carpet; obstacle (generic catch-all for potential obstructions); raised curbs • Measures of environmental risks: Not applicable Vehicle Using the collected dataset, the YOLOv8 algorithm was trained for a course of 100 epochs converging at epoch 69 within a timeframe of 4 h. The models achieved a best validation mAP50 of 0.81 at epoch 69, showcasing the potential of this algorithm within real-world deployment. • Fall risk impact: NA Only a small number of PwPD were recruited, and a limited dataset was curated as part of the pilot study. The number of participants for the focus group was small too and they were recruited based on their prior experience of participating in wearable gait research and their familiarity with technology. The focus group participants may have been biased toward the acceptance of technology as they were recruited by purposive sampling to have a good understanding and/or appreciation of commercial technology. Panyakaew et al. (2021) 28 • Type of AI: Machine learning • AI algorithm: XGBoost models for predicting (1) falls (fallers vs. non-fallers) and (2) recurrent falls (recurrent vs. non-recurrent fallers); SHAP used for model interpretability High-risk daily activities including sweeping the floor (ABC-16 item 7); reaching on tiptoes (item 5); walking in a crowded mall (item 12); walking across a parking lot (item 10); getting in/out of a car (item 9); walking up/down stairs (item 2); walking up/down a ramp (item 11) • Measures of environmental risks: Activities-Specific Balance Confidence Scale (ABC-16)—self-reported confidence (0–100%) in 16 daily activities H&Y stage 3, disease duration, age, PD subtype, Postural Instability and Gait Disorder (PIGD), wearing-off, dyskinesia, use of dopamine agonists, use of COMT inhibitors Falls model accuracy: 72% (p = 0.001) Recurrent falls model accuracy: 81% (p = 0.02) • Fall risk impact: The study successfully predicted fallers (72%) and recurrent fallers (81%). Top predictive environmental activities included: Sweeping the floor, Reaching on tiptoes, Walking in a crowded mall, Walking across a parking lot, and Getting in/out of a car. Exclusion of neurological co-morbidities; lack of details on anti-parkinsonian medication use; PD patients had relatively normal cognition with no neuropsychiatric/non-motor details. Only total MMSE score was available, with no breakdown of cognitive domains or executive function. Retrospective fall reporting; exclusion of patients with MMSE < 21; no objective balance measures; medication details lacking. Pérez-Ros et al. (2019) 25 Machine learning • AI algorithm: Binary logistic regression models Lack of stair handrails; poor stair design; lack of bathroom grab bars; dim lighting or glare; obstacles and tripping hazards; slippery or uneven surfaces; improper use of assistive devices • Measures of environmental risks: recorded as presence/absence of specific hazards via nurse-administered geriatric assessment at baseline Advanced age (≥ 80 years); previous falls; muscle weakness; gait and balance problems; poor vision; postural hypotension; chronic conditions (osteoarthritis, diabetes, stroke, Parkinson’s disease, incontinence, dementia); fear of falling (mFES); previous fractures; obesity (BMI ≥ 30 kg/m²); hearing problems; anxiety-depressive syndrome; medication use (alpha-blockers, benzodiazepines, beta-blockers, CNS-acting drugs); polypharmacy Isolated falls model: Sensitivity 10%, Specificity 98.2%, PPV 75.2%, NPV 77.5%, Nagelkerke R² = 47%, Overall correct classification = 77%. Recurrent falls model: Sensitivity 15.2%, Specificity 98.8%, PPV 58.8%, NPV 92.3%, Nagelkerke R² = 69.4%, Overall correct classification = 91.4%. • Fall risk impact: Predictive models identified that prior falls and use of alpha-blockers predicted isolated falls, while previous fractures, obesity (BMI ≥ 30 kg/m²), and use of benzodiazepines and beta-blockers predicted recurrent falls in independent older adults. These findings highlighted modifiable risk factors that may inform fall-prevention strategies. Drug-related data reflected only commonly prescribed pharmacotherapeutic groups; drug doses were not documented. Time, weather, and location of falls were not recorded. Follow-up was limited to 12 months. Self-reported monthly falls may be biased due to missing information. Medication adherence was not assessed. Six studies (66.7%) applied supervised machine learning models to structured clinical, functional, or survey data. Commonly used algorithms included logistic regression, random forest, and gradient boosting–based ensemble methods such as XGBoost, AdaBoost, and LightGBM. Several studies evaluated multiple classifiers with ensemble approaches. One study additionally incorporated natural language processing to extract features from narrative clinical text before model training. Three studies focused on vision- or robotics-based approaches to environmental perception rather than direct fall outcome prediction. These studies employed techniques such as simultaneous localization and mapping (SLAM), depth-based detection pipelines, and deep learning object detection models (e.g., YOLOv8) to identify environmental features and hazards in home-like settings. (1) Incorporation of Environmental Risk Factors The included studies identified a wide range of environmental hazards at home that contribute to fall risk, though methods and comprehensiveness of environmental assessments varied considerably. a. Structural and Housing Characteristics Lathouwers et al. 23 identified housing issues, housing change (i.e., homeownership or home type), and environmental vulnerability as significant predictors for fall occurrence, though specific measurement methods were not reported. Chen et al. 27 incorporated multiple housing characteristics into both fall prediction and fall-related injury prediction models through structured inquiry, including building structure (reinforced concrete versus other materials), presence of handicapped facilities, kitchen and flush toilet availability, Internet access, house tidiness (clear versus unclear), and house temperature (hot, neutral, cold). Millet et al. 29 assessed home-related features, including the presence of stairs, home accessibility, and elevator availability. b. Specific Environmental Risk Factors Perez-Ros et al. 25 documented specific hazards through nurse-administered geriatric assessment, including lack of stair handrails, poor stair design, lack of bathroom grab bars, dim lighting or glare, obstacles and tripping hazards, slippery or uneven surfaces, and improper use of assistive devices. Lo et al. 21 assessed environmental safety issues via structured assessment items completed by home health care clinicians, including clutter or obstacles, inadequate lighting, unsafe flooring, and lack of assistive devices or improper use. Du et al. 22 focused on lighting-related hazards, including poor lighting, cluttered or narrow spaces, obstructed walkways, and limited visibility, and established quantitative luminosity thresholds that were displayed via a color-coded interface. Moore et al. 26 identified 14 distinct environmental hazard categories using computer vision: stairs, doorways, showers, sinks, toilets, tables, beds, signage, chairs, animals, wet surfaces, mats/rugs/carpets, generic obstacles, and raised curbs. Panyakaew et al. 28 integrated environmental components into validated mobility and balance scales, including sweeping the floor, reaching on tiptoes, walking in a crowded mall, and walking up/down stairs and ramps, using activities that directly reflect environmental challenges. (2) Integration with Non-Environmental Risk Factors Across studies, commonly integrated domains included demographic characteristics, clinical conditions, functional status, cognitive and psychological factors, medication use, and social or socioeconomic variables. 21 , 23 , 25 , 27 Several models combined fall risk factors including age, prior falls, gait and balance impairments, chronic diseases, sensory deficits, and polypharmacy with psychosocial variables (e.g., depressive symptoms, fear of falling, social vulnerability) and functional dependence measures (Activities of Daily Living [ADL]/ Instrumental Activities of Daily Living [IADL]). 21 , 29 Some studies further incorporated disease-specific factors (e.g., Parkinson’s disease severity and motor complications) or physiological and performance-based measures (e.g., gait speed, grip strength, laboratory values). 26 , 28 AI Model Performance Among studies using supervised machine learning on structured data, discrimination was generally modest to good. Reported AUC-ROC values ranged from approximately 0.67 to 0.76, with Chen et al. 27 achieving the highest performance using logistic regression (AUC-ROC 0.739 for falls and 0.757 for fall-related injuries). Random forest–based models frequently outperformed traditional scoring tools, as demonstrated by Lo et al. 21 (0.62 vs. lower baseline performance), although precision remained limited due to highly imbalanced outcomes. Perez-Ros et al. 25 reported specificity exceeding 98% but sensitivity below 20%, indicating strong ability to rule in high-risk individuals but limited capacity to identify all fallers. Ensemble and gradient boosting approaches (e.g., XGBoost, LightGBM, bagging random forest) generally improved robustness and balance across performance metrics, with Millet et al. 29 reporting the best overall performance for recurrent fall prediction (AUC ≈ 0.75; accuracy ≈ 76%). Studies employing computer vision or robotics-based approaches primarily evaluated algorithmic feasibility rather than clinical prediction performance. Cloix et al. 24 demonstrated high environmental hazard recognition accuracy (> 90%) for stair and curb detection, while Moore et al. 26 reported strong object-detection performance using YOLOv8 (mAP50 = 0.81). Environmental factors consistently contributed to fall risk prediction when incorporated into AI-based models, although their representation and relative importance varied substantially by data source and modeling approach. Chen et al. 27 demonstrated that house tidiness emerged as one of the most important environmental predictors for both falls and fall-related injuries, alongside prior fall history. Similarly, Lo et al. 21 found that environmental hazards documented in OASIS-C assessments, such as inadequate lighting and household obstacles, contributed to improved discrimination compared with a standard clinical screening tool. Lathouwers et al. 23 further demonstrated that environmental vulnerability indicators contributed alongside biological, behavioral, and socioeconomic factors, with each environmental predictor accounting for a comparable proportion of overall fall risk. Discussion This systematic review provides a focused synthesis of AI-based fall risk prediction models that incorporate environmental factors among community-dwelling older adults, addressing a critical gap in the existing fall prevention literature. One observation from this study emerged during the search and screening stage. Despite identifying over 20,000 records through a comprehensive search, most studies were excluded because environmental factors were not incorporated into AI-based fall prediction models. Among the included studies, while personal, clinical, and functional factors appropriately formed the foundation of model development, environmental factors were typically included as part of broader predictor sets rather than being explicitly examined as distinct contributors to fall risk. These findings highlight a critical gap in AI-based fall prediction research and reveal environmental context as an underdeveloped yet potentially high-impact component. Notably, synthesis across diverse AI applications demonstrated that when environmental features were incorporated, they consistently contributed meaningful information to fall risk prediction, 21,23,27 either by improving model discrimination or by identifying actionable home hazards. These findings suggested that environmental factors are not peripheral but provide complementary and actionable information that enhances fall risk prediction. An important implication of these findings is that environmental factors differ from many personal or clinical predictors in that they are inherently modifiable. 30-32 While variables such as age, chronic conditions, or prior falls primarily support risk stratification, 33,34 environmental hazards can directly inform targeted prevention strategies. Lack of attention to the environmental context in current AI-based models may constrain both predictive insight and the translation of model outputs into actionable prevention strategies. Therefore, further study is warranted to explicitly incorporate environmental risk factors, which could strengthen the link between fall prediction and practical interventions in community settings. Fall hazards can be identified through traditional home environmental assessments conducted by healthcare providers during home visits, which are considered as a gold-standard approach. 35-37 However, their implementation is often constrained by substantial practical limitations. For example, comprehensive validated tools such as the Westmead Home Safety Assessment consist ofa 72-item checklist, 38 require considerable time and clinical effort and may yield variable results depending on provider expertise and perspective. 39 Additionally, access is limited for older adults in remote areas due to shortages of trained personnel, travel barriers, and economic constraints, 31,40 In parallel, this review identified substantial heterogeneity in how environmental risk factors are represented in AI-based models, ranging from simple binary indicators (e.g., presence of handrails) to high-dimensional sensor- or vision-derived data. Such inconsistency limits model comparability, replication, and cumulative knowledge building, ultimately constraining translation into practice. 41 Therefore, the need for AI-based approaches is highlighted as capable of supporting standardized, scalable, and objective environmental risk assessment. In this context, AI-driven methods, such as computer vision, image-based analysis, and multimodal data integration, offer promising opportunities to reduce assessment burden while enabling more standardized, precise, and usable representations of environmental risk information within fall prevention workflows. However, for real-word adoption in routine clinical practice, existing vision- and robotics have largely focused on algorithmic feasibility, have not yet demonstrated clinical performance, and have provided limited participant demographic information. Consequently, further validation across diverse clinical scenarios is necessary before clinical implementation. From an aging and community health perspective, the limited integration of environmental context in AI-based fall prediction models has important implications for equity, real-world relevance, and healthy aging. As people age, their immediate living environments play an increasingly central role in daily functioning, 42,43 making housing conditions and neighborhood infrastructure key determinants of fall risk. 44-46 These environmental conditions vary widely across populations and are closely linked to lifelong socioeconomic circumstances, geographic location, and access to supportive resources. 44-46 Consequently, AI models that insufficiently account for environmental risk may systematically overlook or underestimate key sources of vulnerability among community-dwelling older adults, particularly those aging in place within resource-limited settings. Such omissions risk reinforcing cumulative disadvantages in later life by limiting accurate risk identification and delaying preventive action, even when overall model performance appears acceptable. Incorporating environmental context into AI-based fall prediction models can therefore move these tools beyond risk stratification toward more context-aware and actionable insights, enabling earlier identification of modifiable hazards and supporting targeted interventions that promote safer aging in place and more equitable fall prevention across diverse aging populations. This systematic review has several limitations. First, although a comprehensive search strategy was employed across multiple databases, relevant studies may have been missed due to publication bias, indexing limitations, or exclusion criteria such as restriction to English-language publications and peer-reviewed articles, potentially excluding gray literature. Second, the small number of eligible studies reflects the emerging nature of AI-based fall prediction models that explicitly incorporate environmental factors, limiting the ability to draw definitive conclusions or conduct quantitative synthesis. Third, fall-related outcomes varied across studies, including any falls, recurrent falls, and fall-related injuries, with varying follow-up durations. This variability complicates comparison across models and may influence reported performance metrics. Fourth, several studies did not report external validation or relied on single datasets, raising concerns about model generalizability. The extent to which models incorporating environmental factors perform consistently across different populations and settings remains uncertain. Fifth, most included studies were conducted in high-income countries, potentially limiting the applicability of findings to low- and middle-income settings or to culturally diverse housing environments where environmental risks and living conditions may differ substantially. Finally, while some studies identified environmental hazards or demonstrated improved predictive performance, few evaluated whether incorporating environmental factors altered fall prevention interventions or reduced fall incidence, limiting conclusions about real-world effectiveness. Despite several limitations, this study has a number of important strengths. First, this study is the first systematic review to specifically examine how environmental factors are incorporated into AI-based fall risk prediction models. In doing so, we address a critical gap in the existing literature, which has largely focused on individual- and functional-level predictors. By isolating the environmental dimension, this review advances understanding of how modifiable home-related risks are currently conceptualized and operationalized within AI-based approaches. Second, this study provides a broad and integrative perspective on environmental risk representation by synthesizing evidence across diverse AI paradigms ranging from traditional machine learning models using structured data to computer vision and robotics-based approaches. Finally, this review provides clear evidence that environmental factors contribute meaningful information to AI-based fall risk prediction models. Future Implication Future research on AI-based fall prediction should address several important gaps identified in this review. (1) There is a clear need for standardized definitions and measurement frameworks for environmental risk factors, as current studies vary widely in how environmental hazards are conceptualized, assessed, and represented in models. Without greater consistency, comparison across studies and translation into practice remain limited. (2) Multimodal AI models that integrate clinical, functional, and environmental data are needed to more fully capture the multifactorial nature of fall risk. Falls often occur as the result of dynamic interactions among individual characteristics, functional limitations, and environmental hazards, rather than from any single factor alone. While many existing models emphasize individual- and functional-level predictors, incorporating environmental information alongside these factors may improve predictive performance and clinical relevance. Given that many environmental hazards are inherently visual and context-dependent, image-based or computer vision–driven approaches may be particularly valuable for identifying home environmental risks that are difficult to capture through structured questionnaires or clinical assessments alone. Advances in image recognition, object detection, and depth sensing offer promising opportunities to enhance environmental risk detection in real-world home settings, ultimately informing targeted interventions to reduce fall risk. (3) Environmental characteristics may vary significantly across different geographic regions and cultural contexts. Future research could further explore adaptive modeling approaches for environmental factors in multicultural settings, aiming to develop models that generalize better across regions. In addition, investigating potential fall risks arising from the interaction between cultural behaviors and environmental contexts may contribute to more inclusive, culturally sensitive, and accurate fall risk prediction. (4) Environmental features were generally assessed at a single time point and treated as static predictors. This approach does not reflect the dynamic nature of home environments, where hazards may change over time due to health decline, behavior adaptation, seasonal conditions, or home modifications. Future AI models may benefit from longitudinal or continuously updated environmental data to better reflect real-world fall risk trajectories. (5) Future AI-based fall prediction tools should prioritize scalability, usability, and real-world applicability, ensuring that models can be feasibly deployed in community settings and adapted to diverse home environments. Approaches that leverage mobile devices, smart home technologies, or automated image-based assessments may help bridge the gap between predictive modeling and actionable fall prevention strategies. Conclusions This review shows that environmental factors remain infrequently emphasized in AI-based fall prediction models, despite their established relevance to fall risk among community-dwelling older adults. Environmental context provides complementary and actionable information that supports the need for more standardized, context-aware AI approaches to equitable and effective fall prevention. Declarations Acknowledgements [Financial support] This study is funded by the Boston College Connell School of Nursing Intramural Pilot Grant (SAFE-AI Phase I (Synergistic Approach to Fall-risk Evaluation in older adults using generative AI). PI: Jung). Jiyoun Song is funded by the NHLBI (1K99HL169940). The content is solely the responsibility of the authors and does not necessarily represent the official views of the NHLBI. [Conflict of interest] All authors report no conflicts of interest relevant to this article. [Ethical Conduct of Research] This systematic review utilized only publicly available published data and did not involve human subjects. Therefore, institutional review board (IRB) approval was not required. [Declaration of Generative AI and AI-assisted Technologies in the Writing Process] During the preparation of this work, the authors used WordTune and Grammarly for editorial support. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication. [Clinical Trial Number] Clinical trial number: not applicable. References Center for Disease Control and Prevention. National Center for Injury Prevention and Control: Web–based Injury Statistics Query and Reporting System (WISQARS). Accessed Nov 20th, 2025. https://wisqars.cdc.gov/ Thompson HJ, McCormick WC, Kagan SH. Traumatic brain injury in older adults: epidemiology, outcomes, and future implications. Journal of the American Geriatrics Society . Oct 2006;54(10):1590–5. doi:10.1111/j.1532-5415.2006.00894.x Sterling DA, O’Connor JA, Bonadies J. Geriatric Falls: Injury Severity Is High and Disproportionate to Mechanism. Journal of Trauma and Acute Care Surgery . 2001;50(1) Vaishya R, Vaish A. Falls in Older Adults are Serious. Indian Journal of Orthopaedics . 2020/02/01 2020;54(1):69–74. doi:10.1007/s43465-019-00037-x Montero-Odasso M, van der Velde N, Martin FC, et al. World guidelines for falls prevention and management for older adults: a global initiative. Age Ageing . Sep 2 2022;51(9)doi:10.1093/ageing/afac205 Jahandideh S, Hutchinson AF, Bucknall TK, et al. Using machine learning models to predict falls in hospitalised adults. International journal of medical informatics . 2024/07/01/ 2024;187:105436. doi:https://doi.org/10.1016/j.ijmedinf.2024.105436 Usmani S, Saboor A, Haris M, Khan MA, Park H. Latest Research Trends in Fall Detection and Prevention Using Machine Learning: A Systematic Review. Sensors . 2021;21(15):5134. doi:10.3390/s21155134 Bargiotas I, Wang D, Mantilla J, et al. Preventing falls: the use of machine learning for the prediction of future falls in individuals without history of fall. Journal of Neurology . 2023/02/01 2023;270(2):618–631. doi:10.1007/s00415-022-11251-3 Dubois A, Bihl T, Bresciani J-P. Identifying Fall Risk Predictors by Monitoring Daily Activities at Home Using a Depth Sensor Coupled to Machine Learning Algorithms. Sensors . 2021;21(6):1957. doi:10.3390/s21061957 Lima CA, Ricci NA, Nogueira EC, Perracini MR. The Berg Balance Scale as a clinical screening tool to predict fall risk in older adults: a systematic review. Physiotherapy . 2018/12/01/ 2018;104(4):383–394. doi:https://doi.org/10.1016/j.physio.2018.02.002 Strini V, Schiavolin R, Prendin A. Fall Risk Assessment Scales: A Systematic Literature Review. Nurs Rep . Jun 2 2021;11(2):430–443. doi:10.3390/nursrep11020041 González-Castro A, Leirós-Rodríguez R, Prada-García C, Benítez-Andrades JA. The Applications of Artificial Intelligence for Assessing Fall Risk: Systematic Review. Journal of medical Internet research . 2024/4/29 2024;26:e54934. doi:10.2196/54934 Liu S, Si Y, Peng Y, Wang D, Yuan X, Peng Y. Incidence and risk factors of falls in older people with chronic comorbidities in community: a cross-sectional study. Original Research. Frontiers in Public Health . 2025–September–24 2025;Volume 13 - 2025doi:10.3389/fpubh.2025.1643699 Li Y, Hou L, Zhao H, Xie R, Yi Y, Ding X. Risk factors for falls among community-dwelling older adults: A systematic review and meta-analysis. Front Med (Lausanne) . 2022;9:1019094. doi:10.3389/fmed.2022.1019094 Saunders S, D'Amore C, Hao Q, et al. Risk Factors for Falls in Community-Dwelling Older Adults: An Umbrella Review. Journal of the American Medical Directors Association . 2025;26(9)doi:10.1016/j.jamda.2025.105765 Stevens JA, Mahoney JE, Ehrenreich H. Circumstances and outcomes of falls among high risk community-dwelling older adults. Injury Epidemiology . 2014/03/20 2014;1(1):5. doi:10.1186/2197-1714-1-5 Moreland BL, Kakara R, Haddad YK, Shakya I, Bergen G. A Descriptive Analysis of Location of Older Adult Falls That Resulted in Emergency Department Visits in the United States, 2015. American Journal of Lifestyle Medicine . 2021/11/01 2020;15(6):590–597. doi:10.1177/1559827620942187 Rubenstein LZ. Falls in older people: epidemiology, risk factors and strategies for prevention. Age Ageing . Sep 2006;35 Suppl 2:ii37–ii41. doi:10.1093/ageing/afl084 Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ . 2021;372:n71. doi:10.1136/bmj.n71 The National Heart L, and Blood Institute. Study Quality Assessment Tools. Accessed Nov, 25th, 2024. https://www.nhlbi.nih.gov/health-topics/study-quality-assessment-tools Lo Y, Lynch SF, Urbanowicz RJ, et al. Using Machine Learning on Home Health Care Assessments to Predict Fall Risk. Studies in health technology and informatics . Aug 21 2019;264:684–688. doi:10.3233/shti190310 Du R, Jagtap V, Long Y, Onwuka O, Padir T. Robotics enabled in-home environment screening for fall risks. presented at: Proceedings of the 2014 workshop on Mobile augmented reality and robotic technology-based systems; 2014; Bretton Woods, New Hampshire, USA. https://doi.org/10.1145/2609829.2609836 Lathouwers E, Dillen A, Díaz MA, et al. Characterizing fall risk factors in Belgian older adults through machine learning: a data-driven approach. BMC Public Health . Nov 29 2022;22(1):2210. doi:10.1186/s12889-022-14694-5 Cloix S, Bologna G, Weiss V, Pun T, Hasler D. Low-power depth-based descending stair detection for smart assistive devices. EURASIP Journal on Image and Video Processing . 2016/10/12 2016;2016(1):33. doi:10.1186/s13640-016-0133-6 Pérez-Ros P, Martínez-Arnau FM, Orti-Lucas RM, Tarazona-Santabalbina FJ. A predictive model of isolated and recurrent falls in functionally independent community-dwelling older adults. Braz J Phys Ther . Jan–Feb 2019;23(1):19–26. doi:10.1016/j.bjpt.2018.05.005 Moore J, Celik Y, Stuart S, et al. Using Video Technology and AI within Parkinson’s Disease Free-Living Fall Risk Assessment. Sensors . 2024;24(15):4914. Chen X, He L, Shi K, Wu Y, Lin S, Fang Y. Interpretable Machine Learning for Fall Prediction Among Older Adults in China. Am J Prev Med . Oct 2023;65(4):579–586. doi:10.1016/j.amepre.2023.04.006 Panyakaew P, Pornputtapong N, Bhidayasiri R. Using machine learning-based analytics of daily activities to identify modifiable risk factors for falling in Parkinson's disease. Parkinsonism Relat Disord . Jan 2021;82:77–83. doi:10.1016/j.parkreldis.2020.11.014 Millet A, Madrid A, Alonso-Weber JM, Rodríguez-Mañas L, Pérez-Rodrá-Guez R. Machine Learning Techniques Applied to the Development of a Fall Risk Index for Older Adults. IEEE Access . 2023;11:84795–84809. doi:10.1109/ACCESS.2023.3299489 Campani D, Caristia S, Amariglio A, et al. Home and environmental hazards modification for fall prevention among the elderly. Public health nursing (Boston, Mass) . May 2021;38(3):493–501. doi:10.1111/phn.12852 Pynoos J, Steinman BA, Nguyen AQ. Environmental assessment and modification as fall-prevention strategies for older adults. Clinics in geriatric medicine . Nov 2010;26(4):633–44. doi:10.1016/j.cger.2010.07.001 Faulkner S, Burns D. Identifying modifiable environmental risk factors for falls in older people. Injury Prevention . 2010;16(Suppl 1):A254. doi:10.1136/ip.2010.029215.905 Montero-Odasso M, Pieruccini-Faria F, Son S, et al. Fall risk stratification in older adults: low and not-at-risk status still associated with falls and injuries. Age Ageing . Mar 3 2025;54(3)doi:10.1093/ageing/afaf064 Tang S, Liu M, Yang T, et al. Association between falls in elderly and the number of chronic diseases and health-related behaviors based on CHARLS 2018: health status as a mediating variable. BMC Geriatr . Apr 28 2022;22(1):374. doi:10.1186/s12877-022-03055-x Gillespie LD, Robertson MC, Gillespie WJ, et al. Interventions for preventing falls in older people living in the community. Cochrane Database of Systematic Reviews . 2012;(9)doi:10.1002/14651858.CD007146.pub3 Ziebart C, Dewan N, Tuazon J, MacDermid J. Development of the Home Fall Hazard Checklist. Rehabilitation Research and Practice . 2021/06/01 2021;2021:5362197. doi:10.1155/2021/5362197 Campani D, Caristia S, Amariglio A, et al. Home and environmental hazards modification for fall prevention among the elderly. Public Health Nursing . 2021/05/01 2021;38(3):493–501. doi:https://doi.org/10.1111/phn.12852 Clemson L, Fitzgerald MH, Heard R. Content Validity of an Assessment Tool to Identify Home Fall Hazards: The Westmead Home Safety Assessment. British Journal of Occupational Therapy . 1999/04/01 1999;62(4):171–179. doi:10.1177/030802269906200407 Pighills A, Drummond A, Crossland S, Torgerson DJ. What type of environmental assessment and modification prevents falls in community dwelling older people? Bmj . Mar 14 2019;364:l880. doi:10.1136/bmj.l880 Sadasivam RS, Luger TM, Coley HL, et al. Robot-assisted home hazard assessment for fall prevention: a feasibility study. Journal of Telemedicine and Telecare . 2014/01/01 2013;20(1):3–10. doi:10.1177/1357633X13517350 Blanchet R, Edwards N. A need to improve the assessment of environmental hazards for falls on stairs and in bathrooms: results of a scoping review. BMC Geriatrics . 2018/11/09 2018;18(1):272. doi:10.1186/s12877-018-0958-1 Jiang Y-S, Shi H, Kang Y-T, et al. Impact of age-friendly living environment and intrinsic capacity on functional ability in older adults: a cross-sectional study. BMC Geriatrics . 2023/06/17 2023;23(1):374. doi:10.1186/s12877-023-04089-5 Bayar R, Türkoğlu H. The relationship between living environment and daily life routines of older adults. A/Z ITU Journal of the Faculty of Architecture . 2021;18(1):29–43. Olden K, White SL. Health-Related Disparities: Influence of Environmental Factors. Medical Clinics . 2005;89(4):721–738. doi:10.1016/j.mcna.2005.02.001 Gee GC, Payne-Sturges DC. Environmental health disparities: a framework integrating psychosocial and environmental concepts. Environ Health Perspect . Dec 2004;112(17):1645–53. doi:10.1289/ehp.7074 Jacobs DE. Environmental Health Disparities in Housing. American Journal of Public Health . 2011/12/01 2011;101(S1):S115–S122. doi:10.2105/AJPH.2010.300058 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 05 Feb, 2026 Editor assigned by journal 30 Jan, 2026 Submission checks completed at journal 30 Jan, 2026 First submitted to journal 28 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8723907","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Systematic Review","associatedPublications":[],"authors":[{"id":583062461,"identity":"56ac8919-fc67-45d7-831a-ea9332bea2ba","order_by":0,"name":"Jiyoun Song","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABD0lEQVRIiWNgGAWjYLACCQMGOTAjAUmQsYGAFmMStQBBIoYCnFrk23sPv7AoOJy+XSL38YeHO2oZDG43H3tcwGAju+EAdi0GZ86lWUgYHM7dOSPdTCLxzHEGgzvH0o1nMKQZ49QikWNmANKy4UYaG0Ni2zEGgxs5ZtI8DIcTcWmRnwHRkm5wI435A0RL/jeglv84tTDcyDF+ANSSANTCIJHYVgOyhQ2o5QBOLQZnzpgBAzndcMOZZ2xALQd4JO8cAzrMINl4Ji6HtfcYf5b4Yy1vcDyN+ePPtjo5vtvNz6R5Kuxk+3A5jIGBTVqCoRnGOczDIAG2HadyEGD++IGhDsYBMiTwqh4Fo2AUjIIRCAAwBl7z4tUH6QAAAABJRU5ErkJggg==","orcid":"","institution":"University of Pennsylvania School of Nursing","correspondingAuthor":true,"prefix":"","firstName":"Jiyoun","middleName":"","lastName":"Song","suffix":""},{"id":583062462,"identity":"2783ffeb-9e0d-48f6-99a8-b3e045020737","order_by":1,"name":"Boeun Kim","email":"","orcid":"","institution":"University of Iowa College of Nursing","correspondingAuthor":false,"prefix":"","firstName":"Boeun","middleName":"","lastName":"Kim","suffix":""},{"id":583062463,"identity":"c09c18be-7a6a-46ef-bbbe-b324ed40fc48","order_by":2,"name":"Min-Jeoung Kang","email":"","orcid":"","institution":"Harvard Medical School","correspondingAuthor":false,"prefix":"","firstName":"Min-Jeoung","middleName":"","lastName":"Kang","suffix":""},{"id":583062464,"identity":"48d0d6be-2bc0-43f0-92e9-f573261dd576","order_by":3,"name":"Shuxuan Li","email":"","orcid":"","institution":"University of Pennsylvania School of Social Policy and Practice","correspondingAuthor":false,"prefix":"","firstName":"Shuxuan","middleName":"","lastName":"Li","suffix":""},{"id":583062465,"identity":"f3e6871e-932d-4e81-9fbe-86171b387a30","order_by":4,"name":"Lingjie Liu","email":"","orcid":"","institution":"University of Pennsylvania School of Engineering","correspondingAuthor":false,"prefix":"","firstName":"Lingjie","middleName":"","lastName":"Liu","suffix":""},{"id":583062466,"identity":"d3e7396d-091c-4383-9aba-fc3ee7258486","order_by":5,"name":"Wonkyung Jung","email":"","orcid":"","institution":"Boston College Connell School of Nursing","correspondingAuthor":false,"prefix":"","firstName":"Wonkyung","middleName":"","lastName":"Jung","suffix":""}],"badges":[],"createdAt":"2026-01-28 17:10:11","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8723907/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8723907/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":101633529,"identity":"4f91bb2d-3e50-4a4f-88e8-6645692efe76","added_by":"auto","created_at":"2026-02-02 06:03:10","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":81029,"visible":true,"origin":"","legend":"\u003cp\u003eAn Overview of the Review Process (PRISMA Diagram)\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8723907/v1/616b15cda9bf48f3c28324ba.png"},{"id":101633528,"identity":"e0c35391-cf65-48e0-9511-cb1246a1b171","added_by":"auto","created_at":"2026-02-02 06:03:10","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":352827,"visible":true,"origin":"","legend":"\u003cp\u003eQuality Appraisal\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8723907/v1/24c7d8c3796bc30387afe514.png"},{"id":101633530,"identity":"cfac4dec-5b44-45e9-90a2-6de512bcf21b","added_by":"auto","created_at":"2026-02-02 06:03:16","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1670680,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8723907/v1/5e732ec7-38e8-4d91-9328-dc99e93af43f.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Are We Missing the Environmental Factors in AI-Based Fall Risk Models?: A Systematic Review","fulltext":[{"header":"Introduction","content":"\u003cp\u003eFalls pose a critical challenge to healthy aging and remain one of the leading causes of injury among community-dwelling older adults.\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e The incidence of falls continues to rise as the global population ages, with approximately one in four older adults experiencing at least one fall each year.\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e Among older adults, falls are associated with serious physical, psychological, and social consequences, including fractures, traumatic brain injuries, functional decline, and long-term disability.\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e Older adults aged 75 years and older experience the highest rates of fall-related hospitalization and are particularly vulnerable to prolonged recovery and subsequent loss of independence.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e These outcomes directly undermine well-being and healthy aging and contribute to greater healthcare utilization, long-term care needs, and economic burden.\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eRecent advances in artificial intelligence (AI) and machine learning (ML) have generated growing interest in their application to fall risk prediction and prevention. A rapidly expanding body of literature has explored AI-driven approaches for fall prediction, including machine learning algorithms, deep learning models, and sensor- or wearable-based systems.\u003csup\u003e\u003cspan additionalcitationids=\"CR7 CR8\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e Traditionally, fall risk assessment has relied on standardized clinical tools and screening instruments developed to identify individuals at elevated risk. Commonly used tools include the Berg Balance Scale (BBS) to assess functional balance,\u003csup\u003e10\u003c/sup\u003e Functional Reach Test, and various fall risk questionnaires that assess fall history, mobility limitations, medication use, and comorbidities.\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e These instruments have been widely implemented in clinical and community settings given their simplicity and straightforward administration. Data derived from these traditional assessment tools have also been used as input features in AI\u0026ndash;based fall prediction models, either independently or in combination with additional clinical, functional, and sensor-derived variables.\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e However, despite their widespread adoption, traditional fall risk assessment tools demonstrate significant limitations in capturing dynamic and environmental determinants of fall risk\u003c/p\u003e \u003cp\u003eFalls result from the interaction of multiple factors, including individual characteristics, physical and psychological impairments, and environmental conditions.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e Accurate fall prediction, therefore, requires approaches capable of integrating comprehensive, diverse and interrelated risk factors. While individual-level risk factors (e.g., age, comorbidities, medication use, and prior falls) and functional-level risk factors (e.g., balance, mobility, and functional limitations) have been extensively studied,\u003csup\u003e14,15\u003c/sup\u003e home environmental hazards remain relatively understudied despite being an integral and modifiable component of fall risk among community-dwelling older adults. Notably, a majority of falls occur at home, contributing to 79.2% of fall-related emergency department visits,\u003csup\u003e16,17\u003c/sup\u003e underscoring the central role of home settings in fall prevention efforts. Common hazards include inadequate lighting, slippery surfaces, loose rugs, unsafe handrails, and cluttered pathways,\u003csup\u003e4,18\u003c/sup\u003e which often interact with age-related functional limitations, increasing fall risk and highlighting the need for systematic approaches to environmental risk identification and mitigation.\u003c/p\u003e \u003cp\u003eAI-based models offer unique advantages in handling complex, high-dimensional data and identifying non-linear interactions among multiple risk factors. Despite these advances, it remains unclear to what existing AI-based fall prediction models account for environmental risk factors which play a modifiable role in fall risk. To address this knowledge gap, this study aims to systematically synthesize the existing literature on the inclusion and representation of environmental factors in AI-based fall risk prediction models, and summarize the algorithms and performance metrics reported in studies of falls among community-dwelling older adults.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eThis systematic review followed the guidelines of Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA)\u003csup\u003e19\u003c/sup\u003e The protocol was registered in the PROSPERO (CRD420261279513).\u003c/p\u003e\n\u003ch2\u003eEligibility Criteria\u003c/h2\u003e\n\u003ch2\u003e(1)\u0026nbsp;\u0026nbsp;Inclusion Criteria\u003c/h2\u003e\n\u003cp\u003eStudies were eligible for inclusion if they met the following criteria: (1) focused on community-dwelling older adults aged 60 and above from any country; (2) utilized AI or technology-based models to assess, predict, or prevent fall risks, including machine learning models or mobile health technologies; (3) assessed or predicted fall risk as a primary outcome; (4) incorporated environmental factors as one of the input variables or predictors in the model; (5) were data-based primary studies, including model development, validation, and evaluation studies reporting empirical data on model outcomes, accuracy, or effectiveness; and (6) were written in English.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e(2)\u0026nbsp;\u0026nbsp;Exclusion Criteria\u003c/h2\u003e\n\u003cp\u003eStudies were excluded if they: (1) focused on institutionalized populations such as nursing home residents; (2) employed non-technology or non-AI-based approaches, such as manual assessments or traditional fall risk screening methods; (3) were qualitative studies, dissertations, conference abstracts, editorials, commentaries, or literature reviews.\u003c/p\u003e\n\u003ch2\u003eSearch Strategy\u003c/h2\u003e\n\u003cp\u003eThe literature search strategy was developed in collaboration with a librarian and conducted across six electronic databases\u0026nbsp;(i.e.,\u0026nbsp;PubMed, Embase, CINAHL, Cochrane Library, Web of Science, and Scopus)\u0026nbsp;on\u0026nbsp;December\u0026nbsp;19, 2025. In addition, a manual review of the reference lists of included studies was conducted to identify additional eligible articles. No publication date limits were applied to ensure comprehensive retrieval of all relevant articles.\u003c/p\u003e\n\u003cp\u003eA comprehensive search methodology was designed using both free-text keywords and Medical Subject Headings (MeSH) terms (\u003cstrong\u003eAppendix 1\u003c/strong\u003e). The search strategy incorporated three major concept groups combined with Boolean operators: (1) Falls-related terms (e.g., \"Accidental Falls\"[MeSH], fall, falls, falling, fallen); (2) AI and technology-related terms (e.g., \"Artificial Intelligence\"[Mesh] machine learning, deep learning, telemedicine, digital health, mHealth, sensors, smart home technology); and (3) Older adult-related terms (e.g., aged[MeSH], older adult, elder*, senior*, geriatric*).\u003c/p\u003e\n\u003ch2\u003eStudy Selection and Data Extraction\u003c/h2\u003e\n\u003cp\u003eSearch results were imported into Covidence (Veritas Health Innovation, Melbourne, Australia), and duplicates were removed prior to screening. Four reviewers [JS, BK, MK, and WJ] conducted a two-stage screening process. In the title and abstract screening stage, two reviewers independently assessed each record, and potentially relevant articles advanced to full-text review. During the full-text assessment, two reviewers again independently evaluated each article against eligibility criteria. At both stages, disagreements were resolved through consensus with a third reviewer who had not participated in the initial assessment.\u003c/p\u003e\n\u003cp\u003eIn order to finalize the form, two authors created a standardized data extraction form and pilot-tested it on one article before finalizing the form. Each included study was independently extracted by two reviewers, with discrepancies resolved by a third reviewer. Extracted information included publication details (author, year, country, data source), study characteristics (aim, design, sample size, inclusion/exclusion criteria, recruitment), details of the AI model (type of approach, algorithms used, input variables), fall risk factors (including environmental factors and how they were measured)\u0026nbsp;and model performance metrics (accuracy, sensitivity, specificity, AUC-ROC, F-score). Study limitations and potential conflicts of interest were also recorded.\u003c/p\u003e\n\u003cp\u003eFindings were synthesized into evidence tables summarizing study characteristics, AI approaches and risk factors, and model performance outcomes.\u003c/p\u003e\n\u003ch2\u003eQuality Appraisal\u003c/h2\u003e\n\u003cp\u003e\u0026nbsp;The methodological quality of included observational studies was assessed using the National Heart, Lung, and Blood Institute (NHLBI) Quality Assessment Tool for Observational Cohort and Cross-sectional Studies.\u003csup\u003e20\u003c/sup\u003e The tool consists of 14 criteria evaluating clarity of research objectives, definition of the study population, participation rates, consistency of recruitment methods, adequacy of sample size justification, temporal relationship between exposure and outcome measurement, validity and reliability of exposure and outcome measures, blinding of outcome assessors, completeness of follow-up, and control for confounding variables. Each study was independently appraised by two reviewers, with discrepancies resolved through discussion with a third reviewer.\u003c/p\u003e\n\u003ch2\u003eEthical Considerations\u003c/h2\u003e\n\u003cp\u003eThis systematic review utilized only publicly available published data and did not involve human subjects. Therefore, institutional review board (IRB) approval was not required.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eAn overview of the review process is depicted in Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e. A total of 21,760 articles were identified through the initial literature search, including PubMed (n\u0026thinsp;=\u0026thinsp;2,185), Embase (n\u0026thinsp;=\u0026thinsp;3,804), CINAHL (n\u0026thinsp;=\u0026thinsp;948), Cochrane Library (n\u0026thinsp;=\u0026thinsp;854), Web of Science (n\u0026thinsp;=\u0026thinsp;5,981), and Scopus (n\u0026thinsp;=\u0026thinsp;7,988). No additional studies were identified through manual searching of reference lists and research reports. After the removal of duplicate articles (n\u0026thinsp;=\u0026thinsp;8,918), the remaining articles (n\u0026thinsp;=\u0026thinsp;12,842) were screened. Of these, 12,817 were excluded after title and abstract review. During the full-text screening, an additional 16 articles were excluded for not meeting the inclusion criteria. Throughout the exclusion process, the majority of excluded studies did not include environmental risk factors, which were a key inclusion criterion. Consequently, a total of nine studies were included.\u003c/p\u003e\n\u003cp\u003eOverall, the methodological quality of included studies was variable (Fig. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e). Most studies clearly stated research objectives and adequately defined study populations, but common limitations included incomplete reporting of sample size justification, limited handling of confounding, and insufficient information on exposure and outcome measurement reliability. Studies using vision- and sensing-based approaches, which rely on unstructured visual or depth data captured in real-world environments, showed greater uncertainty across several quality domains compared with structured-data\u0026ndash;based machine learning studies, which used predefined, tabular inputs.\u003c/p\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\n \u003ch2\u003eOverview of Included Studies\u003c/h2\u003e\n \u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e summarized the overview of included studies. The included studies were conducted across multiple countries, with the majority originating from the United States\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e and European countries (Belgium,\u003csup\u003e23\u003c/sup\u003e Switzerland,\u003csup\u003e24\u003c/sup\u003e Spain,\u003csup\u003e25\u003c/sup\u003e and the United Kingdom\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e), alongside representation from Asian countries including China\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e and Thailand.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003cp\u003eSix studies (66.7%) used structured datasets, such as electronic health record (EHR), home health assessments or community-based registries, to train machine-learning models to predict future falls.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e23\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e25\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e27\u003c/span\u003e\u0026ndash;\u003cspan class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e In these studies, longitudinal follow-up periods and large samples were used to evaluate predictive performance across a wide range of demographic, clinical, and environmental variables. These included studies encompassed diverse populations with sample sizes ranging from 304 to 59,028 participants. The largest study by Lo et al.\u003csup\u003e21\u003c/sup\u003e analyzed data from 59,028 unique patients aged over 65 years using retrospective EHR with available OASIS-C assessment and MAHC-10 fall risk assessment data. Lathouwers et al.\u003csup\u003e23\u003c/sup\u003e examined 82,580 community-dwelling older adults aged 60 and over in Belgium, recruited through stratified random sampling by sex and age from census data; after excluding missing data, 24 input features and 33,346 entries remained for analysis.\u003c/p\u003e\n \u003cp\u003eIn contrast, three studies used sensor- or computer-vision\u0026ndash;based experimental designs. \u003csup\u003e\u003cspan class=\"CitationRef\"\u003e22\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e24\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e Du et al. employed robotics-enabled environmental screening incorporating luminosity and spatial measurements,\u003csup\u003e22\u003c/sup\u003e while Cloix et al. validated depth-based stair detection using controlled motion capture settings.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e Moore et al. developed AI algorithms within simulated environments or structured testing conditions to detect environmental hazards or risky activities related to falls.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e These three studies did not provide detailed participant recruitment or demographic information.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eOverview of Included Studies\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"6\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAuthor (Year of publication)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eStudy region\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eStudy design\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAim\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDate Source\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eParticipants characteristics\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eChen et al. (2023)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eChina\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCohort study\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTo build prediction models for falls and fall-related injuries among community-dwelling older adults in China\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eChina Health and Retirement Longitudinal Study (CHARLS), a national cohort of middle-aged and older Chinese adults (mainly\u0026thinsp;\u0026ge;\u0026thinsp;45 years) across 28 provinces. Launched in 2011 with follow-up in 2013, 2015, and 2018. This study used data from the latest two waves (2015 and 2018).\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026bull; Recruitment: Not applicable.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Inclusion criteria: Not applicable.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Exclusion criteria: (1) aged\u0026thinsp;\u0026lt;\u0026thinsp;60 years, (2) living in a care home, (3) did not participate in 2015 and 2018 follow-ups, (4) incomplete information about falls and blood indices, and (5)\u0026thinsp;\u0026gt;\u0026thinsp;10% missing data in individual variables.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Total participants: 5,818.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCloix et al. (2016)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSwitzerland\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOther: Technology Evaluation Study\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTo evaluate a low-complexity algorithm to detect descending stairs and curbs of any shape, specifically designed for low-power real-time embedded platforms\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eThe cameras were located at 78 cm height with a tilted angle of 35\u0026deg;. Images were captured at 512 \u0026times; 384 and 640 \u0026times; 480 pixel resolution with the stereo camera and the Red Green Blue-Depth (RGBD) sensor.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026bull; The assessment of our approach was carried out using thirteen scenes of descending stairs and curbs.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Recruitment: Not applicable.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Inclusion criteria: Not applicable.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Exclusion criteria: Not applicable.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Total participants: Not applicable.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDu et al. (2014)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUSA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOther: Engineering Feasibility and Prototype Validation Study\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTo develop a robot-based system for in-home environment screening that supports both manual and autonomous fall risk assessment.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eReal-time sensor data collected by a mobile robot (Turtlebot2) in a simulated in-home environment.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026bull; Recruitment: Not applicable.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Inclusion criteria: Not applicable.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Exclusion criteria: Not applicable.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Total participants: Not applicable.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLathouwer et al. (2022)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBelgium\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCohort study\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTo identify fall risk factors using a quality-of-life questionnaire covering biological, behavioral, environmental, and socio-economic domains.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eQuestionnaire developed by the Belgian Ageing Studies research group of the Vrije Universiteit Brussel.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026bull; Recruitment: Stratified random sampling (by sex and age) from census data of community-dwelling adults aged\u0026thinsp;\u0026ge;\u0026thinsp;60 in Belgian municipalities.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Inclusion criteria: Community-dwelling older adults aged\u0026thinsp;\u0026ge;\u0026thinsp;60 living in Belgium.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Exclusion criteria: Not applicable.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Total participants: 33,346 community-dwelling older adults; 84 questionnaire items excluded due to missing data, resulting in 24 input features and 33,346 valid entries.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLo et al. (2019)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUSA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCross sectional study\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTo devise a machine learning pipeline using existing home health care data to predict fall risk among older adults, with the goal of building clinical decision support tools for fall prevention\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOutcome and Assessment Information Set (OASIS-C) \u0026ndash; mandatory assessment for home health care patients; Electronic Health Records (EHR) \u0026ndash; supplemental demographic data.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026bull; Recruitment: Not applicable (retrospective EHR-based study).\u003c/p\u003e\n \u003cp\u003e\u0026bull; Inclusion criteria: Age\u0026thinsp;\u0026ge;\u0026thinsp;65 years; available OASIS-C assessment data and MAHC-10 fall risk assessment data.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Exclusion criteria: Not applicable.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Total participants: 59,028 unique patients.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMillet et al. (2023)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSpain\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCohort study\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTo predict recurrent falls in the older population using machine learning techniques, with the aim of reducing the number of falls and their consequences.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGetafe University Hospital\u0026rsquo;s Geriatric Falls Unit (data obtained from the Hospital\u0026rsquo;s Electronic Health Records) between January 2017 and December 2021. Older adults who experienced a fall requiring medical attention were referred to this unit for personalized treatment and monitoring to prevent repeated falls.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026bull; Recruitment: Not applicable.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Inclusion criteria: Adults who were treated at the Geriatric Falls Unit for at least one fall (mean age 80.3\u0026plusmn;7.7).\u003c/p\u003e\n \u003cp\u003e\u0026bull; Exclusion criteria: Not applicable.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Total participants: 304 older adults.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMoore et al. (2024)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUK\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOther: A mixed-method study\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTo develop an AI algorithm using wearable video-based eye-tracking and IMU gait data, and to explore how people with Parkinson\u0026rsquo;s disease perceive the use of video and AI technologies in daily life for understanding their fall risk.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFor developing the AI model, pre-trained MS COCO weights were initialized and fine-tuned on a local dataset (80:20 split). Ten hours of local video data (\u0026gt;\u0026thinsp;1500 frames) were manually extracted and annotated using the LabelImg tool, with 4 categories and 18 classes selected for relevance to fall risk or privacy. A total of 1542 frames were selected and annotated, including: (1) scripted route data from 7 young adults covering 10 interior and 10 exterior environments; (2) 240 manually extracted frames from 3 PwPD, ensuring unseen data for testing; and (3) 4 additional first-person-view exterior videos (~\u0026thinsp;240 min) downloaded under CC-BY licence and annotated.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026bull; Recruitment:\u003c/p\u003e\n \u003cp\u003e\u0026ndash; PwPD participants were recruited locally through networks within Northumbria University.\u003c/p\u003e\n \u003cp\u003e\u0026ndash; Seven young adults (6M:1F, 23\u0026ndash;32 years) were recruited to wear video glasses solely for generating AI training data; young adults were recruited through word of mouth.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Inclusion criteria:\u003c/p\u003e\n \u003cp\u003e\u0026ndash; For PwPD:\u003c/p\u003e\n \u003cp\u003e\u0026bull; Clinical diagnosis of Parkinson\u0026rsquo;s disease;\u003c/p\u003e\n \u003cp\u003e\u0026bull; Prior experience in wearable-gait research;\u003c/p\u003e\n \u003cp\u003e\u0026bull; Familiarity with technology (e.g., smartphones, tablets, applications);\u003c/p\u003e\n \u003cp\u003e\u0026bull; Willingness to attend focus groups and consent to audio-recording;\u003c/p\u003e\n \u003cp\u003e\u0026bull; English-speaking and literate.\u003c/p\u003e\n \u003cp\u003e\u0026ndash; For young adults: Not formally specified beyond age and functional independence.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Exclusion criteria:\u003c/p\u003e\n \u003cp\u003e\u0026ndash; For PwPD: Significant cognitive impairment.\u003c/p\u003e\n \u003cp\u003e\u0026ndash; For young adults: Any functional impairments.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Total participants: Video dataset included frames from 7 young adults and 3 PwPD; focus group size was n\u0026thinsp;=\u0026thinsp;4.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePanyakaew et al. (2021)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eThailand\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCross sectional study\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTo explore the prediction of falling in Parkinson\u0026apos;s disease patients using a machine learning-based approach.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRecruited from the Chulalongkorn Center of Excellence for Parkinson\u0026rsquo;s Disease and Related Disorders, Faculty of Medicine, Chulalongkorn University, Thailand, between January and December 2019.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026bull; Recruitment: Patients with a clinical diagnosis of PD with Hoehn \u0026amp; Yahr (H\u0026amp;Y) stage 1\u0026ndash;4, with or without a history of falls, were recruited from the same center (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ewww.chulapd.org\u003c/span\u003e\u003c/span\u003e) between January and December 2019.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Inclusion criteria: (1) Clinical diagnosis of PD based on standard diagnostic criteria; (2) Hoehn \u0026amp; Yahr (H\u0026amp;Y) stage 1\u0026ndash;4; (3) able to ambulate within community residences; (4) able to follow simple commands with a score\u0026thinsp;\u0026ge;\u0026thinsp;21 on the Thai version of the Mini-Mental Status Examination (MMSE); and (5) able to complete the ABC-16 scale independently.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Exclusion criteria: (1) Hoehn \u0026amp; Yahr stage 5; (2) coexisting disorders contributing to postural instability and falls, including stroke, ataxia, neuropathy, and visual, vestibular, or proprioceptive problems; (3) patients taking sedative medications.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Total participants: 305.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eP\u0026eacute;rez-Ros et al. (2019)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSpain\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOther: Cohort trial nested case-control design\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTo determine the predictive factors for falls in functionally independent community dwelling older adults.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eData was collected from a prospective, longitudinal study of 374 community-dwelling older adults aged\u0026thinsp;\u0026ge;\u0026thinsp;70 years in La Ribera, Valencia, Spain, recruited from primary care centers between December 2013 and May 2014, with a 12-month follow-up to assess falls.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026bull; Recruitment: A publicity strategy was used from December 2013 to May 2014, including posters and notifications in retirement centers, meetings with managing authorities of retirement associations and centers, and telephone calls to individuals enrolled in primary care centers. Family telephone calls were also used to reduce dropouts among individuals lacking the physical means to attend fall-prevention education sessions.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Inclusion criteria: (1) Aged\u0026thinsp;\u0026ge;\u0026thinsp;70 years; (2) Barthel Index\u0026thinsp;\u0026ge;\u0026thinsp;60 (functionally independent); (3) Independent walking (with or without technical aids, but not assisted by another person); (4) Resident in the La Ribera region.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Exclusion criteria: (1) Life expectancy\u0026thinsp;\u0026lt;\u0026thinsp;6 months; (2) Blindness or deafness; (3) Serious psychiatric illness (schizophrenia, major depression, bipolar disorder, panic disorder); (4) Moderate to severe cognitive impairment (MEC-Lobo\u0026thinsp;\u0026lt;\u0026thinsp;24); (5) Did not sign informed consent.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Total participants: 374.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\n \u003ch2\u003eAI-based Fall Risk Prediction Models\u003c/h2\u003e\n \u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e summarizes the characteristics of AI-based fall risk prediction models included in this review, including the type of AI algorithms used, environmental and non-environmental risk factors incorporated, reported model performance metrics, and key study limitations.\u003c/p\u003e\n \u003cdiv class=\"gridtable\"\u003e\u0026nbsp;\u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eAI-Based Fall Risk Models and Environmental Risk Factors\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"6\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAuthor (Year of publication)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eType of AI algorithm\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eEnvironmental Risk Factors\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eNon-Environmental Risk Factors\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eModel Performance\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eLimitation\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eChen et al. (2023)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMachine learning\u003c/p\u003e\n \u003cp\u003e\u0026bull; AI algorithm: logistic regression (LR); support vector machine (SVM); random forest (RF); adaptive boosting (AdaBoost); light gradient boosting machine (LightGBM)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStructure of building; handicapped facilities; kitchen availability; flush toilets; cooking fuel type; Internet availability; house tidiness; house temperature\u003c/p\u003e\n \u003cp\u003e\u0026bull; Measures of environmental risks: inquiry-based survey measures\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBiological, behavioral, and socioeconomic predictors including sex, vision, diabetes, liver disease, memory-related disease, disabilities (brain, vision, physical), ADL, IADL, experience of falling, hip fracture history, smoking, alcohol use, sleep duration, medications (antihypertensive, lung, heart, kidney, memory, arthritis), depressive symptoms, dental care, relationship with children, health satisfaction, systolic BP, walking speed, hand strength, white blood cell count, blood urea nitrogen, glucose; injury model includes marital status, stroke, dyslipidemia medication, digestive medication, income, life satisfaction, lung function, abdominal obesity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eThe LR model showed good performance with an AUC-ROC of 0.739 (sensitivity 0.707) for falls and an AUC-ROC of 0.757\u003c/p\u003e\n \u003cp\u003e(sensitivity 0.654) for fall-related injuries.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Fall risk impact: Baseline falling experience was the most important predictor. A history of falls was the strongest predictor of future falls, and house tidiness was an important environmental predictor in both fall and fall-related injury models.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVariables were treated as categorical; causal inference was not possible; predictors were only measured at baseline, without accounting for changes over time.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCloix et al. (2016)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eComputer vision for stairs and curbs detection\u003c/p\u003e\n \u003cp\u003e\u0026bull; AI algorithm: Binary prediction algorithm (Algorithm 1) using upper/lower depth map (Dp) from stereo images; Three-class prediction algorithm (Algorithm 2) using full depth map (D) divided into sub-depth maps; thresholds TG/TR and TGi/TRi applied to ground distance and ratio of valid pixels\u003c/p\u003e\n \u003cp\u003e\u0026bull; Environmental risks identified:\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStairs and curbs\u003c/p\u003e\n \u003cp\u003e\u0026bull; Measures of environmental risks: depth difference from stereo disparity; ratio of pixels below ground level; spatial location of drop (upper vs. lower half of image); predicted \u0026ldquo;danger zone\u0026rdquo; proximity using trapezoid projection\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNot applicable\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eThe system distinguished more than 94% dangerous scenes from safe scenes with an overall recognition rate of 91% at very low resolution, operating in real time and robust to indoor/outdoor lighting conditions.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Fall risk impact: Not applicable.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHDR cameras can better capture scenes under bright sunlight to avoid saturation that causes stereo image matching to fail, and DSPs may assist with motion blur removal before disparity computation to improve the SAD stereo image matching.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDu et al. (2014)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eComputer vision, robotics, and shared control systems\u003c/p\u003e\n \u003cp\u003e\u0026bull; AI algorithm: Gmapping (Rao-Blackwellized Particle Filter SLAM); OctoMap (3D mapping using octrees); costmap-based obstacle avoidance planning; AMCL (Adaptive Monte Carlo Localization)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePoor lighting; cluttered or narrow spaces; obstructed walkways; limited visibility (telepresence)\u003c/p\u003e\n \u003cp\u003e\u0026bull; Measures of environmental risks: luminosity (lux) thresholds: \u0026ge;401\u0026thinsp;=\u0026thinsp;bright, 201\u0026ndash;400\u0026thinsp;=\u0026thinsp;medium, \u0026lt;\u0026thinsp;200\u0026thinsp;=\u0026thinsp;dark; lighting condition displayed on web interface (green\u0026thinsp;=\u0026thinsp;bright, yellow\u0026thinsp;=\u0026thinsp;medium, red\u0026thinsp;=\u0026thinsp;dark)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNot applicable\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026bull; Model performance: Not reported.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Fall risk impact: Not applicable.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eThe system is a prototype and requires major improvements, including automation of assessment tasks, addition of object recognition functions for independent hazard detection, and better error handling and recovery methods to ensure reliability and stability for use in real world.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLathouwer et al. (2022)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMachine learning\u003c/p\u003e\n \u003cp\u003e\u0026bull; AI algorithm: Random Forest model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHousing issues, housing change, environmental vulnerability\u003c/p\u003e\n \u003cp\u003e\u0026bull; Measures of environmental risks: not reported\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e24 variables including biological (loneliness, sex), behavioral (physical vulnerability, physical effort, physical activity, mental activity, help required, help available, mode of transportation, psychological vulnerability), and socioeconomic factors (age class, education, civil status, surrounding density, homeownership, home type, number of children and grandchildren, insecurity, neighborhood organization, social vulnerability)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eThe model reached an average of 73% accuracy.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Fall risk impact: The classification model identified 24 predictors of falling (2 biological, 8 behavioral, 11 socioeconomic, and 3 environmental factors), each contributing 4.5\u0026ndash;6.5% to the overall fall risk.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFurther work is needed to improve data quality, address multicollinearity, explore additional ML methods (e.g., SVM, neural networks), and enhance accuracy through hyperparameter tuning and larger datasets.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLo et al. (2019)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMachine learning\u003c/p\u003e\n \u003cp\u003e\u0026bull; AI algorithm: Random Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eClutter or obstacles; inadequate lighting; unsafe flooring; lack of assistive devices or improper use\u003c/p\u003e\n \u003cp\u003e\u0026bull; Measures of environmental risks: assessed via OASIS-C structured items recorded by home health care clinicians during mandatory assessments\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eage; sex; language group; borough of residence; number and severity of diagnoses; pain levels; cognitive impairment; visual impairment; frequency of assistance with ADL/IADL; therapy visit frequencies; living situation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eThe OASIS model achieved a balanced accuracy of 0.62, an AUC-ROC of 0.67 (95% CI: 0.66\u0026ndash;0.68), and an average precision of 0.10, outperforming the baseline MAHC-10 scoring system.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Fall risk impact: A random forest model using OASIS-C and EHR data provided improved fall risk prediction compared with the MAHC-10 tool, with age, severity of diagnoses, therapy visits, pain, and ADL/IADL assistance identified as top predictors supporting individualized fall-prevention strategies.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eThe study relied on data from a single home care agency, and model precision remained low due to highly imbalanced outcomes; nearly 95% of patients in the cohort had no reported falls.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMillet et al. (2023)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMachine learning\u0026thinsp;+\u0026thinsp;Natural language processing\u003c/p\u003e\n \u003cp\u003e\u0026bull; AI algorithms: Random Forest; Decision Tree; Logistic Regression; LightGBM; Support Vector Machine (SVM); K-Nearest Neighbors (KNN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStairs at home; teleassistance; home accessibility; elevator; stairs management (from selected 25 features)\u003c/p\u003e\n \u003cp\u003e\u0026bull; Measures of environmental risks: Not clarified\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBMI; weight; age; gait speed; height; standing balance; walking cane; grip strength; sitting balance; foot support; tele-assistance; lift; antiplatelet drugs; cognitive impairment; anticoagulants; hip fracture; dizziness; depressive syndrome; fear; memory loss; stable turns; disorientation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBest performance achieved by the Bagging ensemble using Random Forest, which attained 75.8% accuracy, 70.0% sensitivity, 80.5% specificity, and an AUC of 75.3%. Among base models, Random Forest alone performed strongly with 76.9% accuracy and 85.0% specificity, though with lower sensitivity (65.5%). Ensemble methods generally improved robustness and balance across metrics.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Fall risk impact:\u003c/p\u003e\n \u003cp\u003eAccurate prediction of recurrent falls in older adults using routine clinical data.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eReliance on retrospective data from a single hospital, the semi-structured nature of clinical notes requiring extensive NLP preprocessing, relatively low model sensitivity, and the absence of dynamic risk updates or validation in prospective or interventional settings.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMoore et al. (2024)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDeep learning model\u003c/p\u003e\n \u003cp\u003e\u0026bull; AI algorithm: YOLOv8 object detection algorithm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStairs; doorway; shower; sink; toilet; table; bed; signage; chair; animal; wet surface; mat/rug/carpet; obstacle (generic catch-all for potential obstructions); raised curbs\u003c/p\u003e\n \u003cp\u003e\u0026bull; Measures of environmental risks: Not applicable\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVehicle\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUsing the collected dataset, the YOLOv8 algorithm was trained for a course of 100 epochs converging at epoch 69 within a timeframe of 4 h. The models achieved a best validation mAP50 of 0.81 at epoch 69, showcasing the potential of this algorithm within real-world deployment.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Fall risk impact: NA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOnly a small number of PwPD were recruited, and a limited dataset was curated as part of the pilot study.\u003c/p\u003e\n \u003cp\u003eThe number of participants for the focus group was small too and they were recruited based on their prior experience of participating in wearable gait research and their familiarity with technology.\u003c/p\u003e\n \u003cp\u003eThe focus group participants may have been biased toward the acceptance of technology as they were recruited by purposive sampling to have a good understanding and/or appreciation of commercial technology.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePanyakaew et al. (2021)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026bull; Type of AI: Machine learning\u003c/p\u003e\n \u003cp\u003e\u0026bull; AI algorithm: XGBoost models for predicting (1) falls (fallers vs. non-fallers) and (2) recurrent falls (recurrent vs. non-recurrent fallers); SHAP used for model interpretability\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHigh-risk daily activities including sweeping the floor (ABC-16 item 7); reaching on tiptoes (item 5); walking in a crowded mall (item 12); walking across a parking lot (item 10); getting in/out of a car (item 9); walking up/down stairs (item 2); walking up/down a ramp (item 11)\u003c/p\u003e\n \u003cp\u003e\u0026bull; Measures of environmental risks: Activities-Specific Balance Confidence Scale (ABC-16)\u0026mdash;self-reported confidence (0\u0026ndash;100%) in 16 daily activities\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eH\u0026amp;Y stage 3, disease duration, age, PD subtype, Postural Instability and Gait Disorder (PIGD), wearing-off, dyskinesia, use of dopamine agonists, use of COMT inhibitors\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFalls model accuracy: 72% (p\u0026thinsp;=\u0026thinsp;0.001)\u003c/p\u003e\n \u003cp\u003eRecurrent falls model accuracy: 81% (p\u0026thinsp;=\u0026thinsp;0.02)\u003c/p\u003e\n \u003cp\u003e\u0026bull; Fall risk impact:\u003c/p\u003e\n \u003cp\u003eThe study successfully predicted fallers (72%) and recurrent fallers (81%).\u003c/p\u003e\n \u003cp\u003eTop predictive environmental activities included: Sweeping the floor, Reaching on tiptoes, Walking in a crowded mall, Walking across a parking lot, and Getting in/out of a car.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eExclusion of neurological co-morbidities; lack of details on anti-parkinsonian medication use; PD patients had relatively normal cognition with no neuropsychiatric/non-motor details.\u003c/p\u003e\n \u003cp\u003eOnly total MMSE score was available, with no breakdown of cognitive domains or executive function.\u003c/p\u003e\n \u003cp\u003eRetrospective fall reporting; exclusion of patients with MMSE\u0026thinsp;\u0026lt;\u0026thinsp;21; no objective balance measures; medication details lacking.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eP\u0026eacute;rez-Ros et al. (2019)\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMachine learning\u003c/p\u003e\n \u003cp\u003e\u0026bull; AI algorithm: Binary logistic regression models\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLack of stair handrails; poor stair design; lack of bathroom grab bars; dim lighting or glare; obstacles and tripping hazards; slippery or uneven surfaces; improper use of assistive devices\u003c/p\u003e\n \u003cp\u003e\u0026bull; Measures of environmental risks: recorded as presence/absence of specific hazards via nurse-administered geriatric assessment at baseline\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAdvanced age (\u0026ge;\u0026thinsp;80 years); previous falls; muscle weakness; gait and balance problems; poor vision; postural hypotension; chronic conditions (osteoarthritis, diabetes, stroke, Parkinson\u0026rsquo;s disease, incontinence, dementia); fear of falling (mFES); previous fractures; obesity (BMI\u0026thinsp;\u0026ge;\u0026thinsp;30 kg/m\u0026sup2;); hearing problems; anxiety-depressive syndrome; medication use (alpha-blockers, benzodiazepines, beta-blockers, CNS-acting drugs); polypharmacy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eIsolated falls model: Sensitivity 10%, Specificity 98.2%, PPV 75.2%, NPV 77.5%, Nagelkerke R\u0026sup2; = 47%, Overall correct classification\u0026thinsp;=\u0026thinsp;77%.\u003c/p\u003e\n \u003cp\u003eRecurrent falls model: Sensitivity 15.2%, Specificity 98.8%, PPV 58.8%, NPV 92.3%, Nagelkerke R\u0026sup2; = 69.4%, Overall correct classification\u0026thinsp;=\u0026thinsp;91.4%.\u003c/p\u003e\n \u003cp\u003e\u0026bull; Fall risk impact:\u003c/p\u003e\n \u003cp\u003ePredictive models identified that prior falls and use of alpha-blockers predicted isolated falls, while previous fractures, obesity (BMI\u0026thinsp;\u0026ge;\u0026thinsp;30 kg/m\u0026sup2;), and use of benzodiazepines and beta-blockers predicted recurrent falls in independent older adults. These findings highlighted modifiable risk factors that may inform fall-prevention strategies.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDrug-related data reflected only commonly prescribed pharmacotherapeutic groups; drug doses were not documented. Time, weather, and location of falls were not recorded. Follow-up was limited to 12 months. Self-reported monthly falls may be biased due to missing information. Medication adherence was not assessed.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003eSix studies (66.7%) applied supervised machine learning models to structured clinical, functional, or survey data. Commonly used algorithms included logistic regression, random forest, and gradient boosting\u0026ndash;based ensemble methods such as XGBoost, AdaBoost, and LightGBM. Several studies evaluated multiple classifiers with ensemble approaches. One study additionally incorporated natural language processing to extract features from narrative clinical text before model training.\u003c/p\u003e\n \u003cp\u003eThree studies focused on vision- or robotics-based approaches to environmental perception rather than direct fall outcome prediction. These studies employed techniques such as simultaneous localization and mapping (SLAM), depth-based detection pipelines, and deep learning object detection models (e.g., YOLOv8) to identify environmental features and hazards in home-like settings.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\n \u003ch2\u003e(1) Incorporation of Environmental Risk Factors\u003c/h2\u003e\n \u003cp\u003eThe included studies identified a wide range of environmental hazards at home that contribute to fall risk, though methods and comprehensiveness of environmental assessments varied considerably.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\n \u003ch2\u003ea. Structural and Housing Characteristics\u003c/h2\u003e\n \u003cp\u003eLathouwers et al.\u003csup\u003e23\u003c/sup\u003e identified housing issues, housing change (i.e., homeownership or home type), and environmental vulnerability as significant predictors for fall occurrence, though specific measurement methods were not reported. Chen et al.\u003csup\u003e27\u003c/sup\u003e incorporated multiple housing characteristics into both fall prediction and fall-related injury prediction models through structured inquiry, including building structure (reinforced concrete versus other materials), presence of handicapped facilities, kitchen and flush toilet availability, Internet access, house tidiness (clear versus unclear), and house temperature (hot, neutral, cold). Millet et al.\u003csup\u003e29\u003c/sup\u003e assessed home-related features, including the presence of stairs, home accessibility, and elevator availability.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\n \u003ch2\u003eb. Specific Environmental Risk Factors\u003c/h2\u003e\n \u003cp\u003ePerez-Ros et al.\u003csup\u003e25\u003c/sup\u003e documented specific hazards through nurse-administered geriatric assessment, including lack of stair handrails, poor stair design, lack of bathroom grab bars, dim lighting or glare, obstacles and tripping hazards, slippery or uneven surfaces, and improper use of assistive devices. Lo et al.\u003csup\u003e21\u003c/sup\u003e assessed environmental safety issues via structured assessment items completed by home health care clinicians, including clutter or obstacles, inadequate lighting, unsafe flooring, and lack of assistive devices or improper use. Du et al.\u003csup\u003e22\u003c/sup\u003e focused on lighting-related hazards, including poor lighting, cluttered or narrow spaces, obstructed walkways, and limited visibility, and established quantitative luminosity thresholds that were displayed via a color-coded interface. Moore et al.\u003csup\u003e26\u003c/sup\u003e identified 14 distinct environmental hazard categories using computer vision: stairs, doorways, showers, sinks, toilets, tables, beds, signage, chairs, animals, wet surfaces, mats/rugs/carpets, generic obstacles, and raised curbs. Panyakaew et al.\u003csup\u003e28\u003c/sup\u003e integrated environmental components into validated mobility and balance scales, including sweeping the floor, reaching on tiptoes, walking in a crowded mall, and walking up/down stairs and ramps, using activities that directly reflect environmental challenges.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\n \u003ch2\u003e(2) Integration with Non-Environmental Risk Factors\u003c/h2\u003e\n \u003cp\u003eAcross studies, commonly integrated domains included demographic characteristics, clinical conditions, functional status, cognitive and psychological factors, medication use, and social or socioeconomic variables.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e23\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e25\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e Several models combined fall risk factors including age, prior falls, gait and balance impairments, chronic diseases, sensory deficits, and polypharmacy with psychosocial variables (e.g., depressive symptoms, fear of falling, social vulnerability) and functional dependence measures (Activities of Daily Living [ADL]/ Instrumental Activities of Daily Living [IADL]).\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e Some studies further incorporated disease-specific factors (e.g., Parkinson\u0026rsquo;s disease severity and motor complications) or physiological and performance-based measures (e.g., gait speed, grip strength, laboratory values).\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e26\u003c/span\u003e,\u003cspan class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\n \u003ch2\u003eAI Model Performance\u003c/h2\u003e\n \u003cp\u003eAmong studies using supervised machine learning on structured data, discrimination was generally modest to good. Reported AUC-ROC values ranged from approximately 0.67 to 0.76, with Chen et al.\u003csup\u003e27\u003c/sup\u003e achieving the highest performance using logistic regression (AUC-ROC 0.739 for falls and 0.757 for fall-related injuries). Random forest\u0026ndash;based models frequently outperformed traditional scoring tools, as demonstrated by Lo et al.\u003csup\u003e21\u003c/sup\u003e (0.62 vs. lower baseline performance), although precision remained limited due to highly imbalanced outcomes. Perez-Ros et al.\u003csup\u003e25\u003c/sup\u003e reported specificity exceeding 98% but sensitivity below 20%, indicating strong ability to rule in high-risk individuals but limited capacity to identify all fallers. Ensemble and gradient boosting approaches (e.g., XGBoost, LightGBM, bagging random forest) generally improved robustness and balance across performance metrics, with Millet et al.\u003csup\u003e29\u003c/sup\u003e reporting the best overall performance for recurrent fall prediction (AUC\u0026thinsp;\u0026asymp;\u0026thinsp;0.75; accuracy\u0026thinsp;\u0026asymp;\u0026thinsp;76%).\u003c/p\u003e\n \u003cp\u003eStudies employing computer vision or robotics-based approaches primarily evaluated algorithmic feasibility rather than clinical prediction performance. Cloix et al.\u003csup\u003e24\u003c/sup\u003e demonstrated high environmental hazard recognition accuracy (\u0026gt;\u0026thinsp;90%) for stair and curb detection, while Moore et al.\u003csup\u003e26\u003c/sup\u003e reported strong object-detection performance using YOLOv8 (mAP50\u0026thinsp;=\u0026thinsp;0.81).\u003c/p\u003e\n \u003cp\u003eEnvironmental factors consistently contributed to fall risk prediction when incorporated into AI-based models, although their representation and relative importance varied substantially by data source and modeling approach. Chen et al.\u003csup\u003e27\u003c/sup\u003e demonstrated that house tidiness emerged as one of the most important environmental predictors for both falls and fall-related injuries, alongside prior fall history. Similarly, Lo et al.\u003csup\u003e21\u003c/sup\u003e found that environmental hazards documented in OASIS-C assessments, such as inadequate lighting and household obstacles, contributed to improved discrimination compared with a standard clinical screening tool. Lathouwers et al.\u003csup\u003e23\u003c/sup\u003e further demonstrated that environmental vulnerability indicators contributed alongside biological, behavioral, and socioeconomic factors, with each environmental predictor accounting for a comparable proportion of overall fall risk.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis systematic review provides a focused synthesis of AI-based fall risk prediction models that incorporate environmental factors among community-dwelling older adults, addressing a critical gap in the existing fall prevention literature. One observation from this study emerged during the search and screening stage. Despite identifying over 20,000 records through a comprehensive search, most studies were excluded because environmental factors were not incorporated into AI-based fall prediction models. Among the included studies, while personal, clinical, and functional factors appropriately formed the foundation of model development, environmental factors were typically included as part of broader predictor sets rather than being explicitly examined as distinct contributors to fall risk. These findings highlight a critical gap in AI-based fall prediction research and reveal environmental context as an underdeveloped yet potentially high-impact component.\u003c/p\u003e\n\u003cp\u003eNotably, synthesis across diverse AI applications demonstrated that when environmental features were incorporated, they consistently contributed meaningful information to fall risk prediction,\u003csup\u003e21,23,27\u003c/sup\u003e either by improving model discrimination or by identifying actionable home hazards. These findings suggested that environmental factors are not peripheral but provide complementary and actionable information that enhances fall risk prediction. An important implication of these findings is that environmental factors differ from many personal or clinical predictors in that they are inherently modifiable.\u003csup\u003e30-32\u003c/sup\u003e While variables such as age, chronic conditions, or prior falls primarily support risk stratification,\u003csup\u003e33,34\u003c/sup\u003e environmental hazards can directly inform targeted prevention strategies. Lack of attention to the environmental context in current AI-based models may constrain both predictive insight and the translation of model outputs into actionable prevention strategies. Therefore, further study is warranted to explicitly incorporate environmental risk factors, which could strengthen the link between fall prediction and practical interventions in community settings.\u003c/p\u003e\n\u003cp\u003eFall hazards can be identified through traditional home environmental assessments\u0026nbsp;conducted by healthcare providers during home visits, which are considered as a gold-standard approach.\u003csup\u003e35-37\u003c/sup\u003e However, their implementation is often constrained by substantial practical limitations. For example, comprehensive validated tools such as the \u003cem\u003eWestmead Home Safety Assessment\u0026nbsp;\u003c/em\u003econsist ofa 72-item checklist,\u003csup\u003e38\u003c/sup\u003e require considerable time and clinical effort and may yield variable results depending on provider expertise and perspective.\u003csup\u003e39\u003c/sup\u003e Additionally, access is limited for older adults in remote areas due to shortages of trained personnel, travel barriers, and economic constraints,\u003csup\u003e31,40\u003c/sup\u003e In parallel, this review identified substantial heterogeneity in how environmental risk factors are represented in AI-based models, ranging from simple binary indicators (e.g., presence of handrails) to high-dimensional sensor- or vision-derived data. Such inconsistency limits model comparability, replication, and cumulative knowledge building, ultimately constraining translation into practice.\u003csup\u003e41\u003c/sup\u003e Therefore, the need for AI-based approaches is highlighted as capable of supporting standardized, scalable, and objective environmental risk assessment. In this context, AI-driven methods, such as computer vision, image-based analysis, and multimodal data integration, offer promising opportunities to reduce assessment burden while enabling more standardized, precise, and usable representations of environmental risk information within fall prevention workflows. However, for real-word adoption in routine clinical practice, existing vision- and robotics have largely focused on algorithmic feasibility, have not yet demonstrated clinical performance, and have provided limited participant demographic information. Consequently, further validation across diverse clinical scenarios is necessary before clinical implementation.\u003c/p\u003e\n\u003cp\u003eFrom an aging and community health perspective, the limited integration of environmental context in AI-based fall prediction models has important implications for equity, real-world relevance, and healthy aging. As people age, their immediate living environments play an increasingly central role in daily functioning,\u003csup\u003e42,43\u003c/sup\u003e making housing conditions and neighborhood infrastructure key determinants of fall risk.\u003csup\u003e44-46\u003c/sup\u003e These environmental conditions vary widely across populations and are closely linked to lifelong socioeconomic circumstances, geographic location, and access to supportive resources.\u003csup\u003e44-46\u003c/sup\u003e Consequently, AI models that insufficiently account for environmental risk may systematically overlook or underestimate key sources of vulnerability among community-dwelling older adults, particularly those aging in place within resource-limited settings. Such omissions risk reinforcing cumulative disadvantages in later life by limiting accurate risk identification and delaying preventive action, even when overall model performance appears acceptable. Incorporating environmental context into AI-based fall prediction models can therefore move these tools beyond risk stratification toward more context-aware and actionable insights, enabling earlier identification of modifiable hazards and supporting targeted interventions that promote safer aging in place and more equitable fall prevention across diverse aging populations.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;This systematic review has several limitations. First, although a comprehensive search strategy was employed across multiple databases, relevant studies may have been missed due to publication bias, indexing limitations, or exclusion criteria such as restriction to English-language publications and peer-reviewed articles, potentially excluding gray literature. Second, the small number of eligible studies reflects the emerging nature of AI-based fall prediction models that explicitly incorporate environmental factors, limiting the ability to draw definitive conclusions or conduct quantitative synthesis. Third, fall-related outcomes varied across studies, including any falls, recurrent falls, and fall-related injuries, with varying follow-up durations. This variability complicates comparison across models and may influence reported performance metrics. Fourth, several studies did not report external validation or relied on single datasets, raising concerns about model generalizability. The extent to which models incorporating environmental factors perform consistently across different populations and settings remains uncertain. Fifth, most included studies were conducted in high-income countries, potentially limiting the applicability of findings to low- and middle-income settings or to culturally diverse housing environments where environmental risks and living conditions may differ substantially. Finally, while some studies identified environmental hazards or demonstrated improved predictive performance, few evaluated whether incorporating environmental factors altered fall prevention interventions or reduced fall incidence, limiting conclusions about real-world effectiveness.\u003c/p\u003e\n\u003cp\u003eDespite several limitations, this study has a number of important strengths. First, this study is the first systematic review to specifically examine how environmental factors are incorporated into AI-based fall risk prediction models. In doing so, we address a critical gap in the existing literature, which has largely focused on individual- and functional-level predictors. By isolating the environmental dimension, this review advances understanding of how modifiable home-related risks are currently conceptualized and operationalized within AI-based approaches. Second, this study provides a broad and integrative perspective on environmental risk representation by synthesizing evidence across diverse AI paradigms ranging from traditional machine learning models using structured data to computer vision and robotics-based approaches. Finally, this review provides clear evidence that environmental factors contribute meaningful information to AI-based fall risk prediction models.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003eFuture Implication\u003c/h2\u003e\n\u003cp\u003eFuture research on AI-based fall prediction should address several important gaps identified in this review.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e(1)\u0026nbsp;\u0026nbsp;There is a clear need for standardized definitions and measurement frameworks for environmental risk factors, as current studies vary widely in how environmental hazards are conceptualized, assessed, and represented in models. Without greater consistency, comparison across studies and translation into practice remain limited.\u003c/p\u003e\n\u003cp\u003e(2)\u0026nbsp;\u0026nbsp;Multimodal AI models that integrate clinical, functional, and environmental data are needed to more fully capture the multifactorial nature of fall risk. Falls often occur as the result of dynamic interactions among individual characteristics, functional limitations, and environmental hazards, rather than from any single factor alone. While many existing models emphasize individual- and functional-level predictors, incorporating environmental information alongside these factors may improve predictive performance and clinical relevance. Given that many environmental hazards are inherently visual and context-dependent, image-based or computer vision–driven approaches may be particularly valuable for identifying home environmental risks that are difficult to capture through structured questionnaires or clinical assessments alone. Advances in image recognition, object detection, and depth sensing offer promising opportunities to enhance environmental risk detection in real-world home settings, ultimately informing targeted interventions to reduce fall risk.\u003c/p\u003e\n\u003cp\u003e(3)\u0026nbsp;\u0026nbsp;Environmental characteristics may vary significantly across different geographic regions and cultural contexts. Future research could further explore adaptive modeling approaches for environmental factors in multicultural settings, aiming to develop models that generalize better across regions. In addition, investigating potential fall risks arising from the interaction between cultural behaviors and environmental contexts may contribute to more inclusive, culturally sensitive, and accurate fall risk prediction.\u003c/p\u003e\n\u003cp\u003e(4)\u0026nbsp;\u0026nbsp;Environmental features were generally assessed at a single time point and treated as static predictors. This approach does not reflect the dynamic nature of home environments, where hazards may change over time due to health decline, behavior adaptation, seasonal conditions, or home modifications. Future AI models may benefit from longitudinal or continuously updated environmental data to better reflect real-world fall risk trajectories.\u003c/p\u003e\n\u003cp\u003e(5) \u0026nbsp;Future AI-based fall prediction tools should prioritize scalability, usability, and real-world applicability, ensuring that models can be feasibly deployed in community settings and adapted to diverse home environments. Approaches that leverage mobile devices, smart home technologies, or automated image-based assessments may help bridge the gap between predictive modeling and actionable fall prevention strategies.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eThis review shows that environmental factors remain infrequently emphasized in AI-based fall prediction models, despite their established relevance to fall risk among community-dwelling older adults. Environmental context provides complementary and actionable information that supports the need for more standardized, context-aware AI approaches to equitable and effective fall prevention.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e[Financial support]\u003c/p\u003e\n\u003cp\u003eThis study is funded by the Boston College Connell School of Nursing Intramural Pilot Grant (SAFE-AI Phase I (Synergistic Approach to Fall-risk Evaluation in older adults using generative AI). PI: Jung). Jiyoun Song is funded by the NHLBI (1K99HL169940). The content is solely the responsibility of the authors and does not necessarily represent the official views of the NHLBI. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e[Conflict of interest]\u003c/p\u003e\n\u003cp\u003eAll authors report no conflicts of interest relevant to this article.\u003c/p\u003e\n\u003cp\u003e[Ethical Conduct of Research]\u003c/p\u003e\n\u003cp\u003eThis systematic review utilized only publicly available published data and did not involve human subjects. Therefore, institutional review board (IRB) approval was not required.\u003c/p\u003e\n\u003cp\u003e[Declaration of Generative AI and AI-assisted Technologies in the Writing Process]\u003c/p\u003e\n\u003cp\u003eDuring the preparation of this work, the authors used WordTune and Grammarly for editorial support. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.\u003c/p\u003e\n\u003cp\u003e[Clinical Trial Number]\u003c/p\u003e\n\u003cp\u003eClinical trial number: not applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eCenter for Disease Control and Prevention. National Center for Injury Prevention and Control: Web\u0026ndash;based Injury Statistics Query and Reporting System (WISQARS). Accessed Nov 20th, 2025. https://wisqars.cdc.gov/\u003c/li\u003e\n\u003cli\u003eThompson HJ, McCormick WC, Kagan SH. Traumatic brain injury in older adults: epidemiology, outcomes, and future implications. \u003cem\u003eJournal of the American Geriatrics Society\u003c/em\u003e. Oct 2006;54(10):1590\u0026ndash;5. doi:10.1111/j.1532-5415.2006.00894.x\u003c/li\u003e\n\u003cli\u003eSterling DA, O\u0026rsquo;Connor JA, Bonadies J. Geriatric Falls: Injury Severity Is High and Disproportionate to Mechanism. \u003cem\u003eJournal of Trauma and Acute Care Surgery\u003c/em\u003e. 2001;50(1)\u003c/li\u003e\n\u003cli\u003eVaishya R, Vaish A. Falls in Older Adults are Serious. \u003cem\u003eIndian Journal of Orthopaedics\u003c/em\u003e. 2020/02/01 2020;54(1):69\u0026ndash;74. doi:10.1007/s43465-019-00037-x\u003c/li\u003e\n\u003cli\u003eMontero-Odasso M, van der Velde N, Martin FC, et al. World guidelines for falls prevention and management for older adults: a global initiative. \u003cem\u003eAge Ageing\u003c/em\u003e. Sep 2 2022;51(9)doi:10.1093/ageing/afac205\u003c/li\u003e\n\u003cli\u003eJahandideh S, Hutchinson AF, Bucknall TK, et al. Using machine learning models to predict falls in hospitalised adults. \u003cem\u003eInternational journal of medical informatics\u003c/em\u003e. 2024/07/01/ 2024;187:105436. doi:https://doi.org/10.1016/j.ijmedinf.2024.105436\u003c/li\u003e\n\u003cli\u003eUsmani S, Saboor A, Haris M, Khan MA, Park H. Latest Research Trends in Fall Detection and Prevention Using Machine Learning: A Systematic Review. \u003cem\u003eSensors\u003c/em\u003e. 2021;21(15):5134. doi:10.3390/s21155134 \u003c/li\u003e\n\u003cli\u003eBargiotas I, Wang D, Mantilla J, et al. Preventing falls: the use of machine learning for the prediction of future falls in individuals without history of fall. \u003cem\u003eJournal of Neurology\u003c/em\u003e. 2023/02/01 2023;270(2):618\u0026ndash;631. doi:10.1007/s00415-022-11251-3\u003c/li\u003e\n\u003cli\u003eDubois A, Bihl T, Bresciani J-P. Identifying Fall Risk Predictors by Monitoring Daily Activities at Home Using a Depth Sensor Coupled to Machine Learning Algorithms. \u003cem\u003eSensors\u003c/em\u003e. 2021;21(6):1957. doi:10.3390/s21061957 \u003c/li\u003e\n\u003cli\u003eLima CA, Ricci NA, Nogueira EC, Perracini MR. The Berg Balance Scale as a clinical screening tool to predict fall risk in older adults: a systematic review. \u003cem\u003ePhysiotherapy\u003c/em\u003e. 2018/12/01/ 2018;104(4):383\u0026ndash;394. doi:https://doi.org/10.1016/j.physio.2018.02.002\u003c/li\u003e\n\u003cli\u003eStrini V, Schiavolin R, Prendin A. Fall Risk Assessment Scales: A Systematic Literature Review. \u003cem\u003eNurs Rep\u003c/em\u003e. Jun 2 2021;11(2):430\u0026ndash;443. doi:10.3390/nursrep11020041\u003c/li\u003e\n\u003cli\u003eGonz\u0026aacute;lez-Castro A, Leir\u0026oacute;s-Rodr\u0026iacute;guez R, Prada-Garc\u0026iacute;a C, Ben\u0026iacute;tez-Andrades JA. The Applications of Artificial Intelligence for Assessing Fall Risk: Systematic Review. \u003cem\u003eJournal of medical Internet research\u003c/em\u003e. 2024/4/29 2024;26:e54934. doi:10.2196/54934\u003c/li\u003e\n\u003cli\u003eLiu S, Si Y, Peng Y, Wang D, Yuan X, Peng Y. Incidence and risk factors of falls in older people with chronic comorbidities in community: a cross-sectional study. Original Research. \u003cem\u003eFrontiers in Public Health\u003c/em\u003e. 2025\u0026ndash;September\u0026ndash;24 2025;Volume 13 - 2025doi:10.3389/fpubh.2025.1643699\u003c/li\u003e\n\u003cli\u003eLi Y, Hou L, Zhao H, Xie R, Yi Y, Ding X. Risk factors for falls among community-dwelling older adults: A systematic review and meta-analysis. \u003cem\u003eFront Med (Lausanne)\u003c/em\u003e. 2022;9:1019094. doi:10.3389/fmed.2022.1019094\u003c/li\u003e\n\u003cli\u003eSaunders S, D\u0026apos;Amore C, Hao Q, et al. Risk Factors for Falls in Community-Dwelling Older Adults: An Umbrella Review. \u003cem\u003eJournal of the American Medical Directors Association\u003c/em\u003e. 2025;26(9)doi:10.1016/j.jamda.2025.105765\u003c/li\u003e\n\u003cli\u003eStevens JA, Mahoney JE, Ehrenreich H. Circumstances and outcomes of falls among high risk community-dwelling older adults. \u003cem\u003eInjury Epidemiology\u003c/em\u003e. 2014/03/20 2014;1(1):5. doi:10.1186/2197-1714-1-5\u003c/li\u003e\n\u003cli\u003eMoreland BL, Kakara R, Haddad YK, Shakya I, Bergen G. A Descriptive Analysis of Location of Older Adult Falls That Resulted in Emergency Department Visits in the United States, 2015. \u003cem\u003eAmerican Journal of Lifestyle Medicine\u003c/em\u003e. 2021/11/01 2020;15(6):590\u0026ndash;597. doi:10.1177/1559827620942187\u003c/li\u003e\n\u003cli\u003eRubenstein LZ. Falls in older people: epidemiology, risk factors and strategies for prevention. \u003cem\u003eAge Ageing\u003c/em\u003e. Sep 2006;35 Suppl 2:ii37\u0026ndash;ii41. doi:10.1093/ageing/afl084\u003c/li\u003e\n\u003cli\u003ePage MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. \u003cem\u003eBMJ\u003c/em\u003e. 2021;372:n71. doi:10.1136/bmj.n71\u003c/li\u003e\n\u003cli\u003eThe National Heart L, and Blood Institute. Study Quality Assessment Tools. Accessed Nov, 25th, 2024. https://www.nhlbi.nih.gov/health-topics/study-quality-assessment-tools\u003c/li\u003e\n\u003cli\u003eLo Y, Lynch SF, Urbanowicz RJ, et al. Using Machine Learning on Home Health Care Assessments to Predict Fall Risk. \u003cem\u003eStudies in health technology and informatics\u003c/em\u003e. Aug 21 2019;264:684\u0026ndash;688. doi:10.3233/shti190310\u003c/li\u003e\n\u003cli\u003eDu R, Jagtap V, Long Y, Onwuka O, Padir T. Robotics enabled in-home environment screening for fall risks. presented at: Proceedings of the 2014 workshop on Mobile augmented reality and robotic technology-based systems; 2014; Bretton Woods, New Hampshire, USA. https://doi.org/10.1145/2609829.2609836 \u003c/li\u003e\n\u003cli\u003eLathouwers E, Dillen A, D\u0026iacute;az MA, et al. Characterizing fall risk factors in Belgian older adults through machine learning: a data-driven approach. \u003cem\u003eBMC Public Health\u003c/em\u003e. Nov 29 2022;22(1):2210. doi:10.1186/s12889-022-14694-5\u003c/li\u003e\n\u003cli\u003eCloix S, Bologna G, Weiss V, Pun T, Hasler D. Low-power depth-based descending stair detection for smart assistive devices. \u003cem\u003eEURASIP Journal on Image and Video Processing\u003c/em\u003e. 2016/10/12 2016;2016(1):33. doi:10.1186/s13640-016-0133-6\u003c/li\u003e\n\u003cli\u003eP\u0026eacute;rez-Ros P, Mart\u0026iacute;nez-Arnau FM, Orti-Lucas RM, Tarazona-Santabalbina FJ. A predictive model of isolated and recurrent falls in functionally independent community-dwelling older adults. \u003cem\u003eBraz J Phys Ther\u003c/em\u003e. Jan\u0026ndash;Feb 2019;23(1):19\u0026ndash;26. doi:10.1016/j.bjpt.2018.05.005\u003c/li\u003e\n\u003cli\u003eMoore J, Celik Y, Stuart S, et al. Using Video Technology and AI within Parkinson\u0026rsquo;s Disease Free-Living Fall Risk Assessment. \u003cem\u003eSensors\u003c/em\u003e. 2024;24(15):4914. \u003c/li\u003e\n\u003cli\u003eChen X, He L, Shi K, Wu Y, Lin S, Fang Y. Interpretable Machine Learning for Fall Prediction Among Older Adults in China. \u003cem\u003eAm J Prev Med\u003c/em\u003e. Oct 2023;65(4):579\u0026ndash;586. doi:10.1016/j.amepre.2023.04.006\u003c/li\u003e\n\u003cli\u003ePanyakaew P, Pornputtapong N, Bhidayasiri R. Using machine learning-based analytics of daily activities to identify modifiable risk factors for falling in Parkinson\u0026apos;s disease. \u003cem\u003eParkinsonism Relat Disord\u003c/em\u003e. Jan 2021;82:77\u0026ndash;83. doi:10.1016/j.parkreldis.2020.11.014\u003c/li\u003e\n\u003cli\u003eMillet A, Madrid A, Alonso-Weber JM, Rodr\u0026iacute;guez-Ma\u0026ntilde;as L, P\u0026eacute;rez-Rodr\u0026aacute;-Guez R. Machine Learning Techniques Applied to the Development of a Fall Risk Index for Older Adults. \u003cem\u003eIEEE Access\u003c/em\u003e. 2023;11:84795\u0026ndash;84809. doi:10.1109/ACCESS.2023.3299489\u003c/li\u003e\n\u003cli\u003eCampani D, Caristia S, Amariglio A, et al. Home and environmental hazards modification for fall prevention among the elderly. \u003cem\u003ePublic health nursing (Boston, Mass)\u003c/em\u003e. May 2021;38(3):493\u0026ndash;501. doi:10.1111/phn.12852\u003c/li\u003e\n\u003cli\u003ePynoos J, Steinman BA, Nguyen AQ. Environmental assessment and modification as fall-prevention strategies for older adults. \u003cem\u003eClinics in geriatric medicine\u003c/em\u003e. Nov 2010;26(4):633\u0026ndash;44. doi:10.1016/j.cger.2010.07.001\u003c/li\u003e\n\u003cli\u003eFaulkner S, Burns D. Identifying modifiable environmental risk factors for falls in older people. \u003cem\u003eInjury Prevention\u003c/em\u003e. 2010;16(Suppl 1):A254. doi:10.1136/ip.2010.029215.905\u003c/li\u003e\n\u003cli\u003eMontero-Odasso M, Pieruccini-Faria F, Son S, et al. Fall risk stratification in older adults: low and not-at-risk status still associated with falls and injuries. \u003cem\u003eAge Ageing\u003c/em\u003e. Mar 3 2025;54(3)doi:10.1093/ageing/afaf064\u003c/li\u003e\n\u003cli\u003eTang S, Liu M, Yang T, et al. Association between falls in elderly and the number of chronic diseases and health-related behaviors based on CHARLS 2018: health status as a mediating variable. \u003cem\u003eBMC Geriatr\u003c/em\u003e. Apr 28 2022;22(1):374. doi:10.1186/s12877-022-03055-x\u003c/li\u003e\n\u003cli\u003eGillespie LD, Robertson MC, Gillespie WJ, et al. Interventions for preventing falls in older people living in the community. \u003cem\u003eCochrane Database of Systematic Reviews\u003c/em\u003e. 2012;(9)doi:10.1002/14651858.CD007146.pub3\u003c/li\u003e\n\u003cli\u003eZiebart C, Dewan N, Tuazon J, MacDermid J. Development of the Home Fall Hazard Checklist. \u003cem\u003eRehabilitation Research and Practice\u003c/em\u003e. 2021/06/01 2021;2021:5362197. doi:10.1155/2021/5362197\u003c/li\u003e\n\u003cli\u003eCampani D, Caristia S, Amariglio A, et al. Home and environmental hazards modification for fall prevention among the elderly. \u003cem\u003ePublic Health Nursing\u003c/em\u003e. 2021/05/01 2021;38(3):493\u0026ndash;501. doi:https://doi.org/10.1111/phn.12852\u003c/li\u003e\n\u003cli\u003eClemson L, Fitzgerald MH, Heard R. Content Validity of an Assessment Tool to Identify Home Fall Hazards: The Westmead Home Safety Assessment. \u003cem\u003eBritish Journal of Occupational Therapy\u003c/em\u003e. 1999/04/01 1999;62(4):171\u0026ndash;179. doi:10.1177/030802269906200407\u003c/li\u003e\n\u003cli\u003ePighills A, Drummond A, Crossland S, Torgerson DJ. What type of environmental assessment and modification prevents falls in community dwelling older people? \u003cem\u003eBmj\u003c/em\u003e. Mar 14 2019;364:l880. doi:10.1136/bmj.l880\u003c/li\u003e\n\u003cli\u003eSadasivam RS, Luger TM, Coley HL, et al. Robot-assisted home hazard assessment for fall prevention: a feasibility study. \u003cem\u003eJournal of Telemedicine and Telecare\u003c/em\u003e. 2014/01/01 2013;20(1):3\u0026ndash;10. doi:10.1177/1357633X13517350\u003c/li\u003e\n\u003cli\u003eBlanchet R, Edwards N. A need to improve the assessment of environmental hazards for falls on stairs and in bathrooms: results of a scoping review. \u003cem\u003eBMC Geriatrics\u003c/em\u003e. 2018/11/09 2018;18(1):272. doi:10.1186/s12877-018-0958-1\u003c/li\u003e\n\u003cli\u003eJiang Y-S, Shi H, Kang Y-T, et al. Impact of age-friendly living environment and intrinsic capacity on functional ability in older adults: a cross-sectional study. \u003cem\u003eBMC Geriatrics\u003c/em\u003e. 2023/06/17 2023;23(1):374. doi:10.1186/s12877-023-04089-5\u003c/li\u003e\n\u003cli\u003eBayar R, T\u0026uuml;rkoğlu H. The relationship between living environment and daily life routines of older adults. \u003cem\u003eA/Z ITU Journal of the Faculty of Architecture\u003c/em\u003e. 2021;18(1):29\u0026ndash;43. \u003c/li\u003e\n\u003cli\u003eOlden K, White SL. Health-Related Disparities: Influence of Environmental Factors. \u003cem\u003eMedical Clinics\u003c/em\u003e. 2005;89(4):721\u0026ndash;738. doi:10.1016/j.mcna.2005.02.001\u003c/li\u003e\n\u003cli\u003eGee GC, Payne-Sturges DC. Environmental health disparities: a framework integrating psychosocial and environmental concepts. \u003cem\u003eEnviron Health Perspect\u003c/em\u003e. Dec 2004;112(17):1645\u0026ndash;53. doi:10.1289/ehp.7074\u003c/li\u003e\n\u003cli\u003eJacobs DE. Environmental Health Disparities in Housing. \u003cem\u003eAmerican Journal of Public Health\u003c/em\u003e. 2011/12/01 2011;101(S1):S115\u0026ndash;S122. doi:10.2105/AJPH.2010.300058\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"discover-artificial-intelligence","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"diai","sideBox":"Learn more about [Discover Artificial Intelligence](https://www.springer.com/44163)","snPcode":"","submissionUrl":"","title":"Discover Artificial Intelligence","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Fall Risk Assessment, Predictive Modeling, Environmental Factors, Systematic Review, Nursing Informatics","lastPublishedDoi":"10.21203/rs.3.rs-8723907/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8723907/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cb\u003eBackground\u003c/b\u003e\u003c/p\u003e \u003cp\u003eFalls commonly occur in home environments where environmental conditions can contribute to fall risk. Identification and mitigation of environmental hazards are critical components of fall prevention. However, artificial intelligence (AI)-based fall prediction models have largely focused on individual-level predictors, with limited attention to home environmental hazards despite their modifiable role in fall risk.\u003c/p\u003e\u003cp\u003e\u003cb\u003eObjective\u003c/b\u003e\u003c/p\u003e \u003cp\u003eTo systematically review how environmental factors are incorporated into existing AI-based fall risk prediction models and summarize reported AI approaches and model performance among community-dwelling older adults.\u003c/p\u003e\u003cp\u003e\u003cb\u003eMethods\u003c/b\u003e\u003c/p\u003e \u003cp\u003eThis systematic review followed PRISMA guidelines. Six electronic databases (PubMed, Embase, CINAHL, Cochrane Library, Web of Science, and Scopus) were searched from inception through December 2025. Eligible studies applied AI-based models to predict falls among older adults in community settings and incorporated environmental factors as model inputs.\u003c/p\u003e\u003cp\u003e\u003cb\u003eResults\u003c/b\u003e\u003c/p\u003e \u003cp\u003eOf more than 10,000 records identified, nine studies met final inclusion criteria. Six used supervised machine learning with structured data, while three employed computer vision or robotics-based approaches. Environmental factors were heterogeneously represented, ranging from checklist-based indicators to sensor- and vision-derived measures. When included, environmental features contributed meaningful information by improving discrimination or identifying actionable home hazards (AUC-ROC ranged from 0.67 to 0.76).\u003c/p\u003e\u003cp\u003e\u003cb\u003eConclusion\u003c/b\u003e\u003c/p\u003e \u003cp\u003eEnvironmental factors remain underemphasized in AI-based fall prediction models. Greater integration of standardized and context-aware environmental information may enhance the relevance and preventive utility of AI-based fall risk prediction in community settings.\u003c/p\u003e","manuscriptTitle":"Are We Missing the Environmental Factors in AI-Based Fall Risk Models?: A Systematic Review","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-02 06:02:48","doi":"10.21203/rs.3.rs-8723907/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-02-05T10:08:09+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-30T13:15:31+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-01-30T13:12:16+00:00","index":"","fulltext":""},{"type":"submitted","content":"Discover Artificial Intelligence","date":"2026-01-28T16:44:47+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"discover-artificial-intelligence","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"diai","sideBox":"Learn more about [Discover Artificial Intelligence](https://www.springer.com/44163)","snPcode":"","submissionUrl":"","title":"Discover Artificial Intelligence","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"0d1348f0-d04c-49cb-ab1e-e6cbe924f60a","owner":[],"postedDate":"February 2nd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-02-07T14:08:56+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-02 06:02:48","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8723907","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8723907","identity":"rs-8723907","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}