Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Am Med Dir Assoc . Author manuscript; available in PMC: 2026 Apr 15. Published in final edited form as: J Am Med Dir Assoc. 2026 Apr 30;27(6):106189. doi: 10.1016/j.jamda.2026.106189 Search in PMC Search in PubMed View in NLM Catalog Add to search Using Natural Language Processing to Improve Fall Documentation in VA Nursing Home Residents Laura A Graham Laura A Graham , PhD 1. Health Economics Resource Center, VA Palo Alto Health Care System, Palo Alto, California, USA. 2. S-SPIRE, Department of Surgery, Stanford University, Stanford, CA Find articles by Laura A Graham 1, 2 , Xiaojuan Liu Xiaojuan Liu , PhD 3. Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School, Boston, Massachusetts, USA. Find articles by Xiaojuan Liu 3 , Bocheng Jing Bocheng Jing , MS 4. Geriatrics, Palliative, and Extended Care Service Line, San Francisco VA Medical Center, San Francisco, California, USA. Find articles by Bocheng Jing 4 , Sei J Lee Sei J Lee , MD 4. Geriatrics, Palliative, and Extended Care Service Line, San Francisco VA Medical Center, San Francisco, California, USA. 5. Division of Geriatrics, Department of Medicine, University of California of San Francisco, San Francisco, California, USA. Find articles by Sei J Lee 4, 5 , Michael A Steinman Michael A Steinman , MD 4. Geriatrics, Palliative, and Extended Care Service Line, San Francisco VA Medical Center, San Francisco, California, USA. 5. Division of Geriatrics, Department of Medicine, University of California of San Francisco, San Francisco, California, USA. Find articles by Michael A Steinman 4, 5 , Christine Kee Liu Christine Kee Liu , MD 6. Geriatric Research Education and Clinical Center, VA Palo Alto Health Care System, Palo Alto, California, USA. 7. Section of Geriatrics, Division of Primary Care and Population Health, Stanford University, School of Medicine, Stanford, California, USA. Find articles by Christine Kee Liu 6, 7 , Chintan V Dave Chintan V Dave , PharmD, PhD 8. Department of Pharmacy Practice and Administration, Ernest Mario School of Pharmacy, Rutgers University, Piscataway, New Jersey, USA. 9. Center for Pharmacoepidemiology and Treatment Science, Institute for Health, Health Care Policy and Aging Research, Rutgers University, New Brunswick, New Jersey, USA. Find articles by Chintan V Dave 8, 9 , Yongmei Li Yongmei Li , PhD 6. Geriatric Research Education and Clinical Center, VA Palo Alto Health Care System, Palo Alto, California, USA. 10. Department of Epidemiology and Population Health, Stanford University, Stanford, California, USA. Find articles by Yongmei Li 6, 10 , Kathy Fung Kathy Fung , MS 4. Geriatrics, Palliative, and Extended Care Service Line, San Francisco VA Medical Center, San Francisco, California, USA. 5. Division of Geriatrics, Department of Medicine, University of California of San Francisco, San Francisco, California, USA. Find articles by Kathy Fung 4, 5 , Michelle C Odden Michelle C Odden , PhD 6. Geriatric Research Education and Clinical Center, VA Palo Alto Health Care System, Palo Alto, California, USA. 10. Department of Epidemiology and Population Health, Stanford University, Stanford, California, USA. Find articles by Michelle C Odden 6, 10 Author information Article notes Copyright and License information 1. Health Economics Resource Center, VA Palo Alto Health Care System, Palo Alto, California, USA. 2. S-SPIRE, Department of Surgery, Stanford University, Stanford, CA 3. Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital and Harvard Medical School, Boston, Massachusetts, USA. 4. Geriatrics, Palliative, and Extended Care Service Line, San Francisco VA Medical Center, San Francisco, California, USA. 5. Division of Geriatrics, Department of Medicine, University of California of San Francisco, San Francisco, California, USA. 6. Geriatric Research Education and Clinical Center, VA Palo Alto Health Care System, Palo Alto, California, USA. 7. Section of Geriatrics, Division of Primary Care and Population Health, Stanford University, School of Medicine, Stanford, California, USA. 8. Department of Pharmacy Practice and Administration, Ernest Mario School of Pharmacy, Rutgers University, Piscataway, New Jersey, USA. 9. Center for Pharmacoepidemiology and Treatment Science, Institute for Health, Health Care Policy and Aging Research, Rutgers University, New Brunswick, New Jersey, USA. 10. Department of Epidemiology and Population Health, Stanford University, Stanford, California, USA. ✉ Corresponding Author: Laura A. Graham, PhD, MPH, 795 Willow Road, Bldg. 324, Rm D-146, Menlo Park, CA 94025, [email protected] ; [email protected] Issue date 2026 Jun. PMC Copyright notice PMCID: PMC13078714 NIHMSID: NIHMS2156201 PMID: 41956437 The publisher's version of this article is available at J Am Med Dir Assoc Abstract Objective Falls can be deadly for older adults, making accurate documentation essential for prevention and quality improvement. The Minimum Data Set (MDS) is used in nursing homes to document falls. However, it fails to capture fall dates and has been critiqued for underreporting and misclassification. We developed a natural language processing (NLP) algorithm to extract fall dates from electronic health records (EHRs). This study outlines the NLP algorithm development and compares its findings to the MDS to evaluate misclassification and underreporting. Design Retrospective cohort study Setting and Participants All Veterans residing in long-term care between 07/01/2012 and 12/31/2024 Methods A rule-based NLP algorithm was developed to identify fall dates in the EHR and validated with manual chart review. The NLP results (reference) were compared to the MDS v3.0 assessment to assess the concordance between the two data sources. Results We identified 38,852 long-term care residents, with a median length of stay of 204 days. About half (49.0%) of residents experienced at least one fall during their stay, as identified by either data source. The NLP algorithm achieved 96% accuracy against manual chart review. Using the NLP output as the reference standard, the MDS correctly identified 86.1% of residents with a fall in the EHR (F1-score = 0.885, and specificity = 0.930). However, 2,430 (6.3%) residents had falls documented in their medical records that were never reported in the MDS, and 3.9% of MDS falls were not identified in the EHR (misclassification). The median time between a fall and an MDS assessment documenting a fall was 17 days. Conclusions and Implications The MDS identified 86% of NH residents with EHR documentation of a fall, but it did not provide the specific date of each fall. By supplementing MDS fall reporting with NLP, we may improve the sensitivity of fall detection and gain more precise information regarding when a fall occurs. Keywords: falls, NLP, nursing home Summary: We built an NLP algorithm to extract fall dates from medical records and compared it with MDS. It was 96% accurate vs chart review. MDS missed 6.3% of falls and reported them 17 days late. Introduction More than 40% of nursing home residents will experience at least one fall during their stay. These falls can result in severe injuries, such as fractures, head trauma, and even death. 1 , 2 Falls are among the leading causes of disability and loss of independence among older adults. 3 Documentation of falls, including incident reporting, is essential for understanding their prevalence and informing fall prevention efforts. This practice is often required at long-term care centers. One example of a nationally mandated fall reporting system is the Centers for Medicare and Medicaid Services (CMS) Minimum Data Set (MDS). 4 The MDS is mandatory for all Medicare- or Medicaid-certified nursing homes and is used to assess various quality metrics that influence reimbursement levels for long-term care facilities in the U.S. 5 - 8 The current MDS fall documentation process has several limitations and has been critiqued for underreporting and misclassification. The process may miss as many as 20% of falls documented in the medical record and is extremely resource-intensive, requiring chart reviews. 49 Additionally, MDS does not capture the exact date of a fall. It only indicates that a fall occurred during an assessment period, which may span up to 90 days. These limitations contribute to the under-reporting of falls, hindering quality measurement, research, and safety efforts. 10 Supplementing the current process with natural language processing (NLP) to extract fall information from the medical chart could improve the MDS fall documentation process. NLP has been shown to improve the efficiency and accuracy of event reporting across multiple quality metrics, including post-traumatic stress disorder and heart failure. 11 , 12 In the case of falls, NLP could improve the accuracy of MDS fall documentation, reduce the time and resources required for mandated chart reviews, and provide more detailed information on the date of the fall. 13 , 14 This study describes the development of a rule-based NLP algorithm to identify falls documented in the Veterans Health Administration (VA) electronic health record (EHR). Following validation through chart reviews, we compared the algorithm's results to the MDS fall documentation, using the NLP findings as the reference standard. Our research objectives were to assess the misclassification rate and quantify the lag between fall occurrences and MDS documentation. Methods Study Population, Data Sources, and Ethical Considerations The study population includes Veterans aged 65 years and older who resided in a VA nursing home (Community Living Center, CLC) between July 1, 2012, and December 31, 2024. Residents with fewer than 14 days in the nursing home were excluded to allow at least 2 weeks observation time for all participants (n=1,617/40,469, 4.0%). This study was reviewed and approved by the Stanford Institutional Review Board (Protocol #49235) with a waiver of consent. Demographic and comorbidity data were obtained from the CDW SPatient, Inpatient, and Outpatient domains. Clinical documents were extracted from the VA Corporate Data Warehouse (CDW) Text Integration Utility (TIU) domain, including nursing notes, physician notes, and incident reports (“Post Fall Notes”). MDS v3.0 assessments were obtained for all residents. The following variables were used to assess falls documented in MDS: J1900A. Falls Since Admit/Prior Assessment: No Injury J1900B: Falls Since Admit/Prior Assessment: Injury (not Major) J1900C: Falls Since Admit/Prior Assessment: Injury (Major) It is important to recognize that reliance on J1900 codes to estimate fall events may undercount fall occurrences, particularly when multiple falls occur between assessments. Each MDS variable is coded as (0) none, (1) one, and (2) two or more falls. Our primary outcome for validation was the occurrence of any fall during the nursing home stay as a binary variable. As a secondary outcome, we also estimated the total number of falls documented in the MDS as the sum of J1900A, J1900B, and J1900C. Algorithm Development and Validation We developed a rule-based NLP algorithm to identify fall dates and validated it using manual chart review ( Supplement ). Step 1: Initiate seed phrases. To begin developing the algorithm, we curated a list of seed phrases commonly associated with fall events in clinical documentation. These seed phrases were derived from clinical expertise, literature review, and exploratory analysis of EHR text data. Examples of seed phrases included “post fall”, “patient fell”, “date of fall”, and “time of fall.” This initial list served as the foundation for identifying notes potentially indicative of fall events. Step 2: Draft the algorithm and expand keywords. We then randomly selected 200 clinical notes that included seed phrases. Each note was manually reviewed to identify linguistic patterns, contextual clues, and structural characteristics indicative of fall-related dates. This information was used to develop the first draft of the NLP algorithm. The majority of fall dates were extracted from a variation of the phase “date/time of event/fall” or from the identification of a date in the note (ex, “patient fell on 6/6/20” or “patient fell at 1200”). When we were unable to extract the fall date from the note, the note date was used as the fall date (see Supplement ). While expanding our list of keywords, we noted variations in phrasing (e.g., “patient fell” vs. “post fall”) and contextual modifiers (e.g., “fall risk assessment” vs. “hx of falls: now”) that were also added to the algorithm. We identified key document titles (“fall risk assessment note” and “post fall assessment”) and semi-structured notes consistent with recommendations from the VHA National Center for Patient Safety’s Falls Toolkit, and noted variations in documentation across facilities. This informed the expansion of our keyword list with specific variations for certain facilities. Since our goal was to identify fall dates, negation (“patient denies history of falls” or “no falls reported”) was ignored. Step 3: Error analysis and rule refinement. After determining our initial set of rules, we randomly selected a set of 200 residents and compared the results of the algorithm against a manual chart review to calculate accuracy, the overall proportion of correct predictions. Errors were categorized as false positives or false negatives, and root causes for each error were analyzed. Based on this error analysis, we refined the rules by adding new keywords or phrases, adjusting contextual rules to better handle ambiguous text, adding facility-specific special cases, and fine-tuning temporal indicators. Step 4: Iterate and revise the algorithm. After each revision, we reapplied the revised algorithm to the full dataset, extracted another random sample of 200 residents, and conducted another manual chart review. This process was repeated until the algorithm met a predetermined accuracy threshold of 95% or higher, which was reached in the fourth iteration. Once complete, we performed event-level chart abstraction validation on a random subset of 200 algorithm-identified falls. The algorithm correctly identified the fall date for 94.5% of the random subset. All incorrectly identified fall dates had evidence of a fall within 7 days before the algorithm-identified date (64% occurred within 2 days prior). Statistical Analysis Descriptive statistics were used to characterize the study population, and fall rates were reported as the number of falls per 1,000 patient-days, as is commonly reported for quality improvement purposes. To assess overlap, we calculated the sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score. 95% confidence intervals were computed for each metric using bootstrapping with 1,000 iterations. To estimate MDS misclassification of falls, we compared the NLP algorithm's detection of any fall in the EHR (reference standard) with those recorded by an MDS v3.0 assessment in the same cohort of residents during the study period. Because MDS assessments do not identify the exact date of the fall or the exact number of falls, performance metrics were calculated only on the presence of any fall during the patient’s stay. To assess the time delay in reporting, the days between an EHR-documented fall and a subsequent MDS assessment were calculated for each MDS assessment. All statistical analyses were performed using SAS version 9.4 and R program. Results Between July 1, 2012, and December 31, 2024, 38,852 residents were in long-term care at 110 VA CLCs. Cohort demographics were consistent with those of the Veteran population: 97.6% male, 77.6% white, and a median age of 76 years at admission. The median length of stay was 204 days (interquartile range (IQR): 95-586 days; Table 1 ). Overall, 49.0% of residents experienced at least one fall, identified by either data source, during their stay. Compared to chart abstractions (n=200 in the final iteration), the NLP algorithm demonstrated sensitivity of 94.0% (95% CI 88.0%-97.5%) and specificity of 98.8% (95% CI 93.5%-99.9%), with an overall accuracy of 96.0% (92.3%-98.3%). When compared with all chart abstractions performed (n=800), the algorithm demonstrated greater sensitivity (85.4% vs. 53.8%) and greater specificity (96.1% vs. 73.4%) than the MDS ( Table 2 ). Table 1. Demographic Characteristics for Long-Term Care Residents at the Time of Admission Characteristics n (%) / median (IQR) Long-term care residents, n 38,852 Patient days in nursing home, n 18,473,233 Female, n (%) 920 (2.4%) White race, n (%) 28,192 (77.6%) Black race, n (%) 7,149 (19.7%) Other race, n (%) 987 (2.7%) Age at admission, years, median (IQR) 76 (70-84) ADLs at admission, median (IQR) 15 (9-19) Charlson Comorbidity Index, median (IQR) 5 (3-7) Open in a new tab IQR = Interquartile Range, ADL = Activities of Daily Living Table 2. Confusion Matrix of Medical Chart Abstraction Results Compared to Natural Language Processing (NLP) Algorithm Results and Minimum Data Set (MDS) Documented Falls Any Fall in the Medical Chart (n=800) * Any Fall in the Medical Chart (n=773) ** Yes No Yes No Any NLP Fall Yes 379 (47.4%) 14 (1.8%) Any CMS MDS Fall Yes 232 (30.0%) 91 (11.8%) No 65 (8.1%) 342 (42.8%) No 199 (25.7% 251 (32.5%) Performance Metrics (95% CI) PPV 0.964 (0.942, 0.978) PPV 0.718 (0.677, 0.756) NPV 0.840 (0.808, 0.868) NPV 0.558 (0.528, 0.587) Sensitivity 0.854 (0.817, 0.885) Sensitivity 0.538 (0.490, 0.586) Specificity 0.961 (0.935, 0.978) Specificity 0.734 (0.684, 0.780) F1-Score 0.906 F1-Score 0.615 Open in a new tab NLP = Natural Language Processing, CMS = Centers for Medicare and Medicaid Services, MDS = Minimum Data Set * Final NLP results compared to all chart abstractions performed (n=800). ** MDS comparison group includes all chart abstractions performed for residents with at least one MDS assessment during their nursing home stay (n=773) Overall, 17,524 (45.1%) residents were identified by the NLP algorithm as experiencing at least one fall during their stay, and 16,601 (42.7%) residents had at least one MDS assessment documenting a recent fall ( Table 3 ). During the study period, 55,142 MDS assessments were completed, documenting at least 67,977 falls (3.7 falls per 1,000 patient days; Table 3 ); 34.3% involved an injury, with 3.8% resulting in a major injury. Using the NLP algorithm, we identified 109,264 falls over the same period (5.9 falls per 1,000 patient days). The median time lag between documenting a fall in the EHR and an MDS assessment documenting a fall was 17 days (IQR=6-40 days). However, of the 55,142 MDS assessments documenting a fall, only 50.5% had a fall documented in the EHR in the 30 days preceding the assessment. Table 3. Fall Documentation by Data Source NLP Results (Reference) CMS MDS Residents (n) 38,852 Resident days (n) 18,473,233 Any Fall (n, %) 16,407 (45.2%) 15,511 (42.7%) Number of Falls (n) 109,264 67,977 * No Injury (J1900A) 44,654 * Minor Injury (J1900B) 20,721 * Major Injury (J1900C) 2,602 * Open in a new tab NLP = Natural Language Processing, CMS = Centers for Medicare and Medicaid Services, MDS = Minimum Data Set * Note that this will result in an underestimate of the total number of falls documented by MDS when a patient experiences more than 2 falls in a single assessment period. 26.2% (n=14,118) of fall-positive MDS assessments were documented as 2 or more falls, 74.8% of which were falls with no injury. Compared with the NLP algorithm results, the MDS assessments demonstrated an F1-score of 0.885 ( Table 4 ), indicating good performance. The MDS correctly identified 86.1% of residents with an EHR fall (sensitivity). However, 2,430 (6.3%) residents had falls documented in their medical charts that were never reported in the MDS, suggesting potentially missed fall reporting or errors in MDS documentation. Conversely, 1,507 (3.9%) residents had falls documented in the MDS that our NLP algorithm did not identify, indicating either an MDS documentation error or a fall not documented in the EHR ( Table 4 ). Table 4. Confusion Matrix & Natural Language Processing (NLP) Algorithm Performance Metrics Any NLP Fall (Reference) Yes No Any MDS Fall Yes 15,094 (38.8%) 1,507 (3.9%) No 2,430 (6.3%) 19,821 (51.0%) Performance Metrics (95% CI) PPV 0.909 (0.905, 0.914) NPV 0.891 (0.887, 0.895) Sensitivity 0.861 (0.856, 0.866) Specificity 0.929 (0.926, 0.933) F1-Score 0.885 Open in a new tab NLP = Natural Language Processing, CMS = Centers for Medicare and Medicaid Services, MDS = Minimum Data Set Falls were more common in older male white residents with longer stays ( Table 5 ). The prevalence of falls was highest early in the study period (2021: 56.3%) and declined to a low of 20.9% in the final year of follow-up (2024). The median time between an NLP-identified fall date and a subsequent MDS assessment with a fall was 10 days (IQR: 3-26) for falls with a major injury, 16 days (IQR: 6-36) for falls with a minor injury, and 19 days (IQR: 6-42) for falls with no injury. The lag time to an MDS assessment of a fall was longer for females and older residents. Table 5. Prevalence of Any Fall and Time Lag Between an NLP Fall Date and a Subsequent MDS Assessment with a Fall by Patient Characteristics % Any Fall p-value Days Between NLP Fall Date and MDS Assessment (Median, IQR) p-value Sex Women 39.8% <0.001 20 (7-42) 0.03 Men 49.2% 17 (6-40) Age at Admission 65-75 45.0% <0.001 16 (5-38) <0.001 76-85 49.3% 18 (6-41) >85 57.0% 19 (7-42) Race White 51.0% <0.001 17 (6-40) 0.06 Black or African American 41.1% 18 (6-41) Other 50.3% 19 (6-43) Length of Stay 14-90 days 26.6% <0.001 5 (2-12) <0.001 91-365 days 47.5% 10 (4-26) >365 days 64.3% 21 (8-45) Calendar Year 2012 56.3% <0.001 15 (6-33) <0.001 2013 53.7% 18 (6-41) 2014 52.4% 17 (6-40) 2015 54.1% 17 (6-40) 2016 54.3% 17 (6-41) 2017 54.6% 16 (5-40) 2018 52.1% 18 (6-41) 2019 49.6% 17 (6-38) 2020 46.5% 18 (6-40) 2021 40.7% 18 (6-41) 2022 35.4% 20 (7-42) 2023 30.5% 18 (6-40) 2024 20.9% 16 (5-39) Open in a new tab The median lag time for residents with a length of stay of 90 days or less was 5 days (IQR: 2-12), but increased to 21 days (IQR: 8-45) for residents with stays exceeding 365 days. While lag time varied by year (p<0.01), there was no significant temporal trend (p=0.47; Table 5 ). Discussion Our results show that an NLP algorithm can accurately identify falls documented in clinical notes and add detailed information on the actual dates of falls, which are not currently available in the MDS. Our findings also quantify the potential magnitude of misclassification and underreporting of falls using MDS data alone versus our NLP algorithm, and highlight variation in the time between an actual fall and its documentation in MDS. We found that 86.1% of residents with at least one fall documented in the EHR also had a fall documented in the MDS; however, 6.3% of residents had falls documented in the EHR that were never reported in the MDS, and 3.9% of MDS falls were not identified in the EHR. Our findings of >80% agreement between falls recorded in the MDS and the EHR are higher than previously reported (57.5%). 6 Our more optimistic finding may be explained by our higher proportion of white residents and our restriction to long-term care residents, both of which have been associated with better MDS documentation and may have biased us toward better agreement. 5 , 6 However, we still identified substantial undercounting of total falls in MDS, with as many as 61% more falls not reported in MDS. Our findings highlight the limitations of the MDS fall documentation process. This approach cannot capture exact fall dates, precisely count more than two falls, or reliably detect mild falls due to recall bias, leading to misclassification and underreporting. 15 , 16 In contrast to the MDS, our algorithm was developed to identify the date of a fall. We found that the MDS look-back period may result in an average lag of approximately 2-3 weeks in documenting a fall. The use of the NLP algorithm to extract exact fall dates, rather than relying solely on MDS assessment dates, could provide valuable insights into fall timing and help clinicians estimate the time until the first fall and the intervals between subsequent falls. Identifying the exact date of each fall can be helpful for research and quality improvement efforts that aim to evaluate what happened immediately before and after a fall, a capability that MDS currently lacks. To our knowledge, this study is the first to develop an NLP algorithm to identify falls in a nursing home setting and the first algorithm explicitly developed for the VA clinical notes. Recently, we have seen a shift from rule-based to machine-learning-based approaches in the use of artificial intelligence methods. While each has its advantages, rule-based NLP algorithms tend to exhibit the best performance, as measured by the F1-score. 14 Rule-based systems also offer greater transparency and interpretability compared to machine learning models. We’ve provided our final NLP algorithm as a supplement to allow other researchers to reproduce and improve upon our findings. While machine learning approaches may offer significant advantages in many NLP tasks, rule-based systems continue to play an important role in specific scenarios where transparency, control, and precision are paramount. Our findings demonstrate not only feasibility but also high accuracy with a simpler, more easily understood rule-based algorithm. These findings of underreporting and delayed reporting have several implications for clinical practice. First, accurate reporting of falls through the MDS is essential, as it directly affects payment, compliance, and the facility’s reputation. 16 Utilizing NLP for MDS fall assessments can lead to more accurate documentation of falls, impacting quality improvement initiatives, public reporting, and pay-for-reporting programs. Second, the current look-back period used by MDS may not provide sufficient granularity for quality improvement initiatives, which require specific information about the date and context of the fall. NLP can provide more accurate information on the actual date of the fall, a significant limitation of the current MDS assessments. 17 In addition, the process of collecting fall data for MDS reporting is resource-intensive and becoming increasingly cumbersome amid growing staffing shortages in long-term care. 18 NLP could be used to identify fall events for further abstraction, reducing the time required to gather this information and freeing up staff for patient care without sacrificing documentation accuracy. Limitations We note that the predominantly male, VA-specific cohort may limit the generalizability of the findings to non-VA populations. While the prevalence rates observed in this study provide valuable insights within this specific context, they may not directly translate to broader populations. To assess the magnitude of misclassification in MDS, we use the NLP algorithm as our reference. Our algorithm was developed using data from a single healthcare system and a specific population and may lack generalizability to other settings. However, our population was sampled from the largest single healthcare system in the US, and our cohort includes residents from 110 facilities nationwide. Furthermore, we cannot determine whether the note text was copied and pasted from previous notes, which may affect the accuracy of our fall date extraction. However, an event-level validation found no evidence of this concerning fall dates. Additionally, of the incorrectly identified fall dates, 64% occurred within 2 days of the algorithm-identified fall date, and 100% occurred within 7 days, suggesting only a minor impact on accuracy. NLP algorithms developed to parse VA-specific templates may not perform well on notes that are differently structured from other systems. In addition, the algorithm was developed specifically for VA nursing home residents and may not be appropriate for other care settings. Future studies will need to be conducted to vet the algorithm in other populations and healthcare settings fully. Nonetheless, our methods to develop the algorithm could be easily repeated in other healthcare settings to tailor our NLP algorithms to a specific healthcare system. Finally, the MDS has acknowledged its limitations in accurately capturing the timing and total number of fall events. While the NLP approach can address these limitations, it currently cannot determine the severity or clinical consequences of a fall. This limitation restricts its usefulness in evaluating the full spectrum of fall-related morbidity. Falls may also still be underreported in the EHR, particularly when staff fail to document the event, which can bias the assessment of fall risk factors. 19 - 21 Research indicates that staff may not consistently report falls in the medical chart, particularly minor incidents that do not result in injury. 5 , 20 This underreporting can bias data used for research and quality improvement, leading to inaccurate estimates of fall rates and of risk factors. 5 , 6 Conclusions and Implications In conclusion, the MDS lacks specific dates for each fall and may underreport the total number of falls documented in the medical record. By supplementing MDS fall reporting with NLP, we can improve fall detection sensitivity and obtain precise timing for each incident. Falls in nursing homes are documented through incident reports and the MDS, but often suffer from underreporting and inconsistent data collection, which complicates fall research and prevention strategies. Researchers should recognize these limitations and consider using additional methods, such as interviews, chart abstraction, or artificial intelligence techniques like NLP, for a more comprehensive view of fall incidents in nursing homes. Supplementary Material 1 NIHMS2156201-supplement-1.docx (37.9KB, docx) Funding: This work was supported by grants R01AG062568 (Odden), R24AG064025 (Steinman), K24AG049057 (Steinman), P30AG04481 (Steinman), R33AG086944 (Steinman), R01AG057751 (Lee), R01HL163163 (Dave) from the NIA and VA Health Services Research grant IIR 15-434 (Lee). Footnotes Disclaimer: The contents of this manuscript are solely the responsibility of the authors and do not necessarily represent the official views of the Department of Veterans Affairs or the NIA. We have no conflicts of interest to report. Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. References 1. Hester Wei F. Falls in the community: state of the science. CIA. Published online June 2013:675. doi: 10.2147/CIA.S44996 [ DOI ] [ Google Scholar ] 2. Shao L, Shi Y, Xie XY, Wang Z, Wang ZA, Zhang JE. Incidence and Risk Factors of Falls Among Older People in Nursing Homes: Systematic Review and Meta-Analysis. Journal of the American Medical Directors Association. 2023;24(11):1708–1717. doi: 10.1016/j.jamda.2023.06.002 [ DOI ] [ PubMed ] [ Google Scholar ] 3. Kimura M, Ruller S, Frank C, et al. Incidence Morbidity and Mortality From Falls in Skilled Nursing Facilities: A Systematic Review and Meta-Analysis. Journal of the American Medical Directors Association. 2023;24(11):1690–1699.e6. doi: 10.1016/j.jamda.2023.07.012 [ DOI ] [ PubMed ] [ Google Scholar ] 4. Mintz J, Lee A, Gold M, Hecker EJ, Colón-Emeric C, Berry SD. Validation of the Minimum Data Set items on falls and injury in two long-stay facilities. J Am Geriatr Soc. 2021;69(4):1099–1100. doi: 10.1111/jgs.16974 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Sanghavi P, Chen Z. Underreporting of Quality Measures and Associated Facility Characteristics and Racial Disparities in US Nursing Home Ratings. JAMA Network Open. 2023;6(5):e2314822. doi: 10.1001/jamanetworkopen.2023.14822 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Sanghavi P, Pan S, Caudry D. Assessment of nursing home reporting of major injury falls for quality measurement on nursing home compare. Health Services Research. 2020;55(2):201–210. doi: 10.1111/1475-6773.13247 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Chen Z, Gleason LJ, Sanghavi P. Accuracy of Pressure Ulcer Events in US Nursing Home Ratings. Medical Care. 2022;60(10):775. doi: 10.1097/MLR.0000000000001763 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Minimum Data Set (MDS) 3.0 for Nursing Homes and Swing Bed Providers ∣ CMS. Centers for Medicare & Medicaid Services. Accessed April 23, 2025. https://www.cms.gov/medicare/quality/nursing-home-improvement/minimum-data-sets-swing-bed-providers [ Google Scholar ] 9. Straker JK, Bailer AJ. A review and characterization of the MDS process in nursing homes. J Gerontol Nurs. 2008;34(10):36–44. doi: 10.3928/00989134-20081001-01 [ DOI ] [ Google Scholar ] 10. Sanghavi P, Pan S, Caudry D. Assessment of nursing home reporting of major injury falls for quality measurement on nursing home compare. Health Services Research. 2020;55(2):201–210. doi: 10.1111/1475-6773.13247 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Shiner B, Levis M, Dufort VM, et al. Improvements to PTSD quality metrics with natural language processing. Journal of Evaluation in Clinical Practice. 2022;28(4):520–530. doi: 10.1111/jep.13587 [ DOI ] [ PubMed ] [ Google Scholar ] 12. Garvin JH, Kim Y, Gobbel GT, et al. Automating Quality Measures for Heart Failure Using Natural Language Processing: A Descriptive Study in the Department of Veterans Affairs. JMIR Medical Informatics. 2018;6(1):e9150. doi: 10.2196/medinform.9150 [ DOI ] [ Google Scholar ] 13. Juhn Y, Liu H. Artificial intelligence approaches using natural language processing to advance EHR-based clinical research. Journal of Allergy and Clinical Immunology. 2020;145(2):463–469. doi: 10.1016/j.jaci.2019.12.897 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Trinh VQN, Zhang S, Kovoor J, et al. The use of natural language processing in detecting and predicting falls within the healthcare setting: a systematic review. Int J Qual Health Care. 2023;35(4):mzad077. doi: 10.1093/intqhc/mzad077 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Minimum Data Set (MDS) 3.0 Resident Assessment Instrument (RAI) Manual ∣ CMS. Accessed June 23, 2025. https://www.cms.gov/medicare/quality/nursing-home-improvement/resident-assessment-instrument-manual [ Google Scholar ] 16. Mintz J, Lee A, Gold M, Hecker EJ, Colón-Emeric C, Berry SD. Validation of the Minimum Data Set Items on Falls and Injury in Two Long-Stay Facilities. J Am Geriatr Soc. 2021;69(4):1099–1100. doi: 10.1111/jgs.16974 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Taseh A, Sasanfar S, Chan JZM, et al. Natural Language Processing Algorithms Outperform ICD Codes in the Development of Fall Injuries Registry. Published online September 27, 2024:2024.09.26.24314444. doi: 10.1101/2024.09.26.24314444 [ DOI ] [ Google Scholar ] 18. American Health Care Association. State Of The Sector: Nursing Home Labor Staffing Shortages Persist Despite Unprecedented Efforts To Attract More Staff. March 5, 2024. Accessed August 18, 2025. https://www.ahcancal.org/News-and-Communications/Press-Releases/Pages/State-Of-The-Sector-Nursing-Home-Staffing-Shortages-Persist-Despite-Unprecedented-Efforts-To-Attract-More-Staff-.aspx [ Google Scholar ] 19. Deandrea S, Bravi F, Turati F, Lucenteforte E, La Vecchia C, Negri E. Risk factors for falls in older people in nursing homes and hospitals. A systematic review and meta-analysis. Archives of Gerontology and Geriatrics. 2013;56(3):407–415. doi: 10.1016/j.archger.2012.12.006 [ DOI ] [ PubMed ] [ Google Scholar ] 20. Pernes M, Agostinho I, Bernardes RA, Belo Fernandes J, Baixinho CL. Documenting fall episodes: a scoping review. Front Public Health. 2023;11. doi: 10.3389/fpubh.2023.1067243 [ DOI ] [ Google Scholar ] 21. Cameron ID, Dyer SM, Panagoda CE, et al. Interventions for preventing falls in older people in care facilities and hospitals - Cameron, ID - 2018 ∣ Cochrane Library. Accessed November 14, 2024. https://www.cochranelibrary.com/cdsr/doi/10.1002/14651858.CD005465.pub4/full [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials 1 NIHMS2156201-supplement-1.docx (37.9KB, docx) ACTIONS View on publisher site PDF (346.9 KB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top