Predicting Falls Using Electronic Health Records: A Time Series Approach

doi:10.21203/rs.3.rs-4712618/v1

Download PDF

Research Article

Predicting Falls Using Electronic Health Records: A Time Series Approach

https://doi.org/10.21203/rs.3.rs-4712618/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background and Aims:

Inpatient falls are a major cause of injury within hospitals and are known to delay recovery and increase patient length of stay. The Morse Fall Scale is a commonly utilized tool to assess fall-risk and is broadly implemented, particularly within Veterans Health Administration. Yet, this scale has limited accuracy. A more precise risk assessment tool is urgently needed to identify those at risk and to ensure targeted fall-risk mitigating interventions. The goal of this work was to develop a more accurate fall prediction model within the Veterans Health Administration.

Methods

The cohort included Veterans admitted to a Veterans Health Administration acute care setting from July 1st, 2020, to June 30th, 2022, with a length of stay between one and seven days. Demographic and clinical data were obtained through VHA electronic health records. Veterans were identified as having a documented fall through clinical progress notes. A transformer model was used to obtain features of this data, which was then used to train a Light Gradient-Boosting Machine for classification and prediction. Area under the precision-recall curve assisted in model tuning, with geometric mean used to define an optimal classification threshold.

Results

Among 242,844 Veterans assessed, 5,965 (2.5%) experienced a fall during their acute inpatient stay. Employing a transformer model with a Light Gradient-Boosting Machine resulted in an area under the curve of 0.851 and an area under the precision-recall curve of 0.285. With an accuracy of 76.3%, the model resulted in a specificity of 76.2% and a sensitivity of 77.3%.

Conclusion

The classification model exhibited a performance greater than the Morse Fall Scale and other risk-assessment tools for predicting risk of fall. Incorporating this type of risk model, which can be automatically calculated from existing data, could provide more efficient and accurate method for identifying at-risk patients.

Fall Risk Assessment

Fall Prediction

Electronic Health Records

Time Series Classification

Prevalent in approximately 2% of all hospital stays, falls are a major cause of injury within hospital care settings.[1] Estimates suggest that up to 50% of all hospital falls result in injury,[1–3] where close to 10% end in serious harm (e.g., fractures, death).[4]

Fall prevention programs are widely implemented to assess and reduce fall-risk through targeted intervention (e.g., increased staffing, the placement of non-skid floor mats, bed alarms, etc.).[5] To assess one’s fall risk, the Morse Fall Scale (MFS) has become common and a widely utilized fall-risk assessment tool.[6] Since 1999, this tool has been consistently administered and documented across all Veterans Health Administration (VHA) inpatient settings.[2, 5] Despite its wide use, and being designed specifically to evaluate fall-risk among hospitalized patients, the MFS has limited accuracy in assessing fall-risk.[7–10]

There are several intrinsic limitations of the MFS that may contribute to its inability to accurately evaluate fall-risk. Importantly, fall risk is multifaceted, and a patient's specific risk may not be optimally represented in the MFS. For example, the MFS does not include demographic variables, which are important factors that influence risk of falls.[2, 11, 12] The MFS also does not account for changeable conditions related to fall-risk, such as changes in patient medications and clinical status.[8, 9] For instance, a surgical patient might have a higher fall risk after initially emerging from anesthesia then that same patient would have 24 hours later. Although alternative fall-risk assessment tools have been suggested,[8, 9, 13–15] they are still limited in their ability to adapt to changing patient conditions, particularly during a complex inpatient stay.

Prior research has expanded upon historical risk-assessment tools by developing models which incorporate electronic health record (EHR) data and employ more advanced machine learning methods.[16–22] Speiser et al. explored the utilization of random forest and decision trees to predict serious injury resulting from falls, but not risk of fall.[22] In addition, Kyu-Nam Heo et al. utilized hospital claims data to train a CatBoost model and showed how it may be used to develop accurate and precise fall prediction models.[21] Although an improvement upon traditional risk models, these studies, and others,[16–18] neglected the nuanced changes in a patients’ presentation over time. Understanding clinical trajectories and incorporating temporal mechanisms into these prediction models is needed for improved risk-assessment acuity and precision that accounts for the dynamic nature of a hospital course.[23]

Accurate prediction of fall risk is central for the appropriate implementation of prevention measures. This is particularly important for the VHA population who are at increased risk due to statistically older age and increased incidence of comorbidities.[24, 25]

1.1 Aim

The intention of this assessment was to leverage EHR data to develop a more accurate fall prediction model intended to enhance the evaluation of falls within VHA acute care settings, and therefore improve care.

2.1 Data Source and Patient Cohort

National EHR data from across the VHA was accessed through DataBricks Data Lakehouse Architecture within the Veterans Affairs (VA) enterprise cloud.[26] Patients were identified who were over the ages of 18 and were admitted to a VHA acute care setting between July 1st, 2020, and June 30th, 2022. If Veterans were admitted multiple times during the assessment period, their most recent stay was utilized. Additionally, only those Veterans who were admitted for at least one day but no more than seven days were included in the assessment. Lastly, Veterans with complete missingness of any variable were excluded from this analysis. Although variables could be padded, as discussed below, at least one observational unit per variable was needed (Supplemental Fig. 1).

2.2 Fall Definition

The VHA National Center for Patient Safety defines a fall as a loss of upright position that results in the patient landing on the floor, or on an object.[5] This definition is utilized to identify and document falls within the VHA using specific “Fall” progress notes.[5] These fall progress notes were used to determine whether patients experienced a fall during their inpatient stay, based upon the notes titles and definitions. Supplemental Table 1 outlines the inclusion/exclusion criteria for identifying falls through EHR progress notes.

2.3 Data & Formatting

Data included patient demographics (i.e., age, sex, race, ethnicity, disability ratings[27]), vital signs (i.e., pulse, respiration, pain rating, temperature, blood pressure), past history of fall, body mass index (BMI), admission source (i.e., where patient was admitted from), inpatient ward specialty, administered mobility assistance devices, acute and chronic diagnosis codes,[28] surgical procedures, medications, lab abnormalities, and Charlson Comorbidity Index.[29] Additionally, the Care Assessment Need (CAN), a VA-specific mortality risk assessment tool, [30] score was included. In total, 320 variables were extracted from the EHR and used for this assessment.

Data were reformatted into a multi-index time series data frame to facilitate analysis.[31] As a hierarchical data format, this enabled the variables to be nested under each patient with respect to time. Changes in a patient’s clinical presentation (e.g., prescription of medications, changes in vitals) can then be evaluated. Regardless of each Veterans’ inpatient stay duration, all data were padded for seven days of observation, with patient intake defined as day one. Padding refers to adding additional data points to missing data. Any missing daily data for time-varying features, such as laboratory results, were padded via forward- and backward-fill. These methods were required to ensure equal length time series data.

2.4 Model

Time series classification, through the utilization of RandOm Convolutional KErnel Transform (ROCKET) model, was used to predict the likelihood of falls among the sample population. As a transformer model, ROCKET leverages a large number of random convolutional kernels to process multivariate time series data into feature maps.[32] The maximum values of the feature maps, similar to global max pooling, and proportion of positive values (PPV), are then generated as aggregate features. These features were then used to train a Light Gradient Boosted Machine (LGBM)[33] to predict the likelihood of a patient falling. LGBM is a gradient boosted framework leveraging tree-based learning algorithms known for its efficiency and accuracy with large data.[34] This approach is designed to capture a wide array of patterns and temporal dependencies within the data, which are not readily identifiable. By transforming the raw time series data into a high-dimensional feature space, ROCKET facilitates the identification of nuanced, predictive signals that are important for accurate fall risk assessment. Figure 1 outlines the data processing and analysis methods.

For model training and evaluation, data were divided using a 70:20:10 split for training, validation, and testing; respectively. Leveraging Optuna v3.4,[35] a random selection of 10,000 training samples were selected to assist in hyperparameter tuning. Area under the receiver operating characteristic curve (AUROC) and precision-recall curve (AUPRC) were used to assist in model tuning and evaluation. Furthermore, the geometric mean was used to determine optimal model thresholds for patient classification. Subgroup analyses were conducted to evaluate model performance among patient demographics through model accuracy and Brier scores. For this subgroup analysis, age was grouped into categorizes. Lastly, probability calibration curves assisted in evaluating model performance. Python v3.10.12 along with pandas v1.4.4[36] was used for data manipulation, with main analysis conducted via scikit-learn v1.1.1 [37] and sktime v0.26.1.[38] This quality improvement and assessment project received a Determination of Non-Research from Stanford IRB #72805 (Stanford University, Stanford, CA, USA).

3.1 Cohort Characteristics

Among 584,536 Veterans admitted to a VHA acute care setting between July 1st, 2020, and June 30th, 2022, a total of 242,844 patients were included in the analysis (Supplemental Fig. 1). The majority of the sample population were male (92.6%), with an average (SD) age of 67.6 (14.3). Most patients self-identified as White (67.9%), followed by Black or African American (22.8%), and identified as not Hispanic or Latino ethnicity (88.5%). Among this sample, 5,965 (2.5%) had a documented fall during their acute care stay (Table 1).

Table 1

Sample Characteristics Among Veterans With a Documented Fall Compared to Those Without Using χ²and Student’s t-tests
	Fall (n = 5965)	No Fall (n = 236879)	Total (n = 242844)	P Value
Sociodemographic characteristics
Sex, n (%)				< 0.001
Male	5644 (94.6)	219166 (92.5)	224810 (92.6)
Female	321 (5.4)	17713 (7.5)	18034 (7.4)
Age, mean (sd)	70.7 (12.2)	67.4 (14.3)	67.6 (14.3)	< 0.001
Race, n (%)				< 0.001
American Indian/Alaskan Native	81 (1.4)	2640 (1.1)	2721 (1.1)
Asian	28 (0.5)	1428 (0.6)	1456 (0.6)
Black or African American	1258 (21.1)	54017 (22.8)	55275 (22.8)
Native Hawaiian or Other Pacific Islander	45 (0.8)	2107 (0.9)	2152 (0.9)
White	4174 (70)	160764 (67.9)	164938 (67.9)
Missing/Unknown	379 (6.4)	15923 (6.7)	16302 (6.7)
Ethnicity, n (%)				< 0.001
Hispanic or Latino	377 (6.3)	17483 (7.4)	17860 (7.4)
Not Hispanic or Latino	5336 (89.5)	209556 (88.5)	214892 (88.5)
Missing/Unknown	252 (4.2)	9840 (4.2)	10092 (4.2)

3.2 Demographic Fall Risk

Significant differences in demographic characteristics were present when comparing those with and without a documented fall (Table 1). Veterans with a documented fall were older (mean (SD): 70.7 (12.2) vs 67.5 (14.3), p < 0.001) and represent a greater proportion of males (94.6% vs 92.5%, p < 0.001). Black or African Americans represented a greater proportion of the no fall group (22.8%) compared to those with a documented fall (21.1%). Conversely, White Veterans represented a greater proportion of the fall group (70%) compared to those without a documented fall (67.9%). In addition, a greater proportion of those who identified as Not Hispanic or Latino were in the fall group (89.5%) compared to those without a documented fall (88.5%).

3.3 Fall Prediction Model Results

By tuning the model with all 320 variables, a ROCKET transformer with 10,000 random convolutional kernels along with a LGBM trained on 10,000 boosting iterations and a learning rate of 0.025 provided optimal fall prediction performance. A complete list of all tuned parameters is provided within Fig. 1. Overall, the model achieved an AUROC of 0.851 on the test set, along with an AUPRC of 0.285 (Fig. 2).

The geometric mean defined an optimal probability threshold of 8.36 x 10^− 8 for classification. This resulted in a model accuracy of 76.3%, a 76.2% in specificity, and a sensitivity of 77.3%. These results suggest that the majority of Veterans who did not experience a fall were correctly classified (76.2%) while 77.3% of those who fell were correctly classified. Despite a high recall (77.3%), the model exhibited a low precision (7.3%). This suggests that only 7.3% of all Veterans who were classified as having a high likelihood of experiencing a fall had a documented fall. Figure 3 provides a confusion matrix detailing the accuracy of the model predictions.

In an evaluation of how well the model performed based upon patient demographic categories, Table 2 displays the model accuracies and Brier scores by demographic group. Minimal differences were observed among patient sex. However, the greatest discrepancies in model accuracy were observed among age groups. Model performance appeared to decrease as the patient’s age increased. This is additionally supported by the Brier scores, in which a lower score suggests greater model performance. Among race, greatest model accuracies were observed among Native Hawaiian and Other Pacific Islander patients (79.1%) while model performance was worse for American Indians/Alaskan Native patients (74.3%). The lowest Brier score was observed when assessed among Asian Veterans (0.017). Interestingly, those Veterans with missing or unknown ethnicities exhibited greatest model accuracies (78.4%) while the model resulted in the lowest Brier score among Hispanic or Latino Veterans (0.019).

Table 2

Subgroup Analysis Displaying Model Accuracy and Brier Score Among Demographics
	Accuracy	Brier Score
Sociodemographic characteristics
Sex
Male	76.9%	0.023
Female	76.2%	0.02
Age Group
<34	81.2%	0.017
35–54	77.3%	0.011
55–74	76.2%	0.23
75–94	75.6%	0.028
>95	73.2%	0.036
Race
American Indian/Alaskan Native	74.3%	0.029
Asian	77.1%	0.017
Black or African American	76.6%	0.023
Native Hawaiian or Other Pacific Islander	79.1%	0.022
White	76.1%	0.023
Missing/Unknown	77%	0.027
Ethnicity
Hispanic or Latino	75.1%	0.019
Not Hispanic or Latino	76.3%	0.023
Missing/Unknown	78.4%	0.023

To better understand how well the model performed overall, Fig. 4 details the probability calibration curve for the reserved test dataset. Comparing the models’ predicted probability against the true probability, we see how the model may not be well-calibrated on this dataset, as suggested by the ill-fitting lines. At predicted probability less than 0.5 and above 0.8, the model appears to underpredict, while it overpredicts at predicted probabilities between 0.5 and 0.8.

Prior work evaluating the efficacy of the MFS has identified key limitations in its ability to accurately assess fall-risk.[7–9] Although alternative fall-risk assessment tools have been proposed, these tools often do not consider changes in a patients’ clinical presentation nor leverage the robust data available within EHR.

In this assessment, we utilized existing EHR data to develop a time series classification model with the goal to more accurately predict the likelihood of falls within VHA acute care settings. Within our finalized model, we obtained an AUROC of 85.1% and an AUPRC of 28.5%. Although an AUROC of 85.1% is considered good, an AUPRC of 28.5% is low.[39] However, considering the severe class imbalance, in which 2.5% of Veterans had a documented fall, an AUPRC of 28.5% is substantially better than identifying those at risk by chance alone. Based upon the optimal geometric mean, we calculated a specificity of 76.2% and a sensitivity of 77.3%. With this, our model outperformed other studies which utilized the MFS. Although the MFS was originally reported to have a sensitivity of 78 and a specificity of 83,[6] recent studies have shown sensitivities as low as 36.9 and specificities as low as 48.3.[8–10] As such, our model has outperformed recent evaluations of the MFS.

Compared to the performance of other tools, our model appears to be an improvement. For example, Nassar et al., evaluated the performance of the Hendrich II Fall Risk Model and obtained a sensitivity of 55.3% and a specificity of 89.3%.[9] In addition, for the evaluation of the John Hopkins Fall Risk Assessment Tool, Kim et al. calculated a sensitivity of 67.8% and a specificity of 80.2%.[8] The greater specificity at the cost of sensitivity for these alternative models are important when considering how to optimally balance care goals and available resources.

Understanding clinical trajectories and disease progressions are known to be important for forecasting patient outcomes,[40, 41] a feature prior fall prediction work has not included. With this knowledge, we incorporated temporal events to assist the model in understanding the time-dependent latent structures and feature representations within the data. This enabled the model to evaluate the magnitude and sequence of events that may impact clinical outcomes.[42] For example, understanding that a medication was prescribed directly following a surgical procedure may provide greater information than a documentation of a patient’s past surgical history and prescription medications.

Our work has revealed and quantified several important demographic risk factors for falls in our patient population. As expected, older patients were at increased risk for fall as were men. Furthermore, we noted race and ethnicity differences in fall risk with greater variability. These types of results emphasized the sometimes subtle and nuanced nature of fall risk. Furthermore, we also observed notable differences when assessing the model's performance for different demographic groups. This is an important reminder that the optimal use of any model may require it to be adjusted or configured for an individual patient.

There are several notable strengths of our work. We utilized existing historical EHR data from a large national cohort. This allowed for prompt analysis representative of our specific VHA population. This is particularly important as creating and training a model on the population that it will be used for, is expected to result in more appropriate and relevant results.[43] Furthermore, building a fall prediction model that utilizes existing and available data will allow for the creation of automated and efficient clinical decision support tool that can empower staff to redirect the time saved to other tasks. In addition, by evaluating falls that occurred within acute care settings only, in which reporting falls is required,[5] we optimized the reliability of the falls data.[2, 44–46]

There are also several limitations, such that this assessment was performed for our specific Veteran population, that are statically older and have a greater likelihood for comorbidities, as well as specific conditions that influence results and limit generalizability. In addition, because 2.5% of patients had a documented fall during their inpatient stay there was a class imbalance for the model. In general, severe class imbalance such as this limits the ability to predict patient outcomes accurately and precisely. This is further emphasized by the ill-fitting probability calibration curve and the low probability threshold as defined by the geometric mean. While the prevalence of falls reflects those documented elsewhere,[47, 48] further work is needed to understand the nuance differences among these two cohorts to enhance model performance. Exploring alternative data sources or extracting more granular data may assist in these efforts. In addition, the performance of this model was measured against historical data. Therefore, although the results are promising for improved care, the performance may not reflect the results from clinical use. Therefore, a more complete understanding of the effectiveness of any risk assessment tool requires rigorous prospective assessment.

As noted, there remains substantial and robust data available within unstructured clinical notes that was not utilized in this model. Nonetheless, the additional data may further improve fall risk identification.[19, 49, 50] Therefore, future work should examine how to leverage emerging tools such as large language models, that can extract insights from unstructured text, will improve performance.

Furthermore, the methods employed within this assessment do not allow for model transparency. Although alternative classification methods may allow us to examine feature importance (e.g., decision trees, logistic regression), these tools often lack the robustness time series methods and transformer models provide. Future work is needed to enhance the interpretability of time series models and identifying feature importance.

By leveraging existing EHR data, we developed a time series classification model to evaluate the likelihood of falls. The performance of the model was an improvement over the commonly used Morse Fall Scale. Incorporating these risk assessments within the electronic health records can alert clinical support staff of at-risk individuals. By providing prompt notification, timely fall-prevention interventions may be implemented, mitigating fall risk.

MFS: Morse Fall Scale

VHA: Veterans Health Administration

EHR: Electronic health record

VA: Veterans Affairs

BMI: Body mass index

CAN: Care Assessment Need

ROCKET: RandOm Convolutional KErnel Transform

PPV: Proportion of positive values

LGBM: Light Gradient Boosted Machine

AUROC: Area under receiver operating curve

AUPRC: Area under precision-recall curve

SD: Standard deviation

Ethics approval and consent to participate: This is a quality improvement and assessment project received a Determination of Non-Research from Stanford IRB #72805 (Stanford University, Stanford, CA, USA). The IRB determined that this project was conducted to assess, analyze, critique, and improve current processes in an institutional setting, involving data-guided activities designed to bring about prompt improvements. Therefore, it does not meet the definition of research nor clinical investigation, and thus participant consent is deemed unnecessary according to national regulations (45 CFR part 46). In addition, no specific human subjects were contacted during this work.

Consent for publication: Not applicable.

Availability of data and materials: Due to US Department of Veterans Affairs (VA) regulations and ethics agreements, the data utilized for this assessment are not permitted to leave the VA firewall without a Data Use Agreement. However, VA data can be made available to researchers with an approved IRB and VA authorized study protocol. For more information, please visit https:// www.virec.research.va.gov or contact the VA Information Resource Center at [email protected].

Competing interests: The authors declare that they have no competing interests.

Funding: None.

Authors' contributions: PH: Data curation, Formal analysis, Methodology, Visualization, Writing - original draft, Writing - review & editing; TB: Data curation, Methodology, Validation, Writing - review & editing; AW: Methodology, Validation, Writing - review & editing; MP: Methodology, Writing - review & editing; ZV: Conceptualization, Methodology, Project administration, Resources, Supervision, Validation; CC: Investigation, Writing - review & editing; TO: Conceptualization, Investigation, Project administration, Resources, Supervision, Writing - review & editing. All authors read and approved the final manuscript.

Acknowledgements: None.

Authors note: Zachary P Veigulis & Thomas F. Osborne participated in the assessment during his employment with the U.S. Department of Veterans Affairs, but he is no longer affiliated with the VA.

Bouldin ELD, Andresen EM, Dunton NE, et al. Falls among adult patients hospitalized in the United States: prevalence and trends. J Patient Saf 2013; 9: 13–17.
Quigley PA, Palacios P, Spehar AM. Veterans’ fall risk profile: a prevalence study. Clin Interv Aging 2006; 1: 169.
Young-Xu Y, Soncrant C, Neily J, et al. Falls in Veterans Healthcare Administration Hospitals: Prevalence and Trends. J Healthc Qual 2020; 42: 113–121.
Oliver D, Healey F, Haines TP. Preventing falls and fall-related injuries in hospitals. Clin Geriatr Med 2010; 26: 645–692.
U.S. Department of Veterans Affairs. Falls Policy, https://www.patientsafety.va.gov/docs/fallsToolkit/05_fallspolicy.pdf (2004, accessed 30 August 2023).
Morse JM, Morse RM, Tylko SJ. Development of a Scale to Identify the Fall-Prone Patient. Can J Aging 1989; 8: 366–377.
O’Connell B, Myers H. The sensitivity and specificity of the Morse Fall Scale in an acute care setting. J Clin Nurs 2002; 11: 134–136.
Kim YJ, Choi KO, Cho SH, et al. Validity of the Morse Fall Scale and the Johns Hopkins Fall Risk Assessment Tool for fall risk assessment in an acute care setting. J Clin Nurs 2022; 31: 3584–3594.
Nassar N, Helou N, Madi C. Predicting falls using two instruments (the Hendrich Fall Risk Model and the Morse Fall Scale) in an acute care setting in Lebanon. J Clin Nurs 2014; 23: 1620–1629.
Kim EAN, Mordiffi SZ, Bee WH, et al. Evaluation of three fall-risk assessment tools in an acute care setting. J Adv Nurs 2007; 60: 427–435.
Ghahramani M, Naghdy F, Stirling D, et al. Fall Risk Assessment in Older People, www.theijes.com (accessed 30 August 2023).
Jewell VD, Capistran K, Flecky K, et al. Prediction of Falls in Acute Care Using The Morse Fall Risk Scale. https://doi.org/101080/0738057720201815928 2020; 34: 307–319.
Oliver D, Britton M, Seed P, et al. Development and evaluation of evidence based risk assessment tool (STRATIFY) to predict which elderly inpatients will fall: case-control and cohort studies. BMJ 1997; 315: 1049–1053.
Hendrich A, Nyhuis A, Kippenbrock T, et al. Hospital falls: development of a predictive model for clinical practice. Appl Nurs Res 1995; 8: 129–139.
Hester AL, Davis DM. Validation of the Hester Davis scale for fall risk assessment in a neurosciences population. Journal of Neuroscience Nursing 2013; 45: 298–305.
Thapa R, Garikipati A, Shokouhi S, et al. Predicting Falls in Long-term Care Facilities: Machine Learning Study. JMIR Aging 2022; 5: e35373.
Patterson BW, Engstrom CJ, Sah V, et al. Training and Interpreting Machine Learning Algorithms to Evaluate Fall Risk After Emergency Department Visits. Med Care 2019; 57: 560–566.
Lindberg DS, Prosperi M, Bjarnadottir RI, et al. Identification of important factors in an inpatient fall risk prediction model to improve the quality of care using EHR and electronic administrative data: A machine-learning approach. Int J Med Inform 2020; 143: 104272.
Nakatani H, Nakao M, Uchiyama H, et al. Predicting Inpatient Falls Using Natural Language Processing of Nursing Records Obtained From Japanese Electronic Medical Records: Case-Control Study. JMIR Med Inform 2020; 8: e16970.
Choi JH, Choi ES, Park D. In-hospital fall prediction using machine learning algorithms and the Morse fall scale in patients with acute stroke: a nested case-control study. BMC Med Inform Decis Mak; 23. Epub ahead of print 1 December 2023. DOI: 10.1186/S12911-023-02330-0.
Heo KN, Seok JY, Ah YM, et al. Development and validation of a machine learning-based fall-related injury risk prediction model using nationwide claims database in Korean community-dwelling older population. BMC Geriatr 2023; 23: 1–10.
Speiser JL, Callahan KE, Houston DK, et al. Machine Learning in Aging: An Example of Developing Prediction Models for Serious Fall Injury in Older Adults. The Journals of Gerontology: Series A 2021; 76: 647–654.
Lee JM, Hauskrecht M. Modeling Multivariate Clinical Event Time-series with Recurrent Temporal Mechanisms. Artif Intell Med 2021; 112: 102021.
Joan Wang AZ, Dhanireddy P, Prince C, et al. Advanced Survey Design, LLC. 2021 Survey of Veteran Enrollees’ Health and Use of Health Care Data Findings Report.
Wong ES, Wang V, Liu CF, et al. Do Veterans Health Administration Enrollees Generalize to Other Populations? Medical Care Research and Review 2016; 73: 493–507.
Data Lakehouse Architecture | Databricks, https://www.databricks.com/product/data-lakehouse (accessed 10 April 2024).
About VA Disability Ratings | Veterans Affairs, https://www.va.gov/disability/about-disability-ratings/ (accessed 15 April 2024).
Chronic Condition Indicator Refined (CCIR) for ICD-10-CM, https://hcup-us.ahrq.gov/toolssoftware/chronic_icd10/chronic_icd10.jsp (accessed 26 February 2024).
Charlson ME, Pompei P, Ales KL, et al. A new method of classifying prognostic comorbidity in longitudinal studies: Development and validation. J Chronic Dis 1987; 40: 373–383.
Fihn S, Mph MD, Box T. Care Assessment Need (CAN) Score and the Patient Care Assessment System (PCAS): Tools for Care Management.
Feng L, Yin Z, Jin B, et al. Time Series Subsequence Retrieval Based on the Properties of Multi-index. International Journal on Advances in Information Sciences and Service Sciences 2011; 3: 300–308.
Dempster A, Petitjean F, Webb GI. ROCKET: exceptionally fast and accurate time series classification using random convolutional kernels. Data Min Knowl Discov 2020; 34: 1454–1495.
Ke G, Meng Q, Finley T, et al. LightGBM: A Highly Efficient Gradient Boosting Decision Tree, https://github.com/Microsoft/LightGBM. (accessed 12 February 2024).
McCarty DA, Kim HW, Lee HK. Evaluation of Light Gradient Boosted Machine Learning Technique in Large Scale Land Use and Land Cover Classification. Environments 2020, Vol 7, Page 84 2020; 7: 84.
Akiba T, Sano S, Yanase T, et al. Optuna: A Next-generation Hyperparameter Optimization Framework. Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining 2019; 2623–2631.
pandas - Python Data Analysis Library, https://pandas.pydata.org/ (accessed 25 February 2024).
Pedregosa FABIANPEDREGOSA F, Michel V, Grisel OLIVIERGRISEL O, et al. Scikit-learn: Machine Learning in Python Gaël Varoquaux Bertrand Thirion Vincent Dubourg Alexandre Passos PEDREGOSA, VAROQUAUX, GRAMFORT ET AL. Matthieu Perrot. Journal of Machine Learning Research 2011; 12: 2825–2830.
Löning M, Bagnall A, Ganesh S, et al. sktime: A Unified Interface for Machine Learning with Time Series, https://github.com/alan-turing-institute/sktime (accessed 20 February 2024).
Saito T, Rehmsmeier M. The Precision-Recall Plot Is More Informative than the ROC Plot When Evaluating Binary Classifiers on Imbalanced Datasets. PLoS One; 10. Epub ahead of print 4 March 2015. DOI: 10.1371/JOURNAL.PONE.0118432.
Verstockt B, Parkes M, Lee JC. How Do We Predict a Patient’s Disease Course and Whether They Will Respond to Specific Treatments? Gastroenterology 2022; 162: 1383–1395.
Sun C, Hong S, Song M, et al. A REVIEW OF DEEP LEARNING METHODS FOR IRREGULARLY SAMPLED MEDICAL TIME SERIES DATA.
Lee JM, Hauskrecht M. Neural Clinical Event Sequence Prediction through Personalized Online Adaptive Learning. Artif Intell Med Conf Artif Intell Med (2005-) 2021; 12721: 175.
Dexter GP, Grannis SJ, Dixon BE, et al. Generalization of Machine Learning Approaches to Identify Notifiable Conditions from a Statewide Health Information Exchange. AMIA Summits on Translational Science Proceedings 2020; 2020: 152.
Hale WA, Delaney MJ, Cable T. Accuracy Of Patient Recall And Chart Documentation Of Falls. DOI: 10.3122/jabfm.6.3.239.
Hoffman GJ, Ha J, Alexander NB, et al. Underreporting of Fall Injuries by Older Adults: Implications for Wellness Visit Fall Risk Screening. J Am Geriatr Soc 2018; 66: 1195.
Ganz DA, Higashi T, Rubenstein LZ. Monitoring Falls in Cohort Studies of Community-Dwelling Older People: Effect of the Recall Interval. J Am Geriatr Soc 2005; 53: 2190–2194.
LeLaurin JH, Shorr RI. Preventing Falls in Hospitalized Patients: State of the Science. Clin Geriatr Med 2019; 35: 273.
Neily J, Quigley PA, Essen K. Implementation Guide for Fall Injury Reduction VA National Center for Patient Safety Reducing Preventable Falls and Fall-Related Injuries VA National Center for Patient Safety Reducing Preventable Falls and Fall-Related Injuries associate director, VA National Center for Patient Safety White River Junction Field Office VA National Center for Patient Safety Reducing Preventable Falls and Fall-Related Injuries.
Kawazoe Y, Shimamoto K, Shibata D, et al. Impact of a Clinical Text-Based Fall Prediction Model on Preventing Extended Hospital Stays for Elderly Inpatients: Model Development and Performance Evaluation. JMIR Med Inform; 10. Epub ahead of print 1 July 2022. DOI: 10.2196/37913.
Fu S, Thorsteinsdottir B, Zhang X, et al. A hybrid model to identify fall occurrence from electronic health records. Int J Med Inform; 162. Epub ahead of print 1 June 2022. DOI: 10.1016/J.IJMEDINF.2022.104736.

No competing interests reported.

SupplementalFigure1.docx
Additional Files/Supplementals: Supplemental Figure 1: File Name: Supplemental_Figure_1.docx Title: Supplemental Figure 1: Flowchart of Patient Cohort Description: Criteria for Patient Cohort Selection. VHA: Veterans Health Administration; CAN: Care Assessment Need
SupplementalTable1.docx
Supplemental Table 2: File Name: Supplemental_Table_2.docx Title: Supplemental Table 1: Progress Note Titles from the Veteran’s Health Administration’s Computerized Patient Record System Used to Define Incidence of Falls. Description: Inclusion/exclusion criteria for identifying falls through EHR progress notes. HBPC: Home Based Primary Care; CBOC: Community Based Outpatient Clinics

Download PDF

Editorial decision: Revision requested
12 Aug, 2024
Reviews received at journal
11 Aug, 2024
Reviews received at journal
06 Aug, 2024
Reviewers agreed at journal
02 Aug, 2024
Reviewers agreed at journal
01 Aug, 2024
Reviewers invited by journal
01 Aug, 2024
Editor invited by journal
31 Jul, 2024
Editor assigned by journal
24 Jul, 2024
Submission checks completed at journal
24 Jul, 2024
First submitted to journal
09 Jul, 2024

You are reading this latest preprint version

Predicting Falls Using Electronic Health Records: A Time Series Approach

Status:

Version 1

Abstract

Background and Aims:

Methods

Results

Conclusion

Figures

1 Background

1.1 Aim

2 Methods

2.1 Data Source and Patient Cohort

2.2 Fall Definition

2.3 Data & Formatting

2.4 Model

3 Results

3.1 Cohort Characteristics

3.2 Demographic Fall Risk

3.3 Fall Prediction Model Results

4 Discussion

5 Conclusion

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1