Development of a Berg Balance Scale Short-Form Using a Machine Learning Approach in Patients With Stroke : Journal of Neurologic Physical Therapy

Secondary Logo

Journal Logo

Research Articles

Development of a Berg Balance Scale Short-Form Using a Machine Learning Approach in Patients With Stroke

Wang, Inga PhD, OTR/L; Li, Pei-Chi MS; Lee, Shih-Chieh PhD; Lee, Ya-Chen PhD; Wang, Chun-Hou BS; Hsieh, Ching-Lin PhD

Author Information
Journal of Neurologic Physical Therapy 47(1):p 44-51, January 2023. | DOI: 10.1097/NPT.0000000000000417


Decreased balance is a common problem associated with people following stroke.1 Poor dynamic balance may hinder individuals from independent walking, one of the critical factors to achieve functional independence in activities of daily living.2 Further adverse consequences may include fear of falling,3,4 subsequent activity avoidance, social isolation, physical and psychological deterioration, or traumatic injuries.5 To improve balance and functional independence, clinicians need a psychometrically sound and efficient balance measure.

The Berg Balance Scale (BBS)6,7 is a widely used clinical performance measure of functional balance. Although the BBS was initially developed to measure balance in older adults, it is now commonly used to measure balance in people with varying conditions and disabilities. To administer the BBS, the participant is instructed to perform a series of predetermined motor tasks and safely balance while attempting the tasks. Each item consists of a 5-point ordinal scale ranging from 0 (unable, needs assistance, or loses balance) to 4 (able). The final measure is a total score between 0 and 56, with a higher score indicating better balance.

The psychometric properties (including reliability,8–12 validity,9,10,12–26 and responsiveness17,26–30) of the BBS are well supported. In the past decade, researchers in health care–related fields have been developing short forms of widely used clinical settings and medical research measures (eg, PROMIS,31 Mini-Mental Status Examination).32 Several plausible reasons stimulate this growth. First, in the primary care setting where schedules are busy, shorter assessments save time and reduce administrative burden. Second, in a rigorous clinical trial, shorter assessments allow researchers to measure multiple health outcomes. Third, short forms have been reported to strongly correlate with the full-length total scores31,33,34 and be as accurate as the original assessments.32–34

Various methods have been used to develop short forms. One of the plausible methods is item response theory ([IRT] or Rasch analysis).35 Notably, IRT has the advantage of invariance of the item parameters, and the standard errors in IRT are calculated separately for each level of the latent trait. However, the resulting scores of the IRT-based short forms are standardized scores, which are different from those of the original measures (raw sum scores). The 2 types of scores are incomparable, which hampers the interchangeability between the short form and the original one. Thus, the utility of an IRT-based short form is limited.

Machine learning (ML) has shown potential to shorten the number of items of a measure. Two advantages of ML-based short-form measures have been demonstrated. First, an ML-based short-form measure can provide scores comparable with those of the original one.36,37 Such an ML-based shortened measure can be an alternative to the original measure to improve the efficiency of assessments. Second, the administrative methods (eg, performance rating for the BBS) and score interpretation of ML-based short-form measures are identical to those of the original measures. Therefore, users can easily adapt to the ML-based short-form measures. Both aforementioned advantages make ML a promising approach to generating efficient short-form measures. Thus, this study aimed to develop a short form of the BBS using an ML approach (BBS-ML).


Study Design

This was a secondary data analysis study. We used data from the Locomotor Experience Applied Post-stroke (LEAPS)38 trial to develop the BBS-ML short-form and externally validate the results using an independent data set.

Data Source

BBS Data From the LEAPS Trial

The objective of the LEAPS trial was to examine the effectiveness of a specialized locomotor training program, conducted at 2 or 6 months poststroke, as compared with a home-based, nonspecific, low-intensity exercise intervention provided at 2 months poststroke. The training protocol included 36 outpatient sessions of locomotor training, 3 times per week, using a body weight support system on a treadmill. The study began in April 2006 and was concluded in April 2010. Participants were included if they were at least 18 years old, had had a stroke within 45 days, had Fugl-Meyer lower extremity motor scores of less than 34, could sit unsupported for 30 seconds, were able to walk at least 10 ft with a maximum of one person assisting, could follow a 3-step command, were able to provide informed consent, had a self-selected 10-m gait speed of less than 0.8 m/s, and had successfully completed an exercise tolerance test.

The 14-item BBS was administered at baseline (2 months poststroke), 6 months poststroke, and 12 months poststroke. In this study, we used only the baseline data. Because this study involved secondary analysis of de-identified data, the institutional review board (IRB) of the University of Wisconsin-Milwaukee determined that this study did not fall within the regulatory definition of research involving human subjects and did not require further IRB review. Note that the researchers in the LEAPS trial slightly modified the BBS to fit their needs.38 Specifically, they replaced “placing the alternate foot on step or stool while standing unsupported” with “dynamic weight shifting while standing unsupported.” Therefore, we decided not to select this item for the BBS-ML to ensure comparability with the original BBS. But its score was used for calculating the sum of the observed BBS scores to obtain the commonly used total score (ie, 0-56).

External Validation

We used data collected in the BBS-3P study to further validate the results. The BBS-3P research study was designed to compare the psychometric properties of 3 commonly used measures assessing poststroke balance function. The study began in December 1999 and was concluded in May 2000. The participants were included if they had a diagnosis of stroke within 14 days, had the first onset with no other major disease, were able to follow commands, and could provide informed consent by themselves or by proxy. The BBS was administered at 14 days and 90 days poststroke. In this study, only the data obtained at 14 days poststroke were used.

Development of the BBS-ML Short-Form

The BBS-ML was developed in 4 steps. First, we applied feature selection algorithms in the artificial neural network (ANN) model to identify candidate items with high predictive power. Second, top-ranked items were used to construct initial short-form (ie, 4-, 5-, 6-, 7-, and 8-item) versions based on the feature selection. The initial short form was decided by half of the items from the original version. To confirm that the 7-item version is capable of training an ML model with sufficient predictive power, we also include an 8-item version to test the hypotheses we set. Third, we compared the performance of the short-form versions and used criteria to select the final version of the BBS-ML. Finally, an independent sample was used to further validate the BBS-ML.

ANN Model and Feature Selection

ANN models were developed using the scikit-learn and Keras package in Python. As part of the ML techniques, neural networks imitate the brain's underlying neural connections with a series of neurons (or nodes) organized in layers like a multilayer perceptron in the central nervous system. Neural networks were used because they offer many advantages, including detecting complex nonlinear relationships between dependent and independent variables implicitly, detecting all possible interactions between predictor variables, and the availability of multiple training algorithms.

Briefly, an ANN model has a set of input variables in the input layer. The next layer is called a hidden layer. Depending on the complexity of the model, there may be 1 or several hidden layers. The final layer is the output layer, which gives the predicted score of the target behavior. Each neuron in one layer is connected to neurons of other layers with associated estimated weights. Activation functions are used to calculate a weighted sum of input and then decide whether the neuron should be activated or not.

In this study, the input attributes of the ANN model included selected observed responses of 14 BBS items. The number of neurons in the hidden layer was set empirically, with 5 hidden layers and 128 neurons. The output layer was the predicted BBS total score, which was compared with the desired output of the observed BBS total score. Figure 1 illustrates the ANN model used for the BBS-ML application.

Figure 1:
The ANN model used for the BBS-ML and an example showing how a patient's responses on the 6-item BBS-ML are transformed into a final score (output layer) as compared with the final score of the BBS (original data). ANN, artificial neural network; BBS, Berg Balance Scale; ML, machine learning.

Split-sample validation was adopted for model training. Specifically, the entire data set (from the LEAPS trial) was divided into training and testing data sets with a 4:1 ratio using the random seed function. The data splitting process was repeated 50 times (ie, 50 runs). Participant responses (ie, scores of the BBS items) were presented to the ANN model one at a time, and the weights associated with the input values were adjusted each time. The iterative learning process continued until the algorithm reached the minimal difference between the predicted output and the desired output (ie, observed BBS total score) no longer decreased.

Construction of Initial Short-Form Versions

Once the model was trained, we applied the feature selection algorithm to select potential candidate items to be included in the short-form versions. The L1 penalty (also known as the lasso penalty) was added to the weights between the ANN model's input and hidden layers to differentiate the importance.

For each run, the top 3 importance items (showing the top 3 weights between the input and hidden layers of the model) were labeled “1” while the rest of the BBS items were labeled “0.” After 50 runs, we summed up the 0s and 1s for each BBS item and ranked items based on their overall importance. The top items, ranging from the top 4 to the top 8, were selected to develop 5 initial short-form versions (ie, 4-, 5-, 6-, 7-, and 8-item short forms).

Evaluation of the Performance of the Initial Short-Form Versions

The predictive power of the short-form versions (optimal weight cutoff) was visualized by the Bland-Altman plot and evaluated using 3 indexes: R2, 95% limit of agreement (LoA), and possible scoring point (PSP). R2 is a goodness-of-fit measure for linear regression models and indicates the percentage of variance in the outcome variable that the predictors explain collectively. The Bland-Altman plot is a graphical method of comparing 2 measurements. The x-axis in the plot is the average of the predicted and observed BBS total scores, while the y-axis is the difference between the predicted and observed BBS total scores (ie, residuals). The 95% LoAs, calculated as 2 × 1.96 standard deviations (SDs) of the residuals, were used to quantify the differences between the 2 sets of scores. The PSPs are the number of unique values generated by the ANN outputs (eg, for a BBS total score of 0-56 points, PSP = 57). The final version of the BBS-ML was chosen by selecting the short form that used a smaller number of items to achieve a higher predictive power R2 (>0.90), a lower 95% LoA (<11.2 [2 × 10% of 56]), and an adequate PSP (>28, half that of the BBS).

External Validation

Finally, we further validated the BBS-ML to examine the generalizability of its performance using data from the BBS-3P sample.


Analytic Sample

The LEAPS trial data included 408 persons with stroke with a mean age of 62 years (SD = 12.7; range, 25-98), 55% of whom were male. About 80% of the sample had ischemic stroke. Approximately half of the sample had brain lesions in the right hemisphere, followed by the left hemisphere (35%), brainstem (15%), and bilateral (2%). Although most of the sample was White (58%), other racial groups were represented: Black (22%), Asian (13%), Native Hawaiian or Other Pacific Islander (5%), and American Indian/Alaska Native (1%). The BBS-3P data used for external validation included 226 persons with stroke from Taiwan with a mean age of 68.0 years (SD = 10.7; range, 32-94), 61% of whom were male. Table 1 presents the demographic information of the sample.

Table 1 - Demographic Characteristics of the Sample From the LEAPS Trial
Data Source
Variable Mean SD Mean SD
Age in years at onset 62.0 (12.7) 68.0 (10.7)
Berg Balance Scale (score 0-56 points) 35.8 (14.0) 23.4 (22.0)
Frequency % Frequency %
Male 224 (55) 137 (61)
Female 184 (45) 89 (39)
Stroke type
Ischemic 327 (80) 167 (74)
Hemorrhagic 76 (19) 59 (26)
Uncertain 5 (1) ...
Stroke lesion
Right hemisphere 197 (48) 120 (53)
Left hemisphere 143 (35) 106 (47)
Brainstem 62 (15) ...
Bilateral 6 (2) ...
Hispanic or Latino 63 (15) ...
Not Hispanic or Latino 345 (85) ...
Abbreviation: SD, standard deviation.

Performance of Initial Short-Form Versions and External Validation

Results of the feature selection suggested that the top 4 useful items were as follows: standing unsupported with feet together, standing on one leg, pick up object from floor, and turn to look behind over left and right shoulders. Hence, these 4 items were included in the 4-item short-form version. The 5-item version included all the previous items and an additional item, transfers. The 6-item version included the previous items and one additional item, standing to sitting. The 7-item version had one added item, reaching forward with outstretched arm, while the 8-item version included an additional item, standing with eyes closed.

In the testing data sets, the average R2s (out of 50 runs) for the 4-, 5-, 6-, 7-, and 8-item short-form versions were high (≥0.91), the 95% LoAs for each short-form version ranged from 9.6 (6-item version) to 15.6 (4-item version), and the PSPs ranged from 32 (4- and 5-item versions) to 36 (9-item version). Table 2 summarizes the details of the results.

Table 2 - Evaluating the Performance of Initial Short-Form Versions: R2, 95% Limits of Agreement, and Possible Scoring Points
Short-Form Versionsa
Data Set 4 Items 5 Items 6 Items 7 Items 8 Items
Training data set R 2 0.96 0.98 0.99 0.99 0.99
LoA 11.1 7.7 6.4 5.5 5.0
PSPb 38 42 46 48 47
Testing data set R 2 0.93 0.95 0.97 0.97 0.97
LoA 1.2 12.2 9.6 9.6 8.9
PSPb 25 35 34 35 36
External validating data set R 2 0.94 0.97 0.99 0.99 0.99
LoA 20.55 13.7 10.6 9.5 10.0
PSPb 22 31 32 37 37
Abbreviations: BBS, Berg Balance Scale; LoA, limits of agreement; ML, machine learning; PSP, possible scoring point.
a4-item BBS-ML includes standing unsupported with feet together, standing on one leg, pick up object from floor, and turn to look behind over left and right shoulders. 5-item BBS-ML includes standing unsupported with feet together, standing on one leg, pick up object from floor, turn to look behind over left and right shoulders, and transfers. 6-item BBS-ML includes standing unsupported with feet together, standing on one leg, pick up object from floor, turn to look behind over left and right shoulders, transfers, and standing to sitting. 7-item BBS-ML includes standing unsupported with feet together, standing on one leg, pick up object from floor, turn to look behind over left and right shoulders, transfers, standing to sitting, and reaching forward with outstretched arm. 8-item BBS-ML includes standing unsupported with feet together, standing on one leg, pick up object from floor, turn to look behind over left and right shoulders, transfers, standing to sitting, reaching forward with outstretched arm, and standing with eyes closed.
bPSP refers to the number of unique scores in the output data.

Figure 2 shows the Bland-Altman plots of the training data sets (left) and the testing data sets (right) for each short-form version. The means of the residuals for the 4-, 5-, 6-, 7-, and 8-item short-form versions with testing data sets were 3.0 (SD = 2.6), 2.4 (SD = 2.1), 2.0 (SD = 1.5), 2.1 (SD = 1.8), and 1.6 (SD = 1.3), respectively.

Figure 2:
Bland-Altman plot: training data set (left) and testing data set (right). This figure is available in color online (

Among these short-form versions, the 6-item version used the smallest number of items to achieve a relatively higher R2 value of 0.97, a lower LoA of 9.6, and an adequate PSP of 35. Therefore, it was chosen as the final BBS-ML.

The results of the external validation are also presented in Table 2. Similar performance (R2 = 0.99, 95% LoA = 10.6, and PSP = 37) were found when the 6-item BBS-ML was examined in the independent sample.


This study aimed to develop a BBS short-form using a ML technique. Because including more than 6 items did not increase the overall performance substantially, the 6-item version was chosen as the final BBS-ML. Overall, the 6-item BBS-ML showed the best combination of evaluation results and excellent agreement with the original BBS, with R2s greater than 0.96, smaller 95% LoAs, and a higher PSP. Preliminary external validation supported the generalizability of its performance in the independent sample. These results support the validity and usage of the 6-item BBS-ML.

The BBS-ML has several advantages over the original BBS in terms of clinical practicality and efficiency. It could allow clinicians or raters to complete the assessment in nearly half of the time required to complete the original BBS. Hence, the short form makes it possible to evaluate patients' progress in rehabilitation settings in a more economical way. In addition, administration of the short form reduces the need to change testing positions, requires fewer assessment tools during the BBS test, and decreases the possibility of incomplete data collection in the examination.39 Therefore, the BBS-ML may provide an alternative to the original BBS for routine assessments to monitor patients' balance function.

The final 6 items in the BBS-ML includes standing unsupported with feet together, standing on one leg, pick up object from floor, turn to look behind over left and right shoulders, transfers, and standing to sitting. Wong et al40 and La Porta et al41 evaluated the item difficulty hierarchy of the BBS using Rasch analysis. Despite the different patient populations (leg amputation40 and Parkinson disease),41 standing on one leg, turn to look behind (turning 360°), and pick up object from floor (retrieving object from floor) were included on the top 5 most challenging items. In contrast, transfers, standing to sitting, and standing unsupported with feet together were the easy items. Thus, the 6 items have wide range of item difficulties, which may support their content representativeness of the original 14 items.

To the best of our knowledge, one BBS short-form was developed by Chou (2006).39 In Chou's study, the best items were determined by selecting the items with the lowest values from an overall item index of each item. The overall item index of each item is the product of the 2 rank orders (ie, the rank order of the corrected item-total correlation for an item and the rank order of the effect size for an item). Their 7-item BBS included reaching forward with outstretched arm, standing with eyes closed, standing with one foot in front, turning to look behind, retrieving object from floor, standing on one foot, and sitting to standing. Our results were overlapping with 3 items, and the sitting to standing is very compatible with standing to sitting.

To improve the accessibility of the BBS-ML, the authors have established an online system for (a) scoring the BBS-ML and (b) converting the score between 6-item BBS-ML and the original 14-item BBS. The ANN is hosted on the server. Therefore, users do not need to install or execute the ANN on their site. In addition, clinicians can use the online system, enter scores on subset BBS items, and obtain the ML-estimated 14-item BBS score. Prospective users can contact the authors for the free use of the online system. As such, similar score interpretations (eg, cutoff score, minimal detectable change, and minimal clinically important difference) apply.

We executed a secondary data analysis of a prospective randomized controlled trial study and a previous study. Secondary analysis of existing data has become an increasingly popular method of enhancing the overall efficiency of the health research enterprise. In accordance with the NIH Data Sharing Policies, lots of data collected by funded projects are deposited in public databases. The goal is to stimulate the use of existing human data sets to conduct additional analyses secondary to a project's primary purpose and to investigate novel scientific ideas or new models, methods, or technologies. The LEAPS project is a multicenter clinical trial with excellent quality control. Our research team believes that the stroke-specific outcome assessment BBS data provide a great resource to develop stroke-relevant short-form measures using a new but promising ML approach.


This study has several limitations. First, the BBS data from the LEAPS trial had a negative skew distribution (skewness = −0.86). The participants had a mean BBS total score of 36 (SD = 14). Fewer participants had lower BBS scores, which may have reduced the ANN's estimation precision at the lower end. The inclusion criteria in the LEAPS trial focused on screening the participants based on lower extremity function, which may have resulted in some degree of selection bias.

Second, we acknowledge that the patient characteristics in the external validation sample were different from those in the LEAPS trial (primarily in race/ethnicity, recovery stages, and BBS scores). The differences might have affected the results of cross-validation.

Third, we developed the short form based on data consisting of people with stroke. Given that the BBS has been widely used in people with various neurologic conditions, our scope seems to be limited because only people with stroke were recruited. Future studies are warranted to examine the BBS-ML in other settings and samples.

Finally, several methods have been proposed to provide evidence and validate the short forms. Pearson correlation coefficients or intraclass correlations between estimated scores of short-form and full-length measures are among those methods used to assess the extent to which short forms capture the information in the full-length scales and compare measurement reliability. An alternative is to evaluate the short form by examining the distribution of the residuals and comparing the psychometric properties, such as the internal consistency, reliability, and convergent validity, of the short form and the original measure. The psychometric properties, clinical usage, and predictive performance of the BBS-ML warrant further research.


The BBS-ML seems to be a promising short-form alternative to improve the efficiency of administration in measuring balance in people with stroke. Future research is needed to examine the psychometric properties and clinical usage of the BBS-ML in various settings and samples.


1. Chen IC, Cheng PT, Hu AL, Liaw MY, Chen LR, Hong WH, et al. Balance evaluation in hemiplegic stroke patients. Chang Gung Med J. 2000;23(6):339–347.
2. Au-Yeung SS, Ng JT, Lo SK. Does balance or motor impairment of limbs discriminate the ambulatory status of stroke survivors? Am J Phys Med Rehabil. 2003;82(4):279–283. doi:10.1097/01.PHM.0000056988.24854.8D.
3. Goh HT, Nadarajah M, Hamzah NB, Varadan P, Tan MP. Falls and fear of falling after stroke: a case-control study. PM R. 2016;8(12):1173–1180. doi:10.1016/j.pmrj.2016.05.012.
4. Miyata K, Kaizu Y, Usuda S. Prediction of falling risk after discharge in ambulatory stroke or history of fracture patients using Balance Evaluation Systems Test (BESTest). J Phys Ther Sci. 2018;30(4):514–519. doi:10.1589/jpts.30.514.
5. Tan KM, Tan MP. Stroke and falls—clash of the two titans in geriatrics. Geriatrics (Basel). 2016;1(4):31. doi:10.3390/geriatrics1040031.
6. Berg K, Wood-Dauphinee S, Williams JI. The Balance Scale: reliability assessment with elderly residents and patients with an acute stroke. Scand J Rehabil Med. 1995;27(1):27–36.
7. Berg KO, Wood-Dauphinee SL, Williams JI, Maki B. Measuring balance in the elderly: validation of an instrument. Can J Public Health. 1992;83(suppl 2):S7–S11.
8. Conradsson M, Lundin-Olsson L, Lindelof N, et al. Berg Balance Scale: intrarater test-retest reliability among older people dependent in activities of daily living and living in residential care facilities. Phys Ther. 2007;87(9):1155–1163. doi:10.2522/ptj.20060343.
9. Pickenbrock HM, Diel A, Zapf A. A comparison between the Static Balance Test and the Berg Balance Scale: validity, reliability, and comparative resource use. Clin Rehabil. 2016;30(3):288–293. doi:10.1177/0269215515578297.
10. Telenius EW, Engedal K, Bergland A. Inter-rater reliability of the Berg Balance Scale, 30 s chair stand test and 6 m walking test, and construct validity of the Berg Balance Scale in nursing home residents with mild-to-moderate dementia. BMJ Open. 2015;5(9):e008321. doi:10.1136/bmjopen-2015-008321.
11. Toomey E, Coote S. Between-rater reliability of the 6-minute walk test, Berg Balance Scale, and handheld dynamometry in people with multiple sclerosis. Int J MS Care. 2013;15(1):1–6. doi:10.7224/1537-2073.2011-036.
12. Viveiro LAP, Gomes GCV, Bacha JMR, et al. Reliability, validity, and ability to identity fall status of the Berg Balance Scale, Balance Evaluation Systems Test (BESTest), Mini-BESTest, and Brief-BESTest in older adults who live in nursing homes. J Geriatr Phys Ther. 2019;42(4):E45–E54. doi:10.1519/JPT.0000000000000215.
13. Alghwiri AA, Alghadir AH, Al-Momani MO, Whitney SL. The Activities-specific Balance Confidence Scale and Berg Balance Scale: reliability and validity in Arabic-speaking vestibular patients. J Vestib Res. 2016;25(5/6):253–259. doi:10.3233/VES-160568.
14. Babaei-Ghazani A, Mohammadi H, Shahidi GA, et al. Reliability and validity of the Persian translation of Berg Balance Scale in Parkinson disease. Aging Clin Exp Res. 2017;29(5):857–862. doi:10.1007/s40520-016-0682-7.
15. Berardi A, Galeoto G, Valente D, Conte A, Fabbrini G, Tofani M. Validity and reliability of the 12-item Berg Balance Scale in an Italian population with Parkinson's disease: a cross sectional study. Arq Neuropsiquiatr. 2020;78(7):419–423. doi:10.1590/0004-282x20200030.
16. Dijkhuizen A, Krijnen WP, van der Schans CP, Waninge A. Validity of the modified Berg Balance Scale in adults with intellectual and visual disabilities. Res Dev Disabil. 2017;62:58–68. doi:10.1016/j.ridd.2017.01.005.
17. Godi M, Franchignoni F, Caligari M, Giordano A, Turcato AM, Nardone A. Comparison of reliability, validity, and responsiveness of the mini-BESTest and Berg Balance Scale in patients with balance disorders. Phys Ther. 2013;93(2):158–167. doi:10.2522/ptj.20120171.
18. Jacome C, Cruz J, Oliveira A, Marques A. Validity, reliability, and ability to identify fall status of the Berg Balance Scale, BESTest, Mini-BESTest, and Brief-BESTest in patients with COPD. Phys Ther. 2016;96(11):1807–1815. doi:10.2522/ptj.20150391.
19. Lemay JF, Nadeau S. Standing balance assessment in ASIA D paraplegic and tetraplegic participants: concurrent validity of the Berg Balance Scale. Spinal Cord. 2010;48(3):245–250. doi:10.1038/sc.2009.119.
20. Major MJ, Fatone S, Roth EJ. Validity and reliability of the Berg Balance Scale for community-dwelling persons with lower-limb amputation. Arch Phys Med Rehabil. 2013;94(11):2194–2202. doi:10.1016/j.apmr.2013.07.002.
21. Mehta T, Young HJ, Lai B, et al. Comparing the convergent and concurrent validity of the Dynamic Gait Index with the Berg Balance Scale in people with multiple sclerosis. Healthcare (Basel). 2019;7(1):27. doi:10.3390/healthcare7010027.
22. Sahin F, Yilmaz F, Ozmaden A, Kotevolu N, Sahin T, Kuran B. Reliability and validity of the Turkish version of the Berg Balance Scale. J Geriatr Phys Ther. 2008;31(1):32–37. doi:10.1519/00139143-200831010-00006.
23. Salavati M, Negahban H, Mazaheri M, et al. The Persian version of the Berg Balance Scale: inter and intra-rater reliability and construct validity in elderly adults. Disabil Rehabil. 2012;34(20):1695–1698. doi:10.3109/09638288.2012.660604.
24. Whitney S, Wrisley D, Furman J. Concurrent validity of the Berg Balance Scale and the Dynamic Gait Index in people with vestibular dysfunction. Physiother Res Int. 2003;8(4):178–186. doi:10.1002/pri.288.
25. Wirz M, Muller R, Bastiaenen C. Falls in persons with spinal cord injury: validity and reliability of the Berg Balance Scale. Neurorehabil Neural Repair. 2010;24(1):70–77. doi:10.1177/1545968309341059.
26. Mao HF, Hsueh IP, Tang PF, Sheu CF, Hsieh CL. Analysis and comparison of the psychometric properties of three balance measures for stroke patients. Stroke. 2002;33(4):1022–1027. doi:10.1161/01.str.0000012516.63191.c5.
27. Chen KL, Chou YT, Yu WH, Chen CT, Shih CL, Hsieh CL. A prospective study of the responsiveness of the original and the short form Berg Balance Scale in people with stroke. Clin Rehabil. 2015;29(5):468–476. doi:10.1177/0269215514549032.
28. Huang YJ, Lin GH, Lee SC, Chen YM, Huang SL, Hsieh CL. Group- and individual-level responsiveness of the 3-point Berg Balance Scale and 3-point Postural Assessment Scale for stroke patients. Arch Phys Med Rehabil. 2018;99(3):529–533. doi:10.1016/j.apmr.2017.08.472.
29. Huang YJ, Lin GH, Lee SC, Hsieh CL. A comparison of the responsiveness of the postural assessment scale for stroke and the Berg Balance Scale in patients with severe balance deficits after stroke. J Geriatr Phys Ther. 2020;43(4):194–198. doi:10.1519/JPT.0000000000000247.
30. Saso A, Moe-Nilssen R, Gunnes M, Askim T. Responsiveness of the Berg Balance Scale in patients early after stroke. Physiother Theory Pract. 2016;32(4):251–261. doi:10.3109/09593985.2016.1138347.
31. Cella D, Choi SW, Condon DM, et al. PROMIS((R)) adult health profiles: efficient short-form measures of seven health domains. Value Health. 2019;22(5):537–544. doi:10.1016/j.jval.2019.02.004.
32. Schultz-Larsen K, Lomholt RK, Kreiner S. Mini-Mental Status Examination: a short form of MMSE was as accurate as the original MMSE in predicting dementia. J Clin Epidemiol. 2007;60(3):260–267. doi:10.1016/j.jclinepi.2006.06.008.
33. Hsueh IP, Lin JH, Jeng JS, Hsieh CL. Comparison of the psychometric characteristics of the functional independence measure, 5 item Barthel index, and 10 item Barthel index in patients with stroke. J Neurol Neurosurg Psychiatry. 2002;73(2):188–190. doi:10.1136/jnnp.73.2.188.
34. Hsieh YW, Hsueh IP, Chou YT, Sheu CF, Hsieh CL, Kwakkel G. Development and validation of a short form of the Fugl-Meyer Motor Scale in patients with stroke. Stroke. 2007;38(11):3052–3054. doi:10.1161/strokeaha.107.490730.
35. Wang YL, Lin GH, Huang YJ, Chen MH, Hsieh CL. Refining 3 measures to construct an efficient functional assessment of stroke. Stroke. 2017;48(6):1630–1635. doi:10.1161/strokeaha.116.015516.
36. Lin GH, Li CY, Sheu CF, Huang CY, Lee SC, Huang YH, et al. Using machine learning to develop a short-form measure assessing 5 functions in patients with stroke. Arch Phys Med Rehabil. 2021. doi:10.1016/j.apmr.2021.12.006.
37. Lin GH, Huang CY, Lee SC, Chen KL, Lien JJ, Chen MH, et al. A 10-item Fugl-Meyer Motor Scale based on machine learning. Phys Ther. 2021;101(4):pzab036. doi:10.1093/ptj/pzab036.
38. Duncan PW, Sullivan KJ, Behrman AL, et al. Protocol for the Locomotor Experience Applied Post-stroke (LEAPS) trial: a randomized controlled trial. BMC Neurol. 2007;7:39. doi:10.1186/1471-2377-7-39.
39. Chou CY, Chien CW, Hsueh IP, Sheu CF, Wang CH, Hsieh CL. Developing a short form of the Berg Balance Scale for people with stroke. Phys Ther. 2006;86(2):195–204.
40. Wong CK, Chen CC, Welsh J. Preliminary assessment of balance with the Berg Balance Scale in adults who have a leg amputation and dwell in the community: Rasch rating scale analysis. Phys Ther. 2013;93(11):1520–1529. doi:10.2522/ptj.20130009.
41. La Porta F, Caselli S, Susassi S, Cavallini P, Tennant A, Franceschini M. Is the Berg Balance Scale an internally valid and reliable measure of balance across different etiologies in neurorehabilitation? A revisited Rasch analysis study. Arch Phys Med Rehabil. 2012;93(7):1209–1216. doi:10.1016/j.apmr.2012.02.020.

balance; machine learning; short form; stroke rehabilitation

Supplemental Digital Content

© 2022 Academy of Neurologic Physical Therapy, APTA.