Table 1.
Objective physical function assessment with ML technology.
| Reference | Main Device to Obtain Input Data | Details of Input Variable or Device | Label Setting/Label Measurement Method | Output Label/Sort of Task | ML Technology | Validation Method | Metrics from Best Model | Baseline Characteristics | Concept | |
|---|---|---|---|---|---|---|---|---|---|---|
| Polus et al. [53] | IMU | 4 sensors during TUG: above and below each knee before and 2 weeks after THA | TUG > 14(6 weeks after THA) | TUG/classification | LDA, SVM | 10-fold CV | LDA: Accuracy 0.87 | 72 patients undergoing THA | Preventing falls by predicting their risk based on TUG | |
| Friedrich et al. [54] | IMU | Single sensor on the right side of the hip | SPPB: score itself TUG: <10, 11–19, 20–29 |
SPPB/regression TUG/classification |
LSTM+CNN | Train–val–test | Accuracy (TUG) 95.9% Accuracy (SPPB) 94.3% |
20 older patients (OTAGO study) | Predicting TUG on real-life IMU data | |
| Bloomfield et al. [55] | IMU+EHR | ·4 sensors: above and below each knee during TUG ·Clinical information ·Patient-reported subjective measures |
(Preoperative TUG—postoperative) >2.27 |
TUG/classification | SVM, NB, RF, | 10-fold CV | RF: Accuracy 0.80 | 82 patients undergoing TKA | Predicting functional recovery for appropriately adjusting patient expectations | |
| Zhuparris et al. [56] | Smartphone | ·Health-related data from smartphone ·Sensor in smartphone |
TUG score itself | TUG/regression | Elastic Net, RF, xgBoost | 5-fold CV | Elastic Net: R2 0.59 |
38 patients with FSHD | Quantifying FSHD progression with TUG | |
| Dubois et al. [57] | Depth sensor | Kinect V2 placed in each room of the rehabilitation center | TUG ≥ 13.5 s | TUG/classification | AdaBoost, NB, KNN, SVM, RF, NN | Leave-one-out CV | KNN, NN: Accuracy 1.0 | 30 older patients in a rehabilitation center | Preventing fall with home-sensor data | |
| Hasegawa et al. [58] | EHR | ·Clinical information mainly from EHR ·Physical measurements |
SPPB ≤ 6(men)/≤9(women) as fall risk | SPPB/classification | Prediction One. Ver3.0.1.3 (SONY) BLRA |
Train–test split | Prediction One: Accuracy 0.74 |
797 older patients at frailty outpatient service | Comparing model performance of predicting fall risk based on SPPB | |
| Kraus et al. [59] | EHR | Clinical information from HER | TUG score itself | TUG/regression | GLM, SVM, RF, xgBoost | 5-fold CV | RF: MAE 2.7 | 103 orthogeriatric patients | Predicting TUG without mobility data | |
| Sasani et al. [60] | Tabular data | Components of GA | TUG < 10 s, TUG ≥ 10 s, uncertain | TUG/classification | Decision Tree Classifier | None | Decision Tree Classifier: Accuracy 78% | 1901 old patients undergoing cancer surgery | Predicting accurately TUG score with ML | |
| Li et al. [61] | Video | Stereo camera | TUG score itself | TUG/regression | Mask R-CNN+ polynomial regression |
None | RE <0.1 (20 participants in 40) | 40 older adults in a daycare facility | Assessing the health status of the older patients with TUG | |
| Hwang et al. [62] | Tabular data | Variables from physical profile and body part measurements (not from EHR) | Grip strength score itself | Grip strength/regression | MLP regression and different polynomial regressions | K-fold CV | MLP regression: correlation 0.88 | 164 healthy young volunteers | Predicting grip strength accurately to reduce the risk of upper extremity disorder | |
| Bae et al. [63] | Big Data | Tabular data from Korean National Fitness Award Data from 2009 to 2019 | Grip strength score itself | Grip strength/regression | LR, LASSO, Ridge, RF, xGBoost, Light GBM, CatBoost | 5-fold CV | CatBoost: MSE 16.6 | 107,290 participants aged over 65 | Predicting grip strength without measuring | |
| Supratak et al. [64] | IMU | Single sensor on the lower back | 25-foot walking test in clinic | Walking speed/ regression |
SVR | Correlation | Correlation 0.98 | 32 young patients with MS | Validating gait speed at home against a 25-foot walking test | |
| Soltani et al. [65] | IMU+GNSS | 2 sensors: on each wrist | Walking speed measured by GNSS | Walking speed/ regression |
LASSO (feature extraction) | CV | RMSE 0.05 | 40 healthy young volunteers | Estimating walking speed with personalization | |
| Dobkin et al. [66] | IMU | 2 sensors: above each ankle | Walking speed measured by stopwatch | Walking speed/ regression |
Sensor system (Medical Daily Activity Wireless Network algorithm) | Correlation | Correlation 0.98 | 12 patients with stroke 6 healthy participants |
Acquiring quantitative data on daily performance | |
| Mannini et al. [67] | IMU | Single sensor on the right shoe | Walking speed manually measured | Walking speed/ regression |
·Hidden Markov model ·Strap-down integration ·LR |
Leave-one-out CV | R2 0.96 | 23 healthy adults | Exploring the ML method to predict walking speed | |
| McGinnis et al. [68] | IMU | 5 sensors: on sacrum, bilateral thigh, and bilateral shank | 6 min walking test on a treadmill | Walking speed/ regression |
SVR | Leave-one-out CV | RMSE 0.12 (patients with MS) | 17 healthy participants 30 patients with MS |
Resolving the hurdle of assessing walking speed | |
| Aziz et al. [69] | IMU | Single sensor inside one shoe | Slow/normal/fast speed | Walking speed/ classification |
RF, xgBoost, SVM | Train–test split | RF: Accuracy 1.0 | 10 healthy men | Analyzing gait patterns of aged people | |
| Atrsaei et al. [70] | IMU | Single sensor on the waist | 10 m walk test | Walking speed/ regression |
GPR | Leave-one-out CV | RMSE 1.10 | 35 participants with MS | Predicting walking speed at home with IMU | |
| Juen et al. [71] | Smartphone | Smartphone in waist belt at L3 | 6 min walking test | Walking speed/ regression |
SVM, GPR | Leave-one-out CV | SVM: Error 3.23 | 28 patients with pulmonary disease 10 healthy participants |
Monitoring individual health status continuously | |
| Aziz et al. [72] | Smartwatch | Smartwatch on the right wrist | Speed during treadmill walking: 0.5, 0.75, 1.0, 1.25, 1.5, 1.75 m/s | Walking speed/ regression |
GPR | None | MAPE 4% (best, 1.0 m/s) | 10 healthy young adults | Assessing walking speed for preventing chronic diseases | |
| Lee et al. [73] | Optical motion capture+ EHR | ·Clinical information from EHR ·Variables extracted from optical motion capture |
The difference between post/pre-operative gait speed | Walking speed/ classification |
GBM | 10-fold CV | AUC 0.86 | 128 female patients undergoing bilateral TKA | Predicting postoperative walking speed by preoperative clinical variable | |
| Davis et al. [74] | Big Data | Tabular data | GSR = MGS—UGS | Walking speed/ regression |
HGBR | 5-fold CV | R2 0.21 | 3925 participants from TILDA wave3 | Predicting gait speed from population statistical data | |
| Sikandar et al. [75] | Image | 5 ratio-based body measurement from marker free video images | Slow (2 to 3 km/h), normal (4 to 5 km/h), and fast (6 to 7 km/h) | Walking speed/ classification |
BiLSTM | 17-fold CV | Accuracy 92.79% | 34 participants (OU-ISIR dataset A) | Classifying walking speed with body measurements | |
| Chen et al. [76] | Image | Plantar region pressure images | (0.8, 1.6, 2.4 m/s) and (10, 20 min) | Walking speed/ classification |
ROI+CNN | Train–test split | F1-score: 1.00 (first toe, 2.4 m/s for 10 min) | 12 healthy young participants | Detecting appropriate exercise intensity | |
| Kidzinski et al. [77] | Video | Timeline keypoint data derived from OpenPose | Walking speed measured by the VICON system | Walking speed/ regression |
OpenPose+ (CNN/RF/Ridge) | Train–val–test | OpenPose+CNN: Correlation 0.73 | 1026 pediatric patients with cerebral palsy | Simplifying the quantitative gait assessment | |
| Lonini et al. [78] | Video | Below-waist videos of patients recorded by normal camera | Walking speed measured by GAITRite | Walking speed/ regression |
DeepLabCut(ResNet based) | Leave-one-out CV | Correlation 0.92 | eight patients with stroke | Predicting the walking speed of patients with stroke without expensive instrument | |
Abbreviations: ML, machine learning; IMU, inertial measurement unit; EHR, electronic health record: GA, geriatric assessment; TUG, Timed Up and Go test; SPPB, Short Physical Performance Battery; LDA, linear discriminant analysis classifier; SVM, support vector machine; LSTM, long short-term memory; BiLSTM, bidirectional long short-term memory; CNN, convolutional neural network; NB, naive Bayes classifier; RF, random forest; xgBoost, eXtreme gradient boosting; AdaBoost, adaptive boosting; KNN, k-nearest neighbors; NN, neural network; BLRA, binomial logistic regression analysis; GLM, generalized linear model; MLP, multilayer perceptron regression; LR, linear regression; GBM, gradient boosting machine; CatBoost, categorical boosting; SVR, support vector regression; GPR, gaussian process regression; HGBR, histogram gradient boosting regression; CV, cross validation; ROI, region of interest; MAE, mean absolute error; MSE, mean squared error; RMSE, root mean square error; RE, relative error; MAPE, mean absolute percentage error; AUC, area under the curve; R2, R-squared value; ICC, intraclass correlation coefficient; GSR, gait speed reserve; MGS, maximum gait speed; UGS, usual gait speed; GNSS, global navigation satellite systems; TKA, total knee arthroplasty; THA, total hip arthroplasty; FSHD, facioscapulohumeral muscular dystrophy; MS, multiple sclerosis; OU-ISIR, Osaka University Institute of Scientific and Industrial Research; TILDA, The Irish Longitudinal Study on Aging.