Table 2.
Summary of studies with multivariate analytic and machine learning approaches in movement assessment and outcome evaluation in CP.
| Study | Subject sample | Data | Methods | Main findings | Other findings |
|---|---|---|---|---|---|
| Meinecke et al. (52) | 22 infants (7 at risk of CP, 15 healthy) | 53 parameters extracted from recorded 3D movement data | Cluster analysis based on Euclidian distances and quadratic discriminant analysis used to find the best combined parameters and separate at risk infants from healthy ones | Overall detection rate (using an optimal combination of 8 parameters): 73% (sensitivity: 1.00; specificity: 0.70) | |
| Berge et al. (53) | 14 infants with CP (who had four types of fidgety movements) | Motion features (1D, 2D, and Wigner-Ville time-frequency virtue/feature) extracted from video recordings of movements | Periodicity (fidgety movements characterized by periodic patterns); principal components analysis (PCA) for data reduction; Pattern recognition (compare movement patterns in video with known visual patterns of fidgety movements) | ENIGMA (a software tool) can assess general movements and detect fidgety movements in CP pts | |
| Adde et al. (54) | 30 high-risk preterm and term infants (13 developed CP in 5 years vs. 17 non-CP) | Movement variables (e.g., quantity of motion, and centroid of motion to identify fidgety movements) extracted from video recordings | Mann–Whitney U test; Logistic regression to identify CP predictors; ROC analysis to assess CP classification accuracy | 1/13 of pts had fidgety movements; predictor of CP: combined variable (centroid of motion STD, quantity of motion mean, quantity of motion STD); prediction accuracy of the combined variable: sensitivity: 85%; specificity: 88% | Combined variable had the highest prediction accuracy; ambulatory and non-ambulatory function was predicted correctly in 90% pts with CP |
| Heinze et al. (55) | 4 infants with CP, vs. 19 healthy infants | 32 features (including velocity and acceleration) extracted from measurement of accelerometers | Optimal parameter combinations selected by genetic algorithm; a decision tree-based classifier used to differentiate between pts’ and controls’ data | Overall detection rate: 88–92% for all measurements | The low-cost movement disorder detection system based on accelerometers is applicable to CP diagnosis in newborns |
| Alaqtash et al. (56) | 4 pts with spastic diplegic CP, vs. 4 pts with multiple sclerosis, vs. 12 healthy controls | Gait features extracted from 3D ground reaction force data | NNC and ANN used to classify gait features into three groups; leave-one-out resampling | Classification accuracy (weighted average): 85% (using a combination of gait features); 95% (using an optimal set of six features) | |
| Karch et al. (57) | 10 infants with spastic CP, vs. 53 non-CP infants | Stereotypy score of limb movements extracted from electromagnetic movement tracking recordings | A multi-segmental chain model used to calculate the joint centers and joint axes; dynamic time warping used to compute stereotype scores; ROC analysis used to assess CP classification accuracy | CP classification accuracy using stereotype score of upper lime movement: sensitivity: 90%; specificity: 96% | Using stereotype score of leg movement could not distinguish pts from controls |
| Stahl et al. (58) | 82 infants (15 with CP, 67 healthy) | Motion features (such as motion distance and relative frequency) extracted from video movement recordings | Motion features selected to identify fidgety movements; SVM used to classify pts from controls; 10-fold cross-validation for classifier validation | Classification accuracy: with features of relative frequency: 93.7 ± 2.1%; sensitivity: 85.3 ± 2.8%; specificity: 95.5 ± 2.5% | Classification with other features (absolute motion distance and wavelet coefficient) had lower accuracy |
| Kanemaru et al. (59) | 145 preterm infants (16 developed CP by 3 years of age, vs. 129 normal) | 6 movement indices (average velocity of limb movement, number of movement units, kurtosis of acceleration, jerk index, etc.) extracted from video recordings | Fisher’s exact test and Mann–Whitney U-test to distinguish pts from controls | CP pts had higher jerk index in the legs (p < 0.01), average velocity of the arms (p < 0.05), and number of movement units of the arms (p < 0.05) than controls | Jerkiness of spontaneous movements in preterm infants at term age is useful for predicting CP |
| Wahid et al. (60) | 51 children with diplegic CP vs. 34 healthy controls | Spatiotemporal gait data (physical properties, walking speed, etc.) | Multiple regression normalization and standard dimensionless equations used for data normalization; multiple regression normalization to identify the effects of AFO on gait in pts | Multiple regression normalization revealed difference in more spatiotemporal parameters in pts who walked with and without an AFO; after multiple regression normalization, most spatiotemporal parameters in pts with AFO became closer to those of controls | Multiple regression normalization may be useful in evaluating CP gait and gait classification |
| Parmar and Morris (61) | 5 healthy subjects (who did exercises correctly, and also mimic the errors/mistakes in exercise made by CP pts) | Features (joint positions, angles) in the time domain (also transformed to the frequency domain) extracted from 78 training samples and 47 testing samples of physical exercises video recording | 4 classifiers (SVM, NN, AdaBoosted decision tree, and DTW) used to distinguish good and erroneous exercises (in five sample exercises such as Blast-Off exercise in CP physical therapy) | Classification accuracy: 94.68% for AdaBoosted tree on joint data (in time domain); 90.89% for SVM on joint data (in frequency domain); 90.65% for SVM on joint data (in time domain); 90.3% for AdaBoosted tree on angle data (in time domain) | Classification accuracy: 90.13% for single-layer NN on joint data (in time domain); 87.63% for SVM on angle data (in frequency domain); 74.03% for DTW on angle data (in time domain) |
| Hemming et al. (62) | 4,007 children with CP | Data from five CP registers (birth characteristics, severity of CP, level of impairment, socioeconomic status, etc.) | Kaplan–Meier survival estimates performed; Multivariate proportional hazards model fitted for survival analysis | Death rate: ~8%; rate of children who survived to 20 years of age: 85–94%; predictors of CP survival: The number and severity of impairment | Birth weight and socioeconomic status might have impact on survival in certain register regions |
| Kim et al. (63) | 174 children with spastic CP who underwent SDR | Clinical data (age at surgery, types of CP, history of prematurity, motor function, history of seizures, etc.) | Univariate and multivariate logistic regression used to identify factors associated with surgical outcome | 6.3% pts had a poor outcome; predictor of outcome: type of CP (diplegia, quadriplegia) | Preoperative diagnosis was a strong predictor; intellectual delay was significant only in univariate analysis |
| Golan et al. (64) | 98 pts with spastic CP who underwent SDR | Data from hospital charts and radiographic spinal studies (preoperative and postoperative) | Univariate and multivariate regression analyses used to identify risk factors for spinal deformity | Risk factors for spinal deformity: CP severity; ambulatory function; age at surgery; gender | Factors associated with a lower rate of hyperlordosis: younger age at surgery and male gender |
| Majnemer et al. (65) | 95 children with CP | Data from Child Health Questionnaire and Pediatric QOL Inventory (by pts and parents), and measurements (impairments, activity limitations, etc.) | Multivariate analysis used to identify determinants of QOL | Indicators of physical well-being: motor and other activity limitations; predictors of social-emotional adaptation: family functioning, behavioral difficulties, and motivation | 47% pts had mild motor impairment |
| White-Koning et al. (66) | 500 children with CP (in 7 countries in Europe) | Data from the Kidscreen questionnaire (by pts and parents) | Multivariate analysis used to identify factors associated the differences in parents’ and pts’ reports | Factors associated with the differences in parents’ and pts’ reports: high levels of stress in parenting (negative influence), self-reported severe child pain | Pts’ self-reports higher than parents’ in 8 domains, lower in the finances domain, and similar in the emotions domain |
| Long et al. (67) | 71 pts with CP vs. 77 non-CP; all subjects underwent orthopedic surgery | Demographic, surgical, and medical data (intraoperative opioid dosing, postoperative ICU admission, postoperative oxygen desaturation, etc.) | Multivariate regression analysis used to determine intraoperative opioid dosing associated outcomes and other variables | CP pts received less intraoperative opioid than non-CP pts; predictors of postoperative ICU admission and postoperative oxygen desaturation: intraoperative opioid dosing | CP associated with decreased opioid dosing |
| Smits et al. (68) | 116 pts with CP | 3-year longitudinal data (motor function, intellectual capacity, etc.) | Univariate and multivariate analyses to investigate associations between the course of capabilities (e.g., in mobility) and CP-, child-, and family characteristics | Predictors of self-care: a model including level of gross motor function and intellectual capacity; predictors of mobility: a model only including level of gross motor function; predictors of social function: a model including level of bimanual function and paternal educational level | Greater increase in capabilities for higher level of functioning, except for level of paternal education |
| Sponseller et al. (69) | 204 pts with CP who underwent spinal fusion surgery (at 7 institutions) | Clinical data of patient, laboratory, and surgical characteristics | Univariate and multivariate regression analysis to identify factors associated with infection development | 6.4% patients developed deep wound infection; factors associated with deep wound infection: presence of a gastrostomy/gastrojejunostomy tube | |
| He et al. (70) | 61 pts with spastic CP | Serial R- and S-baclofen plasma concentrations | Mixed-effects population model and a 2-compartment model used for population pharmacokinetics analysis of oral baclofen; a final multivariable model used to describe oral baclofen profiles | Mean population estimate of apparent clearance/F: 0.273 L/h/kg with 33.4% IIV; apparent volume of distribution (Vss/F): 1.16 L/kg with 43.9% IIV; average baclofen terminal half-life: 4.5 h | Determinants of apparent clearance: body weight, a possible genetic factor, and age |
| Kato et al. (71) | 31 pts with CP and cervical myelopathy; 30 with CSM, all pts underwent posterior decompression surgery | Measurements of pedicle and placement of pedicle screws from CT scans | Multivariate analysis used to evaluate factors associated with the breach of cervical pedicle screws | 23% CP pts and 7% CSM pts had pedicle sclerosis; pedicle sclerosis associated with a higher risk of breach | |
| Kruijsen-Terpstra et al. (72) | 92 pts (2 years old) with CP | Longitudinal data (type of CP, GMFCS level, intellectual capacity, whether epilepsy, etc.) | Multivariate analysis used to identify determinants of development of self-care and mobility activities | Determinants of development of self-care activities: GMFCS and intellectual capacity; determinant of development of mobility activities: GMFCS | Self-care and mobility activity changes were less favorable in pts with severe CP |
| Shore et al. (73) | 320 children with CP who underwent VDRO for treatment of hip displacement | Clinical data (Age, sex, GMFCS, preoperative radiography, use of botulinum toxin, surgical performance, surgeon volume, etc.) | Univariate and multivariate (Cox regression) analyses used to determine effects of the data variables on surgical success; Kaplan–Meier survivorship curve generated | 92% success rate for GMFCS levels I and II vs. 76% success rate for GMFCS level V; predictor of surgical success: soft-tissue release at VDRO | 37% surgical failure; predictors of surgical revision: younger age at surgery, increased GMFCS level, and lower annual surgical hip volume |
| Mo et al. (74) | 206 children with CP who underwent surgical scoliosis correction | Clinical data (age, motor deficits, seizure history, verbal communication, mental retardation, Hydrocephalus severity, etc.) | Univariate and multivariate logistic regression used to identify factors causing poor IONM signals | Predictors of poor IONM signals: PVL, hydrocephalus, encephalomalacia; predictors of no signals: moderate or marked hydrocephalus, encephalomalacia | Predictors of no motor signal: focal PVL, moderate or marked hydrocephalus, encephalomalacia; predictors of no sensory signal: moderate hydrocephalus |
| Grecco et al. (75) | 56 children with spastic CP | Clinical and neurophysiologic data (age, gross motor function, laterality of motor impairment, injury location and MEP) | Univariate and multivariate logistic regression analyses used to identify predictors of tDCS responses | Predictors of good responses to tDCS (and gait training): MEP (for 6-min walk test and gait speed), and subcortical injury (for gait kinematics and gross motor function) | The interaction of MEP and brain injury location predicted the responsiveness of tDCS |
| Minhas et al. (76) | 1,746 pts who underwent orthopedic procedure (345 pts underweight, 952 pts normal weight, 209 overweight, 240 obese) | Clinical data (whether seizure, whether asthma, whether use steroid, surgical procedure, etc.) | Multivariate logistic regressions performed to evaluate the effect of BMI on complications | Risk factors for total and medical complications in spine, hip, and lower extremity procedures: underweight class | Weight was not associated with complications in tendon procedures; overweight and obesity not associated with increased risk for complications |
| Galarraga et al. (77) | 115 children with CP who underwent (hip, ankle, foot, etc.) surgery | Preoperative data (36 physical examination variables and gait kinematics) and surgery data | PCA data dimension reduction; multi-regression analysis used to predict postoperative lower limb kinematics | Based on the kinematic angle, mean prediction errors on test vary from 4° (pelvic obliquity and hip adduction) to 10° (hip rotation and foot progression) | Mean prediction errors are smaller than the variability of gait parameters |
| Mann et al. (78) | 128 pts with CP | Physical activity, physical, psychosocial and total QOL reported by parents, walking performance measured by a StepWatch device | Multivariate regression used to examine the relationship of physical activity and walking performance to QOL | Physical activity positively associated with physical and total QOL; walking performance positively associated with physical QOL | Participation level positively associated with psychosocial QOL |
AFO, ankle foot orthosis; AHA, assisting hand assessment; ANN, artificial neural networks; BMI, body mass index; CP, cerebral palsy; CSM, cervical spondylotic myelopathy; DTW, dynamic time warping; EMG, electromyographic; ENIGMA, enhanced interactive general movement assessment; GMFCS, Gross Motor Function Classification System; ICU, intensive care unit; IIV, interindividual variability; IONM, intraoperative neuromonitoring; MEP, motor-evoked potential; NN, neural networks; NNC, nearest neighbor classifier; Pts, patients; PVL, periventricular leukomalacia; QOL, quality of life; ROC, receiver operating characteristics; SDR, selective dorsal rhizotomy; surgeon volume, the number of procedures performed; SVM, support vector machines; tDCS, transcranial direct current stimulation; VDRO, proximal femoral varus derotation osteotomy.