Background
Methods
Study design
Selection criteria
-
Type of studiesWe searched for all types of primary studies, excluding only reviews and overviews from the search.
-
Types of participantsWe included in the study all adult participants (over 18 years old) with stroke, independently of the type of stroke or the time post-onset (TPO).
-
Types of interventionWe included all the studies evaluating predictive models for outcome prognosis after rehabilitation treatment. We defined predictive models as either ML or theory-based algorithms trained on data and internally or externally validated on new data. Primary studies were excluded when the validation of the models, either internal or external, was not performed. We denoted as external the validation performed on new data, unseen from the model during the training phase and geographically and/or temporally independent from the training set. On the contrary, internal validation refers to methods involving only data from a single data acquisition campaign, eventually split into multiple subsets.Moreover, we considered the outcome of the model as a variable related to the motor functional status of the patient after the rehabilitation treatment, and we considered as predictors any variable related to the patients’ conditions before or during the rehabilitation. So, we included studies that evaluated the relationships between predictors and response, describing the functional recovery of the patient during the rehabilitation.
-
Types of outcomeWe selected studies evaluating motor functional outcomes and excluded studies involving only cognitive or only sensory-related outcomes. Because functional measures are less influenced than cognitive ones by external factors such as social and cultural biases, we preferred to limit our analysis to them. Nevertheless, we decided not to excessively constrain the selection of the outcome, including either upper and lower limb-related outcomes. Both features describing lower and higher-level domains with respect to the International Classification of Functioning, Disability and Health (ICF) were included, e.g. body functions activities and participation. We also discarded all studies considering responses collected more than three months after the end of the rehabilitation treatment to focus on the effective impact of the rehabilitation phase on the outcome.
Search methods for identification of studies
Database | Search string |
---|---|
PubMed | ((“machine learning”[MeSH Terms] OR “regression analysis”[MeSH Terms] OR “automated pattern recognition”[MeSH Terms]) AND (“stroke”[MeSH Terms]) AND (“rehabilitation”[MeSH Terms]) AND (“prognosis”[MeSH Terms] OR “rehabilitation outcome”[MeSH Terms] OR “clinical”[MeSH Terms] OR “efficacy treatment”[MeSH Terms])) OR ((“Machine Learning” OR “pattern recognition” OR “automated pattern recognition” OR “classif*” OR “regress*” OR “regression analysis”) AND (“stroke”) AND (“rehab*”) AND ((“pred*”) AND (“prognosis” OR “rehabilitation outcome” OR “clinical” OR “efficac*” OR “efficacy treatment” OR “treatment effect” OR “treatments effect”))) Sort by: Best Match Filters: English |
Web of Science | (TS = ((“Machine Learning” OR “pattern recognition” OR “automated pattern recognition” OR “classif*” OR “regress*” OR “regression analysis”) AND (“stroke”) AND (“rehab*”) AND ((“pred*”) AND (“prognosis” OR “rehabilitation outcome” OR “clinical” OR “efficac*” OR “efficacy treatment” OR “treatment effect” OR “treatments effect”)))) AND LANGUAGE: (English) |
Scopus | TITLE-ABS-KEY ((“Machine Learning” OR “pattern recognition” OR “automated pattern recognition” OR “classif*” OR “regress*” OR “regression analysis”) AND (“stroke”) AND (“rehab*” OR “rehabilitation”) AND ((“pred*”) AND (“prognosis” OR “rehabilitation outcome” OR “clinical” OR “efficac*” OR “efficacy treatment” OR “treatment effect” OR “treatments effect”))) AND (LIMIT-TO (LANGUAGE, “English”)) |
CENTRAL | ((pred*) AND (prognosis OR “rehabilitation outcome” OR clinical OR efficac* OR “efficacy treatment” OR “treatment effect” OR “treatments effect”)) AND (“Machine Learning” OR “pattern recognition” OR “automated pattern recognition” OR classif* OR regress* OR “regression analysis”) AND (stroke) AND (rehab*) |
CINAHL | ((“Machine Learning” OR “pattern recognition” OR “automated pattern recognition” OR “classif*” OR “regress*” OR “regression analysis”) AND (“stroke”) AND (“rehab*”) AND ((“pred*”) AND (“prognosis” OR “rehabilitation outcome” OR “clinical” OR “efficac*” OR “efficacy treatment” OR “treatment effect” OR “treatments effect”))) |
Data collection
-
Source of data: authors, publication year, study design and DOI.
-
Participant characteristics: age, number, specifications of the stroke event both in terms of aetiology and TPO.
-
Setting: monocentric or multicentric, type.
-
Outcomes: type, measures used, the timing of acquisition with respect to the rehabilitation treatment.
-
Predictors: type, measures used, the timing of acquisition with respect to the rehabilitation treatment, number.
-
Data treatment: number of missing data and treatment of missing data.
-
Methods used: features selection approach, the algorithm used, internal or external validation strategy.
-
Model performances: metrics used for performance evaluation, performance reported, limitations reported.
Assessment of risk of bias of the included studies
Data synthesis
Results
Study | Age (mean (std) or [range]) | Sample size | Further inclusion criteria specifications regarding stroke pathology (time from event or aetiology) | Outcome |
---|---|---|---|---|
Almubark et al. | N/R | 45 | Event happened more than 6 months before the study | Upper extremity home use |
Bates et al. | 70.4 (11.47) | 4020 | N/A | Physical grade achievement |
Berlowitz et al. | 67.7 (11.1) | 2402 | N/A | Functional outcome |
Bland et al. | [21–93] | 269 | N/A | Walking ability |
Cheng et al. | N/R | 82 | Ischemic | Recovery |
Li et al. | 65.6 (12.31) | 271 | First-ever ischemic | Functional status |
De Marchis et al. | [60–83] | 1102 | Acute ischemic | Unfavourable functional outcome |
de Ridder et al. | PAIS: 70.1 (13.4) PRACTISE: 70.6 (13.4) PASS: 71.9 (12.5) | PAIS = training = 1227 PASS = validation = 2125 (2107) PRACTISE = validation = 1657 (1589) | Ischemic | Disability and functional outcome |
George et al. | [24–84] | 35 | Chronic | Extent of motor recovery after constraint-induced movement therapy |
König et al. | Original: 68.1 (12.7) VISTA: 68.8 (12.3) | Original = 1754 VISTA = 5048 | Acute ischemic | Functional independence |
Kuceyeski et al. | 72.0 (12.0) | 41 | Ischemic | Clinical performance |
Abdel Majeed et al. | Control arm: 55.54 (12.63) Treatment arm: 55.23 (9.11) | 26 | Chronic | Change in clinical outcomes |
Masiero et al. | Construction set: 69 (12) Validation set: 68 (11) | 150 | Recent stroke (< 8 weeks post-event) | Ambulation |
Mostafavi et al. | N/R | 126 | Assessment of impairment | |
Sale et al. | N/R | 55 | Subacute (15 ± 10 days from injury) | Motor improvement |
Scrutinio, Lanzillo, et al. | Derivation set: 72 (12) Validation set: 70 (12) | 1592 | N/A | Functional status |
Scrutinio, Guida, et al. | [65–80] | 951 | 30 days from stroke occurrence | Treatment failure |
Sonoda et al. | Prediction group: 63.4 Validation group: 65.2 | 131 | N/A | Stroke outcome |
Zariffa et al. | [60–73] | 9 | Chronic | Measure of upper-limb function |
Criteria | Specification of the review question |
---|---|
Step 1: Specify your systematic review question | |
Intended use of the model: | Prediction of functional outcome after rehabilitation treatment of post-stroke patients |
Participants: | Adults post-stroke participants selected independently on the timing of the event or type of stroke |
Predictors: | Any kind of predictor was included, more specifically any type included in the following categories of stroke assessment: biomechanical assessment, functional assessment, demographic characteristics, medical history, stroke assessment and neurological assessment. The selected predictors are related to the admission or recovery phase only, excluding predictors variables collected at discharge |
Outcome: | Any kind of functional outcome, not exclusively cognitive or sensory-related was selected |
Study | Outcome | Type of prediction study |
---|---|---|
Step 2: Classify the type of prediction model evaluation | ||
Almubark et al. | Upper extremity home use | Development only |
Bates et al. | Physical grade achievement | Development only |
Berlowitz et al. | Functional outcome | Development only |
Bland et al. | Walking ability | Development only |
Cheng et al. | Recovery | Development only |
Li et al. | Functional status | Development only |
De Marchis et al. | Unfavourable functional outcome | Development and validation |
De Ridder et al. | Disability and functional outcome | Development and validation |
George et al. | Extent of motor recovery after constraint-induced movement therapy | Development only |
König et al. | Functional independence | Development and validation |
Kuceyeski et al. | Clinical performance | Development only |
Abdel Majeed et al. | Change in clinical outcomes | Development only |
Masiero et al. | Ambulation | Development only |
Mostafavi et al. | Assessment of impairment | Development only |
Sale et al. | Motor improvement | Development only |
Scrutinio, Lanzillo, et al. | Functional status | Development only |
Scrutinio, Guida, et al. | Treatment failure | Development only |
Sonoda et al. | Stroke outcome | Development only |
Zariffa et al. | Measure of upper-limb function | Development only |
Domain | Risk of bias (number of models) | Applicability (number of models) | ||
---|---|---|---|---|
Dev | Val | Dev | Val | |
Step 3: Assess risk of bias and applicability | ||||
Participants | High = 0 Unclear = 0 Low = 174 | High = 0 Unclear = 0 Low = 174 | High = 0 Unclear = 0 Low = 174 | High = 0 Unclear = 0 Low = 174 |
Predictors | High = 1 Unclear = 0 Low = 173 | High = 1 Unclear = 0 Low = 173 | High = 1 Unclear = 0 Low = 173 | High = 1 Unclear = 0 Low = 173 |
Outcome | High = 24 Unclear = 120 Low = 30 | High = 24 Unclear = 120 Low = 30 | High = 24 Unclear = 119 Low = 31 | High = 24 Unclear = 119 Low = 31 |
Analysis | High = 77 Unclear = 8 Low = 89 | |||
Overall | High = 85 Unclear = 67 Low = 22 | High = 35 Unclear = 110 Low = 29 |
Study | Number of models in the study | Outcomes | Outcome measure (type of outcome, ICF classification) | Outcome at discharge? Yes/no | Predictors (number) | Timing of the measurement | Methods for features selection | Algorithm of the best performing model | Validation approach | Measures and methods used for the description of model performance |
---|---|---|---|---|---|---|---|---|---|---|
Almubark et al. | 102 | Upper extremity use at home | MAL ratio (dichotomous variable, d5d6) | N/R | Trunk compensation, ARAT (3) | N/R | N/A | RF after PCA | Leave-One-Subject-Out | Classification accuracy 93.33% |
Upper extremity use at home | Accel ratio (dichotomous variable, b7) | KNN | Classification accuracy 86.66% | |||||||
Bates et al. | 1 | Physical grade achievement | FIM (numeric variable, d2d3d4d5d7) | Yes | Anagrafic data, clinical data, comorbidities data, acute procedures (38) | N/R | Unadjusted bivariate logistic analyses _ features selected are with p < 0.2 | LogR | 60% -40% split | ROC area on the derivation set = 0.84 ROC area on the validation set = 0.83 + Hosmer–Lemeshow test at p = 0.93 not significant on the derivation cohort |
Berlowitz et al. | 4 | Functional outcome | FIM change (numeric variable, d2d3d4d5d7) | Yes | Age, gender (2) | N/R | N/A | LR | Bootstrap method (1000 samples) | R^2 = 0.75 |
Bland et al. | 2 | Walking ability | 10 m walking speed (dichotomous variable, b7) | Yes | Motricity Index, somatosensation of the dorsum of the foot, Modified Ashworth Scale for plantar flexors, FIM walk item, Berg Balance Scale, 10-m walk speed, age, TPO (8) | Admission | Pearson product-moment correla_ tion | LogR | 110 -159 samples split | Sensitivity (0.94), specificity (0.65), OR (32), positive and negative predictive values (0.70, 0.93) |
10 m walking speed (numeric variable, b7) | LR | Sensitivity (0.94), specificity (0.65), OR (32), positive and negative predictive values (0.70, 0.93) | ||||||||
Cheng et al. | 3 | Recovery | MRS (dichotomous variable) | No, at 3 months | Gender, hypertension, heart disease, diabetes, previous stroke with yes or no nodes, age, OTT, NIHSS (8) | N/R | N/A | NN | 80%—20% split | ROC curve = 0.969,sensitivity = 0.9444,specificity = 0.9565,accuracy = 0.9512 |
De Marchis et al. | 2 | Unfavourable functional outcome | MRS (dichotomous variable, d2d4) | No, at 3 months | Age, NIHSS score, thrombolysis, log10-transformed copeptin levels (4) | Admission | Chosen variables that were independently associated with 3-month functional outcome in the dev and val cohorts | LogR | Model trained on COSMOS dataset (319) and tenfold CV; Ex. validated on CoRisk dataset (783) | Brier score + AUC (0.819) + NRI = continuous net reclassification index (0.46) |
De Ridder et al. | 7 | Functional outcome | MRS (dichotomous variable, d2d4) | No, at 3 months | Gender; age; NIHSS, Diabetes, previous stroke atrial fibrillation and hypertension (7) | N/R | Selected variables that were clinically relevant and/or previously reported to predict outcome in the literature | LR | Model trained on PAIS dataset (1227) and ex. validated on PASS dataset (2107) | AUC = 0.81 |
George et al. | 6 | Extent of motor recovery after constraint-induced movement therapy | WMFT (dichotomous variable, d2d4) | Yes | Side of motor impairment, motor predictors: each of the 15 WMFT natural-log-transformed item times; Sensory-motor predictors: BKT score, TM for the affected side (18) | N/R | All possible combinations of 18 inputs, a total of 262,125 combinations, were generated | NN | 35 different splits at different random ratios (RTT) | Accuracy = 100% |
König et al. | 1 | Functional independence | BI (dichotomous variable, d2d4d5) | No, at 3 months | Single items as well as the overall score of the NIHSS (16) | N/R | Systematic literature search | LogR | Model trained on original dataset (1754); ex. validated on VISTA dataset (5048) | AUC = 72.9% |
Sonoda et al. | 2 | Stroke outcome | Motor FIM (numerical variable, d2d4d5) | Yes | Total cognitive subscore of the FIM, age, days from stroke onset to dmission, motor-FIM (4) | Admission | N/A | LR | 87 -44 samples split | Correlation coefficients = 0.93 |
Kuceyeski et al. | 7 | Clinical performance | Motor FIM (numerical variable, d2d4d5) | N/R | Right inferior occipital and calcarine areas (N/R) | N/R | Jackknife CV | LR | Bootstrap | Akaike Information Criterion (AIC) and R^2 = 0.45 (0.08) |
FIM (numerical variable, d2d3d4d5d7) | Akaike Information Criterion (AIC) and R^2 = 0.37 (0.08) | |||||||||
MI (numerical variable, b7) | Akaike Information Criterion (AIC) and R^2 = 0.54 (0.14) | |||||||||
Li et al. | 2 | Functional status | BI (numerical variable, d2d4d5) | Yes | Demographic information (age, sex and smoking habit), medical history (hypertension, diabetes mellitus, atrial fibrillation and hypercholesterolemia), evaluation at initial admission in the emergency department (blood glucose, blood pressure, laboratory data and the stroke severity) (N/R) | Admission | N/A | LR | CV (90–10% _ split) | R^2 adjusted = 0.573 |
Scrutinio, Lanzillo, et al. | 2 | Functional status | FIS (dichotomous variable, d2d4d5) | Yes | Age, sex, marital staus, employment status, hypertension, diabetes mellitus, COPD, coronary heart disease, atrial fibrillation, TPO, aetiology, side of impairment, aphasia, unilateral neglect, M-FIM, cognitive FIM, blood urea nitrogen, estimated glomerular filtration rate, hemoglobin (19) | Admission | Forward stepwise selection approach with P < 0.05 | LogR | 717–875 samples split | AUC (0.913), Hosmer–Lemeshow test ( 1.20 (P = 0.754)) and calibration plots |
Motor FIM (dichotomous variable, d2d4d5) | AUC (0.883), Hosmer–Lemeshow test ( 4.12 (P = 0.249)) and calibration plots | |||||||||
Mostafavi et al. | 12 | Assessment of impairment | MAS (numerical variable, b7) | Yes | postural hand speed; reaction and its timing; initial movement direction error/ratio, hand speed ratio; number of speed peaks, speed ranges; movement time, hand path length, and maximum hand speed trial-to-trial variability of the active hand; contraction/expansion of the overall spatial area of the active hand relative to the passive hand; systematic shift between the passive and active hand (8) | During every session, they are instrumental attributes | N/A | PCI | tenfold CV, repeated 100 times + external valiudation | R-value, RMSE, NRMSE (0.054, 0.405, 31.2) |
Masiero et al. [29] | 1 | Ambulation | FAC (dichotomous variable, d4) | Yes | Age, gender, arterial hypertension, hypolipoproteinaemia, diabetes, event date and aetiology, paralysed side length of hospital stay, up MI and low MI, TCT, FIM and mot FIM (12) | Admission | N/R | LogR | 100–50 samples split | ROC curves (ROC area = 0.94, CI 95%: 0.86–0.96, p < 0.0001), with sensitivity of 86.5% (CI 95%: 77–96%) and specificity of 95.5% (CI 95%: 75–95%)) |
Abdel Majeed et al. | 8 | Change in clinical outcomes | FM change (numerical variable, b2b7) | Yes | Demographic/physiological characteristics descriptive statistics of movement (51) | Demogr. and physiol. at baseline, movement features | Random forests with 100 repeats of fourfold CV | LR | CV | RMSE and R^2 < 2.24% |
Scrutinio, Guida, et al. | 1 | Treatment failure | FIM-M (dichotomous variable, d2d4d5) | Yes | Age, sex, marital status, diabetes mellitus, TPO, stroke type, side of impairment, FIM-M and cognitive scores, neglect (10) | N/R | Backward stepwise selection (P > 0.157 for exclusion) | LogR | Resampling 200 bootstrap replications | Hosmer–Lemeshow test (7.77 (PZ.456)) and AUC (0.834) |
Mostafavi et al. | 12 | Assessment of impairment | FIM-M (numerical variable, d2d4d5) | Yes | postural hand speed; reaction and its timing; initial movement direction error/ratio, hand speed ratio; number of speed peaks, speed ranges; movement time, hand path length, and maximum hand speed trial-to-trial variability of the active hand; contraction/expansion of the overall spatial area of the active hand relative to the passive hand; systematic shift between the passive and active hand (8) | During every session, they are instrumental attributes | N/A | PCI | Tenfold CV, repeated 100 times | R-value, RMSE, NRMSE (0.562, 16.6, 21.7) |
FIM (numerical variable, d2d3d4d5d7) | R-value, RMSE, NRMSE (0.596, 16.8, 20.5) | |||||||||
Purdue Pegboard score (numerical variable, d2d4) | R-value, RMSE, NRMSE (0.483, 4.1, 14.1) | |||||||||
Abdel Majeed et al. | 8 | Change in clinical outcomes | WMFT change (numerical variable, d2d4) | Yes | Demographic/physiological characteristics descriptive statistics of movement (51) | Demogr. and physiol. at baseline, movement features | Random forests with 100 repeats of fourfold CV | LR | CV | RMSE and R^2 < 4.68% |
Sale et al. | 9 | Motor improvement | FIM-M (numerical variable, d2d4d5) | Yes | Age, gender, aetiology, first event, recombinant tissue plasminogen activator, BI, FIM motor impairment, dysphagia, tracheostomy, neuropsychological impairment, speech impairment, presence of nasogastric feeding tube, length of stay (14) | T0 = admission T1 = discharge | Mutual Information (MI) criterion | SVM | 20 rep. of hold-out approach with 70%—30% split + nested fivefold CV on the training set | Correlation, RMSE and MADP (0.76, 16.32, 26.79%) |
FIM (numerical variable, d2d3d4d5d7) | Correlation, RMSE and MADP (0.79, 18.78, 18.88%) | |||||||||
BI (numerical variable, d2d4d5) | Correlation, RMSE and MADP (0.75, 22.6, 83.96%) | |||||||||
Zariffa et al. | 2 | Measure of upper-limb function | FMA (numerical variable, b2b7) | Yes | Mean velocity, peak velocity, RMS jerk, mean-rectified jerk, number of peaks, path smoothness, speed smoothness, SPARC, passive ROMs, passive ROM Area, Active ROMs, Active ROM Area (14) | During 76 assessments | Exhaustive search of all the combinations of the 14 features | LR | Leave-one-subject-out | R^2 = 0.4390, SRD = 1.4621 |
ARAT (numerical variable, b7) | R^2 = 0.4246, SRD = 2.6803 |
Study characteristics
Risk of bias of the included studies
Participants
Predictors
Outcome
Analysis
Description of the input and output variables
-
Demographic characteristics (age, gender, marital status, employment status…).
-
Medical history (presence of hypertension, presence of diabetes mellitus, presence of chronic obstructive pulmonary disease, presence of chronic heart disease…).
-
Stroke assessment through clinical evaluation (length of stay, presence of dysphagia, presence of nasogastric tube, presence of tracheostomy…).
-
Stroke assessment through laboratory analysis (presence of recombinant tissue plasminogen activator, blood urea nitrogen, haemoglobin…).
-
Stroke assessment through imaging (area of the left supramarginal gyri obtained by MRI, area of the right thalamus obtained by MRI, area of the left superior parietal regions obtained by MRI).
-
Functional assessment (Motricity Index score, Modified Barthel Index score, Berg Balance Scale score…).
-
Neurological assessment through clinical examination (side of the impairment, type of stroke, TPO…).
-
Neurological assessment through instrumental examination (not reported).
-
Biomechanical assessment through clinical examination (10 m walking test speed).
-
Biomechanical assessment through instrumental examination (mean velocity from robotics assessment, peak velocity from robotics assessment, passive range of motion from robotics assessment, active range of motion from robotics assessment…).
Description of the methods
Model performances
Performance measures | Frequency among modelsa |
---|---|
Numerical outcomes | |
R2 | 10 |
RMSE | 9 (7) |
NRMSE | 4 |
R value | 8 |
MADP | 3 |
SRD | 2 |
Categorical outcomes | |
AUC | 9 |
Accuracy | 4 |
Sensitivity and specificity | 3 |
Hosmer–Lemeshow test | 4 |
NRI | 1 |