Table 5.
Study results characteristics
| Authors, Year | Performance Measures (+ Values for Best ML Model) | Predictive Performance of ML Methods | Measures of Feature Importance | Most Important Injury Predictors |
|---|---|---|---|---|
| Ayala et al., 2019 [2] | AUC (0.873), Sensitivity (77.8%), Specificity (83.8%) | An alternating decision tree, combined with synthetic minority oversampling and boosting gave the best results | The frequency with which each of the features appears across the tree classifiers | Sleep Quality |
| Carey et al., 2018 [10] | (Median) AUC (all below 0.65), Sensitivity, Specificity, Precision, False Disovery Rate, Likelihood Ratios | The proposed ML models perform only marginally better than would be expected by random chance | NR | NR |
| López-Valenciano et al., 2018 [25] | AUC (0.747), Sensitivity (65.5%), Specificity (79.1%) | An alternating decision tree, combined with synthetic minority oversampling and boosting gave the best results | The frequency with which each of the features appears across the tree classifiers | sport devaluation, history of muscle injury in last season |
| McCullagh et al., 2013 [27] | Accuracy (82.9%), Sensitivity (94.5%), Specificity (81.1%) | Indication that Artificial Neural Networks are able to derive meaningful information from the vast amount of data available to assist in the injury prediction process | NR | NR |
| Oliver et al., 2020 [32] | AUC (0.663), Sensitivity (55.6%), Specificity (74.2%) | The machine learning model provided improved sensitivity to predict injury | The frequency with which each of the features appears across the tree classifiers | interactions of asymmetry, knee valgus angle and body size |
| Rodas et al., 2019 [34] | Accuracy (52%), Sensitivity (75%), Specificity (23%) | There is low prediction potential for presence or absence of tendinopathy | The number of times that a feature (genetic predictor) received a non-zero coefficient in the LASSO analysis | rs10477683 in the fibrillin 2 gene was the most robust SNP (single-nucleotide polymorphism) |
| Rommers et al., 2020 [35] | F1-score (85%), Sensitivity (85%), Precision (85%) | It is possible to predict injury with high accuracy | SHAP (SHapley Additive exPlanations) summary plot | Higher predicted age at PHV (peak height velocity), longer legs, higher body height, lower body fat percentage |
| Rossi et al., 2018 [36] | (Mean) AUC (0.76), F1-score (64%), Sensitivity (80%), Specificity (87%) Precision (50%), Negative Predicted Value (96%) | The single Decision tree performs best in terms of precision | Mean decrease in Gini coefficient | Previous injury (exponential weighted moving average), total distance (monotony of workload feature) and high-speed running distance (exponential weighted moving average) |
| Ruddy et al., 2018 [38] | (Median) AUC (0.58, 0.57 and 0.52) | Eccentric hamstring strength, age, and previous hamstring strain injury (HSI) data cannot be used to identify athletes at an increased risk of HSI with any consistency | NR | NR |
| Thornton et al., 2017 [41] | AUC (0.74, 0.65, 0.64 and 0.64) | Machine learning techniques can appropriately monitor injury risk amongst professional team sport athletes | Number of times that each feature appears in the ensemble of decision trees | The relative importance of each training load variable varied for each player |
| Whiteside et al., 2016 [46] | Accuracy (75%), Sensitivity (74%), Specificity (75%), Precision (75%), False Omission Rate (26%) | Machine learning models can predict future ulnar collateral ligament surgeries with high accuracy | The frequency with which each feature appeared in the optimized models in the fivefold cross-validation | Mean days between consecutive games, pitches in repertoire, mean pitch speed, horizontal release location |