Explainable artificial intelligence is an emerging research direction helping the user or developer of machine learning models understand why models behave the way they do. The most popular explanation technique is feature importance. However, there are several different approaches how feature importances are being measured, most notably global and local. In this study we compare different feature importance measures using both linear (logistic regression with L1 penalization) and non-linear (random forest) methods and local interpretable model-agnostic explanations on top of them. These methods are applied to two datasets from the medical domain, the openly available breast cancer data from the UCI Archive and a recently collected running injury data. Our results show that the most important features differ depending on the technique. We argue that a combination of several explanation techniques could provide more reliable and trustworthy results. In particular, local explanations should be used in the most critical cases such as false negatives.

Abstract:Clustering is an unsupervised machine learning and pattern recognition method. In general, in addition to revealing hidden groups of similar observations and clusters, their number needs to be determined. Internal clustering validation indices estimate this number without any external information. The purpose of this article is to evaluate, empirically, characteristics of a representative set of internal clustering validation indices with many datasets. The prototype-based clustering framework includes multiple, classical and robust, statistical estimates of cluster location so that the overall setting of the paper is novel. General observations on the quality of validation indices and on the behavior of different variants of clustering algorithms will be given.

Previous studies have suggested that runners can be subgrouped based on homogeneous gait patterns; however, no previous study has assessed the presence of such subgroups in a population of individuals across a wide variety of injuries. Therefore, the purpose of this study was to assess whether distinct subgroups with homogeneous running patterns can be identified among a large group of injured and healthy runners and whether identified subgroups are associated with specific injury location. Three‐dimensional kinematic data from 291 injured and healthy runners, representing both sexes and a wide range of ages (10‐66 years), were clustered using hierarchical cluster analysis. Cluster analysis revealed five distinct subgroups from the data. Kinematic differences between the subgroups were compared using one‐way analysis of variance (ANOVA). Against our hypothesis, runners with the same injury types did not cluster together, but the distribution of different injuries within subgroups was similar across the entire sample. These results suggest that homogeneous gait patterns exist independent of injury location and that it is important to consider these underlying patterns when planning injury prevention or rehabilitation strategies.

The purpose of this article is to present how predictive machine learning methods can be utilized for detecting sport injury risk factors in a data-driven manner. The approach can be used for finding new hypotheses for risk factors and confirming the predictive power of previously recognized ones. We used three-dimensional motion analysis and physical data from 314 young basketball and floorball players (48.4% males, 15.72±1.79 yr, 173.34±9.14 cm, 64.65±10.4 kg). Both linear (L1-regularized logistic regression) and non-linear methods (random forest) were used to predict moderate and severe knee and ankle injuries (N=57) during three-year follow-up. Results were confirmed with permutation tests and predictive risk factors detected with Wilcoxon signed-rank-test (p<0.01). Random forest suggested twelve consistent injury predictors and logistic regression twenty. Ten of these were suggested in both models; sex, body mass index, hamstring flexibility, knee joint laxity, medial knee displacement, height, ankle plantar flexion at initial contact, leg press one-repetition max, and knee valgus at initial contact. Cross-validated areas under receiver operating characteristic curve were 0.65 (logistic regression) and 0.63 (random forest). The results highlight the difficulty of predicting future injuries, but also show that even with models having relatively low predictive power, certain predictive injury risk factors can be consistently detected.

Background: Injury risk prediction is an emerging field in which more research is needed to recognize the best practices for accurate injury risk assessment. Important issues related to predictive machine learning need to be considered, for example, to avoid overinterpreting the observed prediction performance. Purpose: To carefully investigate the predictive potential of multiple predictive machine learning methods on a large set of risk factor data for anterior cruciate ligament (ACL) injury; the proposed approach takes into account the effect of chance and random variations in prediction performance. Study Design: Case-control study; Level of evidence, 3. Methods: The authors used 3-dimensional motion analysis and physical data collected from 791 female elite handball and soccer players. Four common classifiers were used to predict ACL injuries (n = 60). Area under the receiver operating characteristic curve (AUC-ROC) averaged across 100 cross-validation runs (mean AUC-ROC) was used as a performance metric. Results were confirmed with repeated permutation tests (paired Wilcoxon signed-rank-test; P < .05). Additionally, the effect of the most common class imbalance handling techniques was evaluated. Results: For the best classifier (linear support vector machine), the mean AUC-ROC was 0.63. Regardless of the classifier, the results were significantly better than chance, confirming the predictive ability of the data and methods used. AUC-ROC values varied substantially across repetitions and methods (0.51-0.69). Class imbalance handling did not improve the results. Conclusion: The authors’ approach and data showed statistically significant predictive ability, indicating that there exists information in this prospective data set that may be valuable for understanding injury causation. However, the predictive ability remained low from the perspective of clinical assessment, suggesting that included variables cannot be used for ACL prediction in practice.