XAI Prediction of Functional Connectivity Self-Similarity Changes Arising from Head Acceleration Event Exposure

Introduction: Many studies over the past fifteen years have demonstrated negative effects on functionally-derived brain networks resulting from cumulative exposure to head acceleration events (HAEs), even in the absence of clinical diagnosis of concussion. Resting-state functional magnetic resonance imaging (rs-fMRI) has been used to track longitudinal changes in the functional connectome (FC) of individuals. Decreased self-similarity (Iself)—assessed as the correlation between the FC matrix derived across multiple sessions—is usually interpreted as reflecting accrual of deleterious alterations, whereas increasing Iself is often taken as an outcome of active repair mechanisms. Understanding what factors may best facilitate effective repair of injuries associated with HAE exposure would allow for improved care of collision-sport athletes. Explainable artificial intelligence (XAI) methods were used to develop a predictive model of the observed recovery in Iself after a single season of participation in high school football.

Materials and

Methods: Using explainable artificial intelligence (XAI) methods, a decision tree regression model has been developed for predicting the longitudinal alterations in brain resting-state fMRI (rs-fMRI) functional connectomes (FCs) as a function of both HAE exposure and athlete demographics. 58 high school American football athletes (FBA) and 14 peer non-collision sport controls (NCC) underwent serial rs-fMRI around a competition season. FBA were also monitored for HAEs in all practices and games. Following quality assurance, FCs were generated for 42 FBA and 14 controls based on the covariance of mean time-series signals acquired from 245 functional brain regions. FCs are symmetric positive definite (SPD) and lie on a nonlinear Riemannian manifold. A Euclidean tangent space (Tx) based on NCC FCs was constructed to allow Euclidean operations on FBA FCs after mapping onto Tx (t-FCs). A Ledoit-Wolf regularization factor of 0.01 was applied to all FCs before tangent space conversion to ensure FCs were strictly SPD. Correlation-based fingerprint self-similarity (Iself) was evaluated for FBA across sessions during (In1, In2) and after (Post) the season, relative to pre-season (Pre). Change of Iself from In2 to Post was predicted based on HAEs accumulated throughout the season, as well as athlete demographics. ElasticNet was used for feature selection (from a set of 382 variables) via 4-fold cross-validation. Twenty features that appeared at least twice among the four folds were input to a decision tree to predict changes in Iself from In2 to Post. Shapley values were used to evaluate importance and polarity of feature effects.

Results, Conclusions, and Discussions: Individual fingerprint rates, at which specific subjects are differentiable from the cohort, improved significantly among NCCs based on t-FCs. The ID rate based on original FCs among NCC is 9/14 while the ID rate based on t-FCs is 13/14. This indicates that FBA t-FCs provide a robust biomarker for each individual. This validated the longitudinal comparisons of changes in Iself (based on t-FCs) from In2 to Post for FBA. T-SNE embedding (Figure 1) illustrated shifting of individual t-FCs across four sessions for FBA, indicating FBA functional brain connectivity was altered during the season. A set of 20 HAE- and demographic-related features selected by ElasticNet produced a best testing R^2=0.84 through decision tree regression. The most critical features indicate that FBA receiving more than 69 HAEs exceeding 30g within a 2 week period prior to In2 (near the end of the season) had lower Iself at Post. As expected, the greater the period since last contact at In2, the higher the Iself at Post. Full Shapley evaluation is included in figure 2, Decision tree results are shown in figure 3.
In summary, key variables observed to predict the completeness of recovery following a season of exposure to HAEs are measures that may be collected under circumstances of normal athletic participation. This is critical, as medical imaging techniques that have to-date identified deficits and/or recovery from HAE exposure are both expensive, inconvenient, and often contraindicated. The identified thresholds may be used by coaches and athletic training staff members to monitor and moderate athlete exposure to HAEs, potentially improving near-term recovery, and long-term health for these athletes.
While this work provides promising results for Iself prediction, use of additional data could allow for (i) greater sensitivity to demographic features, and (ii) construction of a more “unbiased” tangent space, likely to lead to the shift for each FBA t-FC being more sensitive to HAEs. Both will enhance construction of predictive models that can be used for athlete care.

Acknowledgements (Optional): This research was supported in part by Allied Milk Producers