Machine Learning Revolutionizes Division-1 Women's Basketball Performance Analysis

By Susha Cheriyedath, M.Sc.Jan 19 2024

In an article published in the journal Nature, researchers employed machine learning (ML) to comprehensively assess Division-1 Women's basketball performance at player, team, and conference levels. Utilizing data from training, subjective stress, sleep, and in-game statistics, the predictive models achieved over 90% accuracy in reactive strength index and game score prediction. The study employed an ensemble approach and partial dependence plots (PDP) to identify feature importance, enabling coaches to monitor the readiness of athletes and enhance training strategies for improved sports performance.

Background

In the realm of sports coaching and training, understanding and optimizing the type, quantity, and frequency of training are crucial for enhancing athlete performance. Leveraging artificial intelligence (AI) in forecasting performance offered a promising avenue for data-driven strategies, providing coaches with robust feedback and informed decision-making capabilities. The synergy between AI techniques and coaches' expertise has the potential to significantly improve performance predictions.

While sports data analytics (SDA) powered by big data and ML has gained traction, existing efforts in basketball predominantly focus on individual players and team dynamics, with challenges in predicting performance and mitigating injury risks. Various ML approaches, including decision tree (DT)-based ensemble methods, artificial neural network (ANN), and support vector machine (SVM) have been employed for performance prediction, utilizing technical and tactical analyses.

This paper addressed existing gaps by presenting a multi-level approach, utilizing supervised ML to analyze an entire competitive basketball season. Unlike previous studies primarily concentrating on individual or team performance, this research spanned player, team, and conference levels. The methodology included evaluating player importance using Shapley values, ranking players based on contributions, and building a predictive model that can be generalized for individual athletes or an entire group. By incorporating new features and high sampling rate tracking systems, this study enhanced the understanding of athletic performance factors. This comprehensive approach aimed to provide coaches with valuable insights to optimize practice schedules, training variables, and overall stress, ultimately improving performance at individual, team, and season levels.

Methods

This study employed a three-tiered approach to assess Division-1 Women's basketball performance, focusing on athletes, teams, and conferences. Athlete-level readiness is evaluated using the reactive strength index modified (RSImod), calculated from jump height and contact time. At the team level, the game score was determined using key game metrics, reflecting in-game player contributions. Conference-level performance was measured through the player efficiency rating (PER), indicating a player's efficiency compared to the conference average. The study investigated the relationships between these levels and assessed if athlete readiness correlated with team game scores and if higher game scores led to increased PER.

Sixteen Division-1 female basketball players underwent comprehensive testing and monitoring, including workload data, vertical jump measurements, subjective questionnaires on stress and recovery, and sleep data collected through Whoop straps. ML methods were employed for analysis. The study obtained Institutional Review Board approval, and participants provided informed consent following the Declaration of Helsinki.

The analysis involved calculating training load, resistance training load, practice and game metrics, training monotony, training strain, and jump data. Subjective questionnaires and sleep data provided additional insights. Game scores were calculated using John Hollinger's metric, and PER was determined based on various player performance factors.
The ethical conduct of the study aligned with the Declaration of Helsinki, and detailed methodologies ensure data accuracy and participant welfare. The provided supplementary material includes diagrams illustrating the experimental approach and additional formulas for game score and PER calculations.

Data analysis

Addressing missing data using multiple imputation by chained equation (MICE) and mitigating multicollinearity with factor analysis (FA), the analysis predicted performance metrics using ML. At the player level, RSImod was predicted with 98.67% accuracy. Team-level game scores were predicted with 94.20% accuracy, while conference-level PER was forecasted with a mean squared error (MSE) of 0.026 and R2 of 0.68. Synthetic minority oversampling technique (SMOTE) and ensemble-based feature importance ensure robust predictions. Notable features included Training Strain, Average Speed, and Peak Power, emphasizing the significance of training, in-game statistics, and physiological metrics in performance prediction.

Key performance indicators (KPIs) at each tier, analyzed through partial dependence plots, offered nuanced understanding. At the player level, the RSImod prediction emphasized the impact of training, sleep, and subjective stress. The predicted team-level game scores highlighted the significance of in-game statistics, recovery time, and average speed. The conference-level PER showcased the importance of peak power, maximum speed, sleep consistency, and emotional balance.

The study identified the most important modality at each level, with training data impacting RSImod, in-game statistics dominating game scores, and a combination of modalities influencing PER. While strengths included detailed analysis and a holistic perspective, limitations involve the specificity of female basketball players and a one-year data span.

Conclusion

In conclusion, this study successfully forecasted athletes' performance across individual, team, and conference levels, offering a nuanced evaluation of diverse modalities at each tier. The detailed quantification and elucidation of these predictions serve as invaluable tools for coaches. They enable continuous monitoring of athlete readiness and informed training adjustments on an individual basis.

Additionally, the insights derived from team-level performance predictions aid in strategic decision-making regarding team composition for upcoming matches. At the conference level, identifying the most valuable player facilitates strategic planning for the impending season. This multi-level predictive approach equips coaches with actionable information, fostering a more dynamic and informed approach to athlete development and team strategy.

The authors recommend including multi-modal data collection for informed decision-making, real-time dashboard applications, and ongoing research to test hypotheses and improve athletic performance. Future work entails addressing biases, incorporating time-series approaches, and expanding the model to diverse sports.