Title: Interpretable Athlete Performance Modelling in Collegiate Basketball: A Review of Machine Learning and Computer Vision Methods
Author: Srishti Sharma, Srikrishnan Divakaran, Tolga Kaya, et.al.
Institution: Ahmedabad University, Sacred Heart University
Publication Date: February 2, 2026
The Problem: Traditional collegiate and professional basketball analytics relies heavily on surface level box score metrics or opaque deep learning "black box" models. This creates a severe operational inefficiency for front offices and coaching staffs: while complex neural networks can deliver high raw predictive accuracy, their internal logic is obscured. Consequently, decision makers cannot establish the organizational trust required to act on high-stakes predictions.
Furthermore, conventional evaluation models frequently isolate on-court performance data, ignoring critical, interconnected multi-modal stressors such as sleep quality, training load spikes, travel fatigue, and psychological readiness that directly drive performance degradation and non-contact injury risks.
Methodology: To bridge the gap between predictive precision and operational transparency, this systematic review of 106 studies advocates for pairing interpretable classical machine learning algorithms (e.g., XGBoost, Random Forests) with multi-modal datasets and Explainable AI (XAI) frameworks
To conceptualize the deployment:
Computer Vision (CV): Functions like an automated video coordinator. It systematically breaks down film to track player trajectories and extract lower-limb biomechanical features (such as knee flexion angles) during movement phases like jump landings.
Neural Networks: Act like a highly complex web of scout networks. They are incredibly proficient at recognizing subtle, layered patterns across thousands of video frames or player tracking logs, but their internal logic is notoriously difficult to map.
XAI (SHAP, LIME, PDP): Serves as an algorithmic translator. It takes complex model outputs and provides clear, post-hoc justifications—explicitly detailing exactly how much a sleep deficit or an asymmetrical landing force contributed to a high-fatigue or high-injury warning flag.
Why it Matters:
For a General Manager, this methodology mitigates valuation errors and protects te teams’ most valuable assets. Deep learning models frequently overfit the small-to-moderate sample sizes common in athletic environments, rendering them operationally unreliable. Conversely, classical machine learning coupled with XAI yields transparent, context rich rationale.
From an ROI perspective, integrating multi-source textual scouting reports with performance data has shown the potential to outperform human draft selections by 70% on average, yielding substantial on-court financial surplus over rookie scale contracts. On the court, utilizing XAI to flag non-contact injury risks from wearable metrics (such as acute workload spikes or landing asymmetries) has successfully reduced player injuries by 22%. This ensures maximum return on asset availability and prevents catastrophic contract depreciation
ACTIONABLE TAKEAWAYS
Establish Automated Strain Threshold Limits: GMs and Performance Directors should use Partial Dependence Plots (PDPs) to map team specific fatigue thresholds, automatically triggering roster-load or rotation adjustments before a player crosses into high risk training zones.
Personalize Lineup Readiness Protocols: Head Coaches should move away from rigid, group level baseline generalizations. Implement Smallest Worthwhile Change (SWC) thresholds to personalize daily lineup and rotation decisions based on unique intra-athlete reactive strength and recovery metrics.
Upgrade Scouting Arbitrage Models: Scouting Directors should utilize Natural Language Processing (NLP) to convert qualitative text from historical scouting reports into structured data features. Integrate these features with traditional combine measurements within a Random Forest model to identify undervalued draft assets.