Data-Driven Imputation for Cohort Studies Using Collegiate Basketball Data

Abstract

Missing data remains a critical challenge in cohort studies. This study introduces a novel missing-value imputation technique that integrates feature sensitivity analysis, factor analysis, clustering, and predictive modelling to enhance accuracy, reliability, and interpretability. The dataset comprises 42 features collected from 16 collegiate female basketball athletes over 26 weeks, including sleep and cardiac rhythms, training loads, cognitive states, travel, and countermovement jump performance. The objective is to model the impact of these contextual stressors on athletic readiness, quantified via the Reactive Strength Index modified (RSImod). Our proposed methodology achieved up to an 80.85% reduction in MSE and a 79.99% increase in scores in RSImod predictions on the imputed dataset, demonstrating substantial improvements over existing state-of-the-art approaches. The average reduction in computation time across the evaluated state-of-the-art methods was approximately 21.6%. External validation on an independent wearable-based sleep–HRV dataset (49 participants) confirmed generalisability, with the proposed model achieving an RMSE of 0.643 and outperforming baseline methods by 10–16%. Further, ablation analysis showed clear contributions from each module: clustering, hybrid feature weighting, factor analysis, and the pure XGBoost variant. Missing-data simulations confirmed robustness, with RMSE increasing from 0.74–0.95 (MCAR) to 0.80–1.06 (MAR) and 0.91–1.32 (MNAR), reflecting graceful degradation under structured missingness. Interpretability was enhanced using SHAP (SHapley Additive exPlanations) and ALE (Accumulated Local Effects) analyses, providing actionable insights for coaches and practitioners. © The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2026.