Longitudinal modeling of Post-COVID-19 condition over three years: A machine learning approach using clinical, neuropsychological, and fluid markers

Post-COVID-19 condition (PCC) manifests with prolonged, heterogeneous symptoms challenging both, diagnosis and therapeutic management. This three-year longitudinal study analyzed data from 93 adults (mean age of 48.9 ± 14.0, 60 female) after confirmed SARS-CoV-2 infection. Every follow-up visit included clinical, neuropsychological, and laboratory assessments, capturing multidimensional indicators of patient health. A machine learning framework was implemented to classify temporal stage of patient health status, identify visit-specific predictive markers, and manage incomplete data using both native handling in tree-based models and explicit imputation techniques. Gradient boosting methods consistently achieved the best performance across all visit comparisons, achieving F1-scores close to or above 90%. Classification performance improved with greater time intervals between visits, suggesting progressive divergence in patient phenotypes over time. For discriminating follow-up stages, inflammatory markers emerged as the most informative predictors, followed by SARS-CoV-2 antibody levels and neuropsychiatric measures for fatigue and cognitive performance. Interpretability analyses using SHAP and LIME confirmed the contribution of these features, while revealing shifts in feature relevance across years. These findings highlight the utility of machine learning in characterizing follow-up stage separability in PCC and offer clinically interpretable insights that prioritize immune and neuropsychological measures for monitoring and risk-stratified follow-up.