Predictive survival modelings for HIV-related cryptococcosis: comparing machine learning approaches.

Introduction: HIV-associated cryptococcosis is marked by unpredictable disease trajectories and persistently high mortality rates worldwide. Although improved risk stratification and tailored clinical management are urgently needed to enhance patient survival, such strategies remain limited. Methods: We analyzed clinical and immunological data from 98 HIV-related cryptococcosis cases, employing machine learning techniques to model disease severity and predict survival outcomes. Our approach included unsupervised clustering, elastic net regularized Cox regression, and random survival forests. Model performance was rigorously assessed using the C-index, Brier score, Calibration and time-dependent AUC, with validation executed through a comprehensive, multi-replicated nested cross-validation framework. Results: Through cytokine profiling, we identified an immune phenotype characterized by excessive inflammatory response (EXC), associated with greater disease severity, more frequent neurological symptoms, and poorer survival outcomes compared to the other two immune phenotypes, highlighting its potential significance in risk stratification. To further support clinical decision-making, we developed an elastic net regularized Cox regression model, achieving superior predictive accuracy with a mean C-index of 0.78 for 36-month outcomes and a mean Brier score of 0.13, outperforming both random survival forest and traditional Cox models. Time-dependent AUC analysis validated the model's robustness, with AUC values of 0.84 at 12 months and 0.79 at 36 months, indicating its reliability and potential clinical utility. Discussion: This study presents comprehensive and multidimensional approaches to overcome the challenges commonly encountered in real-world clinical settings. By applying cytokine-based clustering, we illustrate the potential for more nuanced severity stratification, offering a fresh perspective on disease progression. In parallel, our penalized survival model provides a step forward in personalized risk assessment, supporting informed clinical decisions and customized patient management. These findings suggest promising directions for individualized healthcare solutions, leveraging machine learning to enhance survival predictions in HIV-related cryptococcosis.