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

Area of Science:

• Infectious Diseases
• Immunology
• Computational Biology

Background:

• HIV-associated cryptococcosis presents unpredictable disease courses and high mortality rates globally.
• Current risk stratification and clinical management strategies for this condition are limited, necessitating improved approaches.

Purpose of the Study:

• To develop and validate machine learning models for predicting disease severity and survival outcomes in HIV-associated cryptococcosis.
• To identify immune phenotypes associated with disease severity and prognosis.

Main Methods:

• Analysis of clinical and immunological data from 98 HIV-related cryptococcosis cases.
• Application of unsupervised clustering, elastic net regularized Cox regression, and random survival forests.
• Rigorous model performance assessment using C-index, Brier score, and time-dependent AUC with nested cross-validation.

Main Results:

• Identification of an immune phenotype with excessive inflammation (EXC) linked to increased disease severity, neurological symptoms, and poorer survival.
• An elastic net regularized Cox regression model achieved a mean C-index of 0.78 and a mean Brier score of 0.13 for 36-month outcomes.
• Time-dependent AUC values of 0.84 at 12 months and 0.79 at 36 months demonstrated the model's predictive robustness.

Conclusions:

• Cytokine-based clustering offers nuanced severity stratification for HIV-associated cryptococcosis.
• A penalized survival model enhances personalized risk assessment, supporting tailored clinical decision-making.
• Machine learning applications show promise for improving survival predictions and individualized patient management in this disease.