BITES

BITES estimates individual treatment effects from right-censored time-to-event (survival) data using counterfactual survival analysis to inform personalized treatment decisions.

Key Features:

• Counterfactual Survival Analysis: Predicts potential outcomes under alternative treatments using observational data with non-random treatment assignment.
• Right-Censored Time-to-Event Handling: Models survival outcomes in the presence of right-censoring common in clinical studies.
• Treatment-Specific Semi-Parametric Cox Loss: Employs a treatment-specific semi-parametric Cox loss function tailored for survival data.
• Deep Neural Network with IPM Regularization: Uses a deep neural network architecture incorporating Integral Probability Metrics (IPM) to balance distributions between treated and untreated groups.
• Scheduled Hyper-Parameter Optimization: Implements scheduled hyper-parameter optimization for model tuning.
• Performance and Validation: Demonstrated superior performance in simulation studies and validated on a breast cancer cohort where hormone treatment was optimized using six routine clinical parameters with independent cohort confirmation.

Scientific Applications:

• Personalized Medicine: Estimates individualized treatment effects to support tailoring interventions to patient-specific risk and treatment response profiles.
• Breast Cancer Treatment Optimization: Applied to optimize hormone treatment decisions in breast cancer cohorts using six routine clinical parameters with independent validation.
• Observational Time-to-Event Studies: Applicable to clinical research scenarios involving right-censored survival data and non-randomized treatment assignment.

Methodology:

Treatment-specific semi-parametric Cox loss; deep neural network architecture; Integral Probability Metrics (IPM) for balancing treated and untreated distributions; handling of right-censored time-to-event data; scheduled hyper-parameter optimization.