VdistCox: Vertically distributed Cox proportional hazards model with hyperparameter optimization.

Area of Science:

• Distributed computing
• Machine learning
• Biostatistics

Background:

• Vertically partitioned data involves distributing patient information across multiple sites.
• The Cox proportional hazards model (Cox model) is a standard survival analysis tool.
• Existing methods often require data sharing, posing privacy concerns.

Purpose of the Study:

• To propose a novel algorithm, VdistCox, for building an efficient Cox model in a vertically distributed setting.
• To enable hyperparameter tuning without sharing patient data.
• To explore the impact of randomness in neural network weights and biases.

Main Methods:

• VdistCox utilizes a single hidden layer feedforward neural network via extreme learning machine.
• Hyperparameter tuning (hidden nodes, activation function, regularization) is performed with minimal communication.
• Randomness in hidden layer input weights and biases was investigated.

Main Results:

• VdistCox efficiently builds a distributed Cox model for vertically partitioned data.
• The algorithm accurately reflects characteristics of centralized data without information sharing.
• Effective hyperparameter tuning was achieved with a single communication round.

Conclusions:

• VdistCox provides an efficient and privacy-preserving solution for distributed Cox modeling.
• The method facilitates robust hyperparameter optimization in decentralized environments.
• This approach is suitable for survival analysis on sensitive, vertically partitioned health data.