Multi-Modal Deep Clustering Survival Machines for Alzheimer's Disease Subtype Discovery.

Area of Science:

• Biomedical data science
• Artificial intelligence in medicine
• Computational biology

Background:

• Current survival clustering methods often use single data types and separate risk prediction from patient subtyping.
• There is a need for a unified approach to integrate diverse biomarkers for simultaneous subgroup discovery and risk prediction.

Purpose of the Study:

• To introduce a novel unified framework, Multi-Modal Deep Clustering Survival Machines (MMDCSM), for integrated biomarker analysis.
• To simultaneously discover patient subgroups and predict disease conversion risk using multi-modal data.

Main Methods:

• MMDCSM encodes individual data modalities using modality-specific MultiLayer Perceptrons (MLPs).
• Embeddings are fused into a joint representation for integrated analysis.
• Survival outcomes are modeled using a mixture of Weibull expert distributions for subtype definition and survival curve estimation.

Main Results:

• MMDCSM outperformed existing methods in identifying distinct low- and high-risk subgroups for Alzheimer's disease (AD) conversion in a cohort of 382 Mild Cognitive Impairment (MCI) patients.
• The framework achieved competitive accuracy in predicting personalized timelines to disease progression.
• Key brain regions, including the hippocampus, were identified as influential in distinguishing high-risk converters.

Conclusions:

• MMDCSM offers a powerful, unified framework for integrating multi-modal biomarker data for survival analysis.
• The approach enables more accurate, early-stage risk stratification for MCI patients, facilitating targeted interventions.
• This method holds promise for improving the management and potential prevention of AD progression.