Multi-modal mixed-type structural equation modeling with structured sparsity for subgroup discovery from...

Area of Science:

• Health Informatics
• Machine Learning
• Biostatistics

Background:

• Increasing availability of multi-modal, mixed-type health data (biobanks, EHRs, wearables) necessitates advanced analytical models.
• Clustering aims to identify homogenous subgroups from heterogeneous data for targeted studies and interventions.
• Clustering high-dimensional, multi-modal, mixed-type data presents significant challenges for existing statistical and machine learning models.

Purpose of the Study:

• To propose a novel Multi-modal Mixed-type Structural Equation Model (M2-SEM) with structured sparsity for precise subgroup discovery in heterogeneous health data.
• To develop an efficient algorithm (GH-EMM) for estimating models with mixed data types (continuous and categorical).

Main Methods:

• Developed a novel Multi-modal Mixed-type Structural Equation Model (M2-SEM) incorporating structured sparsity.
• Created a Gauss-Hermite-enabled Expectation-Majorization-Minimization (GH-EMM) algorithm to handle mixed data types within the Expectation Maximization (EM) framework.
• Validated the M2-SEM and GH-EMM through extensive simulation studies against benchmark models.

Main Results:

• The M2-SEM and GH-EMM demonstrated effective clustering of high-dimensional, multi-modal, mixed-type health data.
• Applied to cardiometabolic (CM) risk factors, the model identified distinct subgroups of individuals at low and high risk.
• The model successfully integrated diverse CM risk factors including nutrition, mental health, physical activity, and sleep.

Conclusions:

• The proposed M2-SEM with GH-EMM offers a powerful approach for subgroup discovery using complex health data.
• This method facilitates early identification of at-risk individuals for targeted interventions and population health promotion.
• Leveraging multi-modal mixed-type health data holds significant promise for precision medicine and public health strategies.