Learning clinical networks from medical records based on information estimates in mixed-type data

Area of Science:

• Computational biology
• Medical informatics
• Machine learning

Background:

• Diagnosing complex diseases requires integrating heterogeneous clinical and biomedical data.
• Assessing interdependencies between mixed-type variables (continuous/categorical) is challenging.

Purpose of the Study:

• To develop an efficient computational approach for assessing multivariate information significance between mixed-type variables.
• To uncover direct, indirect, and causal relationships in medical records using graphical models.

Main Methods:

• An efficient computational approach to compute and assess multivariate information significance.
• Extension of a machine learning method to reconstruct graphical models for mixed-type data.
• Application to analyze medical records of elderly patients with cognitive disorders.

Main Results:

• The proposed method outperforms existing tools on benchmark mixed-type datasets.
• A clinical network was constructed from patient records, visualizing interdependencies.
• Physiological insights were gained, linking cerebrovascular accidents to brain atrophy and cognitive impairment.

Conclusions:

• The computational approach effectively reconstructs graphical models for mixed-type data.
• The method provides novel insights into cognitive disorders by analyzing clinical data.
• This approach aids in understanding complex disease relationships and diagnostic practices.