Treatment selection using prototyping in latent-space with application to depression treatment

Area of Science:

• Computational psychiatry
• Machine learning in healthcare
• Precision medicine

Background:

• Current machine-assisted treatment selection methods either personalize treatments without considering patient groups or group patients while ignoring individual differences.
• Both approaches have limitations, hindering optimal treatment selection.

Purpose of the Study:

• To propose a novel deep learning-based approach for treatment selection that integrates personalized and sub-grouping paradigms.
• To address limitations of existing methods by balancing individual patient variability and group-level patterns.

Main Methods:

• Developed a deep learning model utilizing latent-space prototyping to identify patient subgroups and individual characteristics.
• Evaluated the approach on synthetic data and real-world clinical data from 4754 Major Depressive Disorder (MDD) patients.

Main Results:

• The proposed model demonstrated significant improvements over state-of-the-art methods.
• Achieved an 8% absolute and 23% relative improvement in treatment allocation compared to random assignment.
• Showed favorable comparisons with existing treatment selection paradigms.

Conclusions:

• The novel deep learning approach offers a balanced strategy for machine-assisted treatment selection.
• This method holds potential for significant clinical impact in treating Major Depressive Disorder (MDD).
• The approach is particularly suited for domains where underlying patient subgroups are not readily apparent.