Generating Complex Explanations for Artificial Intelligence Models: An Application to Clinical Data on Severe Mental...

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Psychiatry

Background:

• Predicting mortality in severe mental illness (SMI) patients is challenging.
• Existing models often lack transparency in their predictions.
• Explainable AI (XAI) offers a path to understanding complex medical predictions.

Purpose of the Study:

• To develop an explainable AI methodology for predicting mortality in SMI patients.
• To utilize class-contrastive reasoning for generating interpretable machine learning models.
• To enhance hypothesis generation and clinical decision-making in personalized medicine.

Main Methods:

• Combined electronic health record data with SMI-relevant factors.
• Applied machine learning with a focus on class-contrastive reasoning.
• Generated complex, data-driven explanations for mortality predictions.

Main Results:

• Demonstrated the ability of class-contrastive reasoning to generate intricate explanations.
• Identified potential protective factors (e.g., diuretics) and risk factors (e.g., delirium and dementia in Alzheimer's) for mortality.
• Showcased how explanations can reveal patient subgroups and inform new research hypotheses.

Conclusions:

• The developed XAI methodology provides interpretable mortality predictions for SMI patients.
• Class-contrastive reasoning facilitates hypothesis generation and understanding of patient data.
• This approach represents a significant step towards explainable AI in personalized medicine and healthcare.