Explainable and externally validated machine learning for neurocognitive diagnosis via ECGs.

Area of Science:

• Cardiology
• Neurology
• Biomarker Discovery

Background:

• Electrocardiogram (ECG) analysis is increasingly recognized for detecting physiological changes in non-cardiac conditions.
• Cardiovascular and neurocognitive health are closely linked, suggesting ECG abnormalities may indicate neurocognitive disorders.
• ECG's potential as a biomarker for neurocognitive disorder detection, therapy monitoring, and risk stratification remains underexplored.

Purpose of the Study:

• To demonstrate the feasibility of predicting neurocognitive disorders using ECG features.
• To validate predictive models across diverse patient populations.

Main Methods:

• Utilized ECG features and demographic data to predict neurocognitive disorders (dementia, delirium, Parkinson's disease) based on ICD-10.
• Performed internal and external validations using the MIMIC-IV and ECG-View datasets.
• Assessed predictive performance using Area Under the Receiver Operating Characteristic Curve (AUROC) and interpreted feature contributions with Shapley values.

Main Results:

• Achieved significant predictive performance for several neurocognitive disorders.
• Highest performance for dementia (internal AUROC 0.848, external AUROC 0.865) and Alzheimer's disease (internal AUROC 0.809, external AUROC 0.863).
• Feature importance analysis identified both known and novel ECG correlates.

Conclusions:

• ECG shows potential as a non-invasive, explainable biomarker for specific neurocognitive disorders.
• The study demonstrates robust cross-cohort performance, supporting future clinical applications.
• Findings pave the way for early detection and personalized monitoring of neurocognitive disorders using ECG.