A Self-Explainable Dynamic Risk Monitoring Framework for Predicting Alzheimer's Disease and Related Dementias

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Neuroscience

Background:

• Alzheimer's Disease and Related Dementias (ADRD) pose a significant global health challenge, often progressing silently for over a decade.
• Early detection is critical as ADRD becomes largely irreversible upon symptom manifestation.
• While neuroimaging aids prediction, its limited accessibility necessitates scalable alternatives like Electronic Health Records (EHR).

Purpose of the Study:

• To develop a temporal deep learning model for early ADRD prediction using EHR data.
• To address challenges of irregular data, sparsity, and interpretability in EHR-based prediction.
• To enable real-time, interpretable risk monitoring for ADRD.

Main Methods:

• Utilized EHR data from 15,172 ADRD cases and 145,443 controls from the UT Physician EHR system.
• Developed GRU-D-RETAIN, a temporal deep learning architecture combining GRU-D imputation with RETAIN's attention mechanism.
• Trained and validated models using 6-fold cross-validation, comparing GRU-D-RETAIN against GRU-D, LSTM, Logit static, and Logit dynamic.

Main Results:

• GRU-D-RETAIN demonstrated strong performance, closely matching GRU-D, with accuracy improving as follow-up time increased.
• Models achieved AUROC of 0.7 at 8-year follow-up without data cut-offs, outperforming other methods.
• Data completeness was more critical than follow-up length for prediction accuracy; 1 year with 15% data yielded performance comparable to 7.5 years with 10% data.

Conclusions:

• EHR data can facilitate dynamic ADRD risk monitoring up to 10 years pre-diagnosis, contingent on data completeness.
• GRU-D-RETAIN provides real-time, interpretable risk assessment, aiding clinicians in identifying high-risk patients and key factors.
• The GRU-D-RETAIN framework is adaptable for other conditions requiring dynamic, interpretable risk prediction with irregular data.