Predicting Onset of Dementia Using Clinical Notes and Machine Learning: Case-Control Study.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Neurodegenerative Disease Research

Background:

• Efficient tools are needed for recruiting patients at risk of Alzheimer disease and related dementias (ADRD).
• Early ADRD detection aids patients in financial planning for long-term care.
• Clinical notes are an underutilized data source for machine learning due to collection costs and analysis complexity.

Purpose of the Study:

• To investigate the use of de-identified clinical notes from multiple hospital systems to augment retrospective machine learning models for ADRD risk.
• To assess the impact of incorporating 10 years of clinical note data on ADRD prediction accuracy.

Main Methods:

• Utilized 2 years of de-identified clinical note data to predict future ADRD onset.
• Embedded clinical note terms into a 100-dimensional vector space to identify clusters and differences across hospital systems.
• Employed natural language processing and postprocessing techniques to merge notes from multiple hospital systems.

Main Results:

• Machine learning models incorporating clinical notes showed improved Area Under the Curve (AUC) from 0.85 to 0.94.
• Positive Predictive Value (PPV) increased from 45.07% to 68.32% at disease onset when clinical notes were used.
• Models demonstrated improved AUC and PPV in years 3-6, with mixed results in later years due to smaller cohort sizes.

Conclusions:

• De-identified clinical notes enhance the accuracy of ADRD risk prediction models.
• The presence of symptomatic terms years before diagnosis suggests undercoding of ADRD by clinicians.
• Natural language processing of multi-system clinical notes can be effectively merged to improve model performance.