Deep neural network models for identifying incident dementia using claims and EHR datasets.

Area of Science:

• Artificial Intelligence
• Neurology
• Medical Informatics

Background:

• Dementia prediction models are crucial for early intervention and clinical trial recruitment.
• Traditional machine learning models have limitations in capturing complex patterns for accurate dementia risk stratification.
• Deep learning offers potential for enhanced predictive accuracy in neurological disorders.

Purpose of the Study:

• To investigate and compare the performance of deep learning models against traditional machine learning for dementia prediction.
• To evaluate the accuracy of predictive models at varying time intervals prior to the index date.
• To assess the feasibility of deploying an accurate predictive model as an initial screening tool.

Main Methods:

• Development and comparison of four models (boosted trees, feed forward network, recurrent neural network, pre-trained recurrent neural network) across seven cohorts.
• Utilized historical data ranging from three to eight years prior to the index date.
• Performance evaluation using Area Under the Curve (AUC) and F1 scores on validation and test datasets.

Main Results:

• The incident dementia model achieved an AUC of 94.4% and an F1 score of 54.1% at the index date.
• Model performance decreased with longer lead times, with an AUC of 80.7% and F1 score of 25.6% eight years prior.
• Deep learning models demonstrated significant performance improvements but required greater computational resources.

Conclusions:

• Deep learning models can effectively stratify patients at risk of dementia, particularly for near-term prediction.
• Sustaining prediction quality at longer lead times is currently limited by data availability and quality, not algorithmic choice.
• The developed models show potential as a first-round screening tool for dementia, guiding clinical follow-up and trial recruitment.