Developing a Bespoke Neural Network Model for Diagnosing Alzheimer's Dementia: A Fit-for-Purpose Machine Learning Study

OBJECTIVES: This study aims to develop a bespoke, fit-for-purpose feedforward neural network (FNN) for diagnosing Alzheimer's dementia (AD). Furthermore, this study aims to identify the most impactful sources of data for FNN diagnostic performance by evaluating multiple FNN data input scenarios.

METHODS: The FNNs were made using Modelist. Data was sourced from the Alzheimer's Disease Neuroimaging Initiative (ADNI) containing retrospective clinical and diagnosis data for 8,409 total unique patients. The model was trained, validated, and tested using unique data sets at each stage, and run at five differing levels of data inputs to identify the impact on accuracy. Data inputs consisted of social determinants of health (SDOH), cerebrospinal fluid (CSF), neurocognitive assessments, and imaging results. Diagnostic accuracy was measured as the model’s ability to assign the same diagnosis as the physician (normal, mild cognitive impairment (MCI), AD).

RESULTS: Of the five total FNN scenarios evaluated, the diagnostic accuracy ranged from 43.4% (217/500) to 92.0% (460/500). The top-performing model utilized SDOH, CSF, and neurocognitive assessments while the next best performing model utilized SDOH, CSF, imaging, and neurocognitive assessments, resulting in a 91.8% (459/500) accuracy. Of the 8.0% incorrectly diagnosed (40/500) in the top-performing model, 3% (15/500) were missed AD diagnoses, 2.2% (11/500) missed MCI, and 2.8% (14/500) missed normal diagnoses. This FNN scenario had a Matthews Correlation Coefficient of 0.88 and F-1 score of 0.92.

CONCLUSIONS: Based on the model’s results, the combined data inputs of neurocognitive assessments, CSF, and SDOH resulted in the highest level of diagnostic accuracy. The addition of imaging data provided no additional benefit to diagnostic accuracy when neurocognitive assessments were also used. This study serves as a real-world example of the ability of FNNs to diagnose AD with a high level of accuracy and less data resources than standard diagnostic approaches.