Predicting Alzheimer's dementia using a multilayer perceptron: gender differences in diagnostic accuracy

Introduction: Alzheimer's dementia is a major cause of cognitive decline in older adults. Although artificial intelligence-based predictive models have great potential to improve early diagnosis, few consider gender differences in their effectiveness. Objective: To analyze the diagnostic accuracy of the multilayer perceptron in predicting Alzheimer's dementia according to biological sex. Materials and methods: A cross-sectional study of a secondary database of 373 participants. Multilayer perceptron-type neural networks were used, the variables were: age, sex, educational level, socioeconomic status, eTIV (total intracranial volume), nWBV (normalized white matter volume) and MMSE (Mini-Mental State Examination). Performance was assessed using AUC (Area Under the Curve) and diagnostic accuracy using classification tables. Results: In perceptron training, women presented a lower cross-entropy error (13.074 vs. 20.461) and percentage of incorrect predictions (5% vs. 9%). In testing, they continued with a lower cross-entropy error (7.888 vs. 16.500) and percentage of incorrect predictions (6.80% vs. 20.80%). The AUC reflected an excellent predictive capacity slightly higher in women (0.992 vs. 0.947). Classification rates were also better in women, with an overall accuracy of 95% in training and 93.20% in testing, compared to 91% and 79.20% in men. Conclusions: Biological sex influences the effectiveness of predictive models for Alzheimer's dementia. The results underline the importance of considering biological and social factors when developing diagnostic tools for Alzheimer's, which can improve the personalization of treatment and prevention.