Generative deep learning models for cognitive performance trajectories in real-world scenarios

Abstract

The increasing prevalence of cognitive disorders, such as Alzheimers disease, imposes significant challenges on healthcare systems and society. The ability to predict the future cognitive performance (CP) is crucial for professionals in neuropsychology, and real-world data emerges as an important source of complete and reliable information. However, its inherent complexities requires the use of advanced models to make predictions. To do so, we have implemented and compared three deep learning predictive strategies from CP trajectories: multilayer perceptron (MLP), convolutional neural networks (CNN) and long short-term memory (LSTM). The three models showed robustness on their predictions in different patient datasets. The CNN was the most suitable architecture due to its local pattern recognition capabilities and its robustness to overfitting. Therefore, professionals can have a complementary support for targeting treatment approaches to patients needs and anticipate undesired outcomes (e.g. cognitive impairment). Nonetheless, further studies are needed to validate whether neuropsychological interventions based on score predictions lead to improved intervention efficacy compared to traditional approaches for controlled patient groups.